paper_id
stringlengths 10
10
⌀ | yes_no
bool 2
classes | paper_index
int64 0
519
⌀ | evidence
stringlengths 0
37.7k
| question
stringlengths 4
11.7k
| answer
stringlengths 1
26k
|
---|---|---|---|---|---|
null | false | null | You have purchased a product with which you are satisfied. Write a short review describing your experience the product: "Unisex Anti Fog No Leaking Swimming Goggles with Adjustable Silicone for Men Women Adult Youth". | These streamlined swimming goggles seem to very comparable in terms of quality with the other brands I've used. The coating on the lenses inhibits fog buildup, and the silicone straps are comfortable and broadly adjustable, providing a comfortable leak-proof seal for all sizes. I also love the universal style that can accommodate men or women. They're about as stylish as something like this could be. |
|
null | false | null | What Are the Gift Ideas for Anime Fans? | Collector edition films, cosplay outfits, limited edition manga comics, tickets to comic-con, flights to Japan. |
|
null | false | 200 | Datasets and Models We evaluate our adversarial attacks on different text classification datasets from tasks such as sentiment classification, subjectivity detection and question type classification. Amazon, Yelp, IMDB are sentence-level sentiment classification datasets which have been used in recent work BIBREF15 while MR BIBREF16 contains movie reviews based on sentiment polarity. MPQA BIBREF17 is a dataset for opinion polarity detection, Subj BIBREF18 for classifying a sentence as subjective or objective and TREC BIBREF19 is a dataset for question type classification.
We use 3 popular text classification models: word-LSTM BIBREF20, word-CNN BIBREF21 and a fine-tuned BERT BIBREF12 base-uncased classifier. For each dataset we train the model on the training data and perform the adversarial attack on the test data. For complete model details refer to Appendix.
As a baseline, we consider TextFooler BIBREF11 which performs synonym replacement using a fixed word embedding space BIBREF22. We only consider the top $K{=}50$ synonyms from the MLM predictions and set a threshold of 0.8 for the cosine similarity between USE based embeddings of the adversarial and input text.
Results We perform the 4 modes of our attack and summarize the results in Table . Across datasets and models, our BAE attacks are almost always more effective than the baseline attack, achieving significant drops of 40-80% in test accuracies, with higher average semantic similarities as shown in parentheses. BAE-R+I is the strongest attack since it allows both replacement and insertion at the same token position, with just one exception. We observe a general trend that the BAE-R and BAE-I attacks often perform comparably, while the BAE-R/I and BAE-R+I attacks are much stronger. We observe that the BERT-based classifier is more robust to the BAE and TextFooler attacks than the word-LSTM and word-CNN models which can be attributed to its large size and pre-training on a large corpus.
The baseline attack is often stronger than the BAE-R and BAE-I attacks for the BERT based classifier. We attribute this to the shared parameter space between the BERT-MLM and the BERT classifier before fine-tuning. The predicted tokens from BERT-MLM may not drastically change the internal representations learned by the BERT classifier, hindering their ability to adversarially affect the classifier prediction.
Effectiveness We study the effectiveness of BAE on limiting the number of R/I operations permitted on the original text. We plot the attack performance as a function of maximum $\%$ perturbation (ratio of number of word replacements and insertions to the length of the original text) for the TREC dataset. From Figure , we clearly observe that the BAE attacks are consistently stronger than TextFooler. The classifier models are relatively robust to perturbations up to 20$\%$, while the effectiveness saturates at 40-50$\%$. Surprisingly, a 50$\%$ perturbation for the TREC dataset translates to replacing or inserting just 3-4 words, due to the short text lengths.
Qualitative Examples We present adversarial examples generated by the attacks on a sentence from the IMDB and Yelp datasets in Table . BAE produces more natural looking examples than TextFooler as tokens predicted by the BERT-MLM fit well in the sentence context. TextFooler tends to replace words with complex synonyms, which can be easily detected. Moreover, BAE's additional degree of freedom to insert tokens allows for a successful attack with fewer perturbations.
Human Evaluation We consider successful adversarial examples generated from the Amazon and IMDB datasets and verify their sentiment and grammatical correctness. Human evaluators annotated the sentiment and the grammar (Likert scale of 1-5) of randomly shuffled adversarial examples and original texts. From Table , BAE and TextFooler have inferior accuracies compared to the Original, showing they are not always perfect. However, BAE has much better grammar scores, suggesting more natural looking adversarial examples.
Ablation Study We analyze the benefits of R/I operations in BAE in Table . From the table, the splits $\mathbb {A}$ and $\mathbb {B}$ are the $\%$ of test points which compulsorily need I and R operations respectively for a successful attack. We can observe that the split $\mathbb {A}$ is larger than $\mathbb {B}$ thereby indicating the importance of the I operation over R. Test points in split require both R and I operations for a successful attack. Interestingly, split is largest for Subj, which is the most robust to attack (Table ) and hence needs both R/I operations. Thus, this study gives positive insights towards the importance of having the flexibility to both replace and insert words.
Refer to the Appendix for additional results, effectiveness graphs and details of human evaluation.
We evaluate BAE on different text classification tasks. Amazon, Yelp, IMDB are sentiment classification datasets used in recent works (Sarma et al., 2018) and MR (Pang and Lee, 2005) contains movie reviews based on sentiment polarity. | Which datasets are sentiment classification datasets used in recent works? | Amazon, Yelp, IMDB are sentiment classification datasets used in recent works. |
null | false | null | Mini survival kits or "Altoids" tin survival kits are small kits that contain a few basic survival tools. These kits often include a small compass, waterproof matches, minimum fishing tackle, large plastic bag, small candle, jigsaw blade, craft knife or scalpel blade, and/or a safety pin/s. Pre-packaged survival kits may also include instructions in survival techniques such as fire-starting or first aid methods. In addition, paracord can be wrapped around the tin. The paracord can be used for setting up an emergency shelter or snaring small animals. They are designed to fit within a container roughly the size of a mint tin. | Extract the essential tools for a mini survival kit from the following Wikipedia article. | Small compass, waterproof matches, minimum fishing tackle, large plastic bag, small candle, jigsaw blade, craft knife, and safety pins. |
null | false | 18 | Neural Machine Translation (NMT) has shown its effectiveness in translation tasks when NMT systems perform best in recent machine translation campaigns BIBREF0 , BIBREF1 . Compared to phrase-based Statistical Machine Translation (SMT) which is basically an ensemble of different features trained and tuned separately, NMT directly modeling the translation relationship between source and target sentences. Unlike SMT, NMT does not require much linguistic information and large monolingual data to achieve good performances.
An NMT consists of an encoder which recursively reads and represents the whole source sentence into a context vector and a recurrent decoder which takes the context vector and its previous state to predict the next target word. It is then trained in an end-to-end fashion to learn parameters which maximizes the likelihood between the outputs and the references. Recently, attention-based NMT has been featured in most state-of-the-art systems. First introduced by BIBREF2 , attention mechanism is integrated in decoder side as feedforward layers. It allows the NMT to decide which source words should take part in the predicting process of the next target words. It helps to improve NMTs significantly. Nevertheless, since the attention mechanism is specific to a particular source sentence and the considering target word, it is also specific to particular language pairs.
Some recent work has focused on extending the NMT framework to multilingual scenarios. By training such network using parallel corpora in number of different languages, NMT could benefit from additional information embedded in a common semantic space across languages. Basically, the proposed NMT are required to employ multiple encoders or multiple decoders to deal with multilinguality. Furthermore, in order to avoid the tight dependency of the attention mechanism to specific language pairs, they also need to modify their architecture to combine either the encoders or the attention layers. These modifications are specific to the purpose of the tasks as well. Thus, those multilingual NMTs are more complicated, much more free parameters to learn and more difficult to perform standard trainings compared to the original NMT.
In this paper, we introduce a unified approach to seamlessly extend the original NMT to multilingual settings. Our approach allows us to integrate any language in any side of the encoder-decoder architecture with only one encoder and one decoder for all the languages involved. Moreover, it is not necessary to do any network modification to enable attention mechanism in our NMT systems. We then apply our proprosed framework in two demanding scenarios: under-resourced translation and zero-resourced translation. The results show that bringing multilinguality to NMT helps to improve individual translations. With some insightful analyses of the results, we set our goal toward a fully multilingual NMT framework.
The paper starts with a detailed introduction to attention-based NMT. In Section SECREF3 , related work about multi-task NMT is reviewed. Section SECREF5 describes our proposed approach and thorough comparisons to the related work. It is followed by a section of evaluating our systems in two aforementioned scenarios, in which different strategies have been employed under a unified approach (Section SECREF4 ). Finally, the paper ends with conclusion and future work.
This work is licenced under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/
An NMT consists of an encoder which recursively reads and represents the whole source sentence into a context vector and a recurrent decoder which takes the context vector and its previous state to predict the next target word. | What does an NMT consist of? | An encoder and a recurrent decoder. |
null | false | null | What causes that feeling of weightlessness when you're on a rollercoaster? | The weightless feeling from a rollercoaster is caused by gravitational force, also known as g-force. Different rollercoaster elements lead to different types of g-force: there is positive g-force, which drives blood downwards towards the feet, and negative g-force, which drives blood upwards towards the head. Negative g-force acts opposite to gravity, and it occurs when you accelerate downwards faster than free-fall. The acceleration causes you to pull out of your seat, evoking a sense of weightlessness. |
|
null | false | null | Boiling water is 100 degree in celsius, is that an obtuse angle? | No. An obtuse angle is any angle greater than 90 degree. The degree here is a unit of measure used to measure the magnitude of an angle, and it's a Geometric concept. Although boiling water is 100 degree, the degree here is a scale of temperature. The degrees refer to different things. Therefore, boiling water is not even an angle, and it's not an obtuse angle. |
|
null | false | null | How many musicians performed in the world's largest rock band? | The world’s largest performing rock band included 953 musicians. |
|
1910.03891 | false | null | In order to address this issue, TransH BIBREF8 models a relation as a relation-specific hyperplane together with a translation on it, allowing entities to have distinct representation in different relations. TransR BIBREF9 models entities and relations in separate spaces, i.e., entity space and relation spaces, and performs translation from entity spaces to relation spaces. TransD BIBREF22 captures the diversity of relations and entities simultaneously by defining dynamic mapping matrix. Recent attempts can be divided into two categories: (i) those which tries to incorporate additional information to further improve the performance of knowledge graph embedding, e.g., entity types or concepts BIBREF13, relations paths BIBREF17, textual descriptions BIBREF11, BIBREF12 and logical rules BIBREF23; (ii) those which tries to design more complicated strategies, e.g., deep neural network models BIBREF24.
Recent attempts can be divided into two categories: (i) those which tries to incorporate additional information to further improve the performance of knowledge graph embedding, e.g., entity types or concepts BIBREF13, relations paths BIBREF17, textual descriptions BIBREF11, BIBREF12 and logical rules BIBREF23; (ii) those which tries to design more complicated strategies, e.g., deep neural network models BIBREF24. | What are recent works on knowedge graph embeddings authors mention? | The answers are shown as follows:
* entity types or concepts BIBREF13
* relations paths BIBREF17
* textual descriptions BIBREF11, BIBREF12
* logical rules BIBREF23
* deep neural network models BIBREF24
|
null | false | null | The chateau has four wings and is surrounded by arcades lining the courtyard. The yard is accessible from the entrance gate and leads to the Italian symmetrical garden. There are also four fountains in the garden by stonemason Jakub Mitthofer. | From the passage provided, extract the name of the person that made some fountains in the garden, and how many. Your answer shall be of the format "{name} built {number of fountains} fountains." | Jakub Mitthofer built four fountains. |
1910.09295 | false | null | We use a siamese neural network, shown to perform state-of-the-art few-shot learning BIBREF11, as our baseline model.
We modify the original to account for sequential data, with each twin composed of an embedding layer, a Long-Short Term Memory (LSTM) BIBREF12 layer, and a feed-forward layer with Rectified Linear Unit (ReLU) activations.
We use a siamese neural network, shown to perform state-of-the-art few-shot learning BIBREF11, as our baseline model.
We modify the original to account for sequential data, with each twin composed of an embedding layer, a Long-Short Term Memory (LSTM) BIBREF12 layer, and a feed-forward layer with Rectified Linear Unit (ReLU) activations. | What were the baselines? | Siamese neural network consisting of an embedding layer, a LSTM layer and a feed-forward layer with ReLU activations |
null | false | 122 | The sentence level classification task is an imbalanced binary classification problem that we address using BERT BIBREF0. We use BERTBASE, uncased, which consists of 12 self-attention layers, and returns a 768-dimension vector that representation a sentence. So as to make use of BERT for sentence classification, we include a fully connected layer on top of the BERT self-attention layers, which classifies the sentence embedding provided by BERT into the two classes of interest (propaganda or non-propaganda).
We attempt to exploit various data augmentation techniques to address the problem of class imbalance. Table TABREF17 shows the results of our experiments for different data augmentation techniques when, after shuffling the training data, we train the model on 75% of the training data and test it on the remaining 25% of the training data and the development data.
We observe that BERT without augmentation consistently outperforms BERT with augmentation in the experiments when the model is trained on 75% of the training data and evaluated on the rest, i.e trained and evaluated on similar data, coming from the same distribution. This is consistent with observations by Wei et al. wei2019eda that contextual word embeddings do not gain from data augmentation. The fact that we shuffle the training data prior to splitting it into training and testing subsets could imply that the model is learning to associate topic words, such as `Mueller', as propaganda. However, when we perform model evaluation using the development set, which is dissimilar to the training, we observe that synonym insertion and word dropping techniques also do not bring performance gains, while random oversampling increases performance over base BERT by 4%. Synonym insertion provides results very similar to base BERT, while random deletion harms model performance producing lower scores. We believe that this could be attributed to the fact that synonym insertion and random word dropping involve the introduction of noise to the data, while oversampling does not. As we are working with natural language data, this type of noise can in fact change the meaning of the sentence. Oversampling on the other hand purely increases the importance of the minority class by repeating training on the unchanged instances.
So as to better understand the aspects of oversampling that contribute to these gains, we perform a class-wise performance analysis of BERT with/without oversampling. The results of these experiments (Table TABREF18) show that oversampling increases the overall recall while maintaining precision. This is achieved by significantly improving the recall of the minority class (propaganda) at the cost of the recall of the majority class.
So far we have been able to establish that a) the training and test sets are dissimilar, thus requiring us to generalise our model, b) oversampling provides a method of generalisation, and c) oversampling does this while maintaining recall on the minority (and thus more interesting) class.
Given this we explore alternative methods of increasing minority class recall without a significant drop in precision. One such method is cost-sensitive classification, which differs from random oversampling in that it provides a more continuous-valued and consistent method of weighting samples of imbalanced training data; for example, random oversampling will inevitably emphasise some training instances at the expense of others. We detail our methods of using cost-sensitive classification in the next section. Further experiments with oversampling might have provided insights into the relationships between these methods, which we leave for future exploration.
One of the major setbacks to our method for identifying sentence fragments was the loss of training data as a result of randomly picking one label when the same fragment had multiple labels. This could have been avoided by training different models for each label and simply concatenating the results. Additionally, training on all sentences, including those that did not contain any fragments labelled as propagandistic would have likely improved our model performance. We intend to perform these experiments as part of our ongoing research. | What are the limitations of the method? | One of the major setbacks to their method for identifying sentence fragments was the loss of training data as a result of randomly picking one label when the same fragment had multiple labels. This could have been avoided by training different models for each label and simply concatenating the results. Additionally, training on all sentences, including those that did not contain any fragments labelled as propagandistic would have likely improved their model performance. They intend to perform these experiments as part of their ongoing research. |
null | false | null | What are the top 10 suggestions for fillings in Mexican street tacos? | 1. Steak (carne asada)
2. Chicken (pollo)
3. Beef Cheeks (cabeza)
4. Braised Beef (barbacoa)
5. Braised Pork (carnitas)
6. Pork Marinated with Pineapple (al pastor)
7. Shrimp (camaron)
8. Sausage (chorizo)
9. Fish (pescado)
10. Beans (frijoles) |
|
null | false | null | Of the following, which are colors of a rainbow: black, red, white, gold, purple, orange, green, blue, yellow, gray | Red, purple, orange, green, blue, yellow |
|
1911.03385 | true | null | We compare the above model to a similar model, where rather than explicitly represent $K$ features as input, we have $K$ features in the form of a genre embedding, i.e. we learn a genre specific embedding for each of the gothic, scifi, and philosophy genres, as studied in BIBREF8 and BIBREF7. To generate in a specific style, we simply set the appropriate embedding. We use genre embeddings of size 850 which is equivalent to the total size of the $K$ feature embeddings in the StyleEQ model.
We compare the above model to a similar model, where rather than explicitly represent $K$ features as input, we have $K$ features in the form of a genre embedding, i.e. we learn a genre specific embedding for each of the gothic, scifi, and philosophy genres, as studied in BIBREF8 and BIBREF7. | Is this style generator compared to some baseline? | Yes. |
null | false | null | Identify which instrument is string or woodwind: Bandol, Soprano clarinet | Soprano clarinet is woodwind, Bandol is string. |
|
null | false | null | How do you determine the severity of vulnerabilities. | The severity of vulnerabilities are commonly determined using the common vulnerability scoring system (CVSS). CVSS scores vulnerabilities from 0-10, with 10 being the most critical severity score. |
|
null | false | null | What US embassy was abandoned on April 12th, 1975 | Phnom Penh was abandoned after being besieged by the Khmer Rouge on April 12th 1975 |
|
null | false | null | What is a good park in Redmond, WA | Marymoor Park in Redmond, WA |
|
null | false | null | Tell me whether each airline is American or European: British Airways, United Airlines, Southwest Airlines, Jet Blue, Easy Jet, Ryan Air | British Airways: European
United Airlines: American
Southwest Airlines: American
Jet Blue: American
Easy Jet: European
Ryan Air: European |
|
null | false | null | Is it possible to perform in local theater while also pursuing a demanding career? | Yes, however there are many implications that should be thought through, as theater (even at the local, unpaid volunteer level) is a time-consuming and rigorous hobby. 1) Career choice. Restaurants, retail and other careers that required nights and weekends regularly often conflict with rehearsal and performance schedules, which are often all on nights and weekends. 2) Travel. Planned, infrequent travel can usually be planned around. But if your career requires spontaneous travel on a moments notice, it may be difficult for you to commit to the rehearsal and performance schedule, which is often planned months in advance. |
|
null | false | null | The movement of electric charge is known as an electric current, the intensity of which is usually measured in amperes. Current can consist of any moving charged particles; most commonly these are electrons, but any charge in motion constitutes a current. Electric current can flow through some things, electrical conductors, but will not flow through an electrical insulator.
By historical convention, a positive current is defined as having the same direction of flow as any positive charge it contains, or to flow from the most positive part of a circuit to the most negative part. Current defined in this manner is called conventional current. The motion of negatively charged electrons around an electric circuit, one of the most familiar forms of current, is thus deemed positive in the opposite direction to that of the electrons. However, depending on the conditions, an electric current can consist of a flow of charged particles in either direction, or even in both directions at once. The positive-to-negative convention is widely used to simplify this situation.
The process by which electric current passes through a material is termed electrical conduction, and its nature varies with that of the charged particles and the material through which they are travelling. Examples of electric currents include metallic conduction, where electrons flow through a conductor such as metal, and electrolysis, where ions (charged atoms) flow through liquids, or through plasmas such as electrical sparks. While the particles themselves can move quite slowly, sometimes with an average drift velocity only fractions of a millimetre per second, the electric field that drives them itself propagates at close to the speed of light, enabling electrical signals to pass rapidly along wires.
In engineering or household applications, current is often described as being either direct current (DC) or alternating current (AC). These terms refer to how the current varies in time. Direct current, as produced by example from a battery and required by most electronic devices, is a unidirectional flow from the positive part of a circuit to the negative. If, as is most common, this flow is carried by electrons, they will be travelling in the opposite direction. Alternating current is any current that reverses direction repeatedly; almost always this takes the form of a sine wave. Alternating current thus pulses back and forth within a conductor without the charge moving any net distance over time. The time-averaged value of an alternating current is zero, but it delivers energy in first one direction, and then the reverse. Alternating current is affected by electrical properties that are not observed under steady state direct current, such as inductance and capacitance. These properties however can become important when circuitry is subjected to transients, such as when first energised. | What do the terms direct current (DC) or alternating current (AC) mean? | In engineering or household applications, current is often described as being either direct current (DC) or alternating current (AC). These terms refer to how the current varies in time. Direct current, as produced by example from a battery and required by most electronic devices, is a unidirectional flow from the positive part of a circuit to the negative. If, as is most common, this flow is carried by electrons, they will be travelling in the opposite direction. Alternating current is any current that reverses direction repeatedly; almost always this takes the form of a sine wave. Alternating current thus pulses back and forth within a conductor without the charge moving any net distance over time. |
null | false | null | Queen Elizabeth II was born on April 21, 1926 in | Mayfair, London |
|
null | false | 57 | Our framework consists of a series of steps for which there are choices and alternatives. Although it is not possible to provide comparisons to the myriad of methods and possibilities available, we have examined quantitatively the robustness of the results to parametric and methodological choices in different steps of the framework: (i) the importance of using Doc2Vec embeddings instead of BoW vectors, (ii) the size of training corpus for Doc2Vec; (iii) the sparsity of the MST-kNN similarity graph construction. We have also carried out quantitative comparisons to other methods, including: (i) LDA-BoW, and (ii) clustering with other community detection methods. We provide a brief summary here and additional material in the SI.
The use of fixed-sized vector embeddings (Doc2Vec) instead of standard bag of words (BoW) is an integral part of our pipeline. Doc2Vec produces lower dimensional vector representations (as compared to BoW) with higher semantic and syntactic content. It has been reported that Doc2Vec outperforms BoW representations in practical benchmarks of semantic similarity, as well as being less sensitive to hyper-parameters BIBREF22 .
To quantify the improvement provided by Doc2Vec in our framework, we constructed a MST-kNN graph following the same steps but starting with TF-iDF vectors for each document. We then ran MS on this TF-iDF similarity graph, and compared the results to those obtained from the Doc2Vec similarity graph. Figure 7 shows that the Doc2Vec version outperforms the BoW version across all resolutions in terms of both $NMI$ and $\widehat{PMI}$ scores.
As shown in Table 1 , we have tested the effect of the size of the training corpus on the Doc2Vec model. We trained Doc2Vec on two additional training sets of 1 million and 2 million records (randomly chosen from the full set of $\sim $ 13 million records). We then followed the same procedure to construct the MST-kNN similarity graph and carried out the MS analysis. The results, presented in Figure S3 in the SI, show that the performance is affected only mildly by the size of the Doc2Vec training set.
To examine the effect of sparsification in the graph construction, we have studied the dependence of quality of the partitions against the number of neighbours, $k$ , in the MST-kNN graph. Our numerics, shown in Figure S4 in the SI, indicate that both the $NMI$ and $\widehat{PMI}$ scores of the MS clusterings reach a similar level of quality for values of $k$ above 13-16, with minor improvement after that. Hence our results are robust to the choice of $k$ , provided it is not too small. Due to computational efficiency, we thus favour a relatively small $k$ , but not too small.
We carried out a comparison with LDA, a widely used methodology for text analysis. A key difference between standard LDA and our MS method is the fact that a different LDA model needs to be trained separately for each number of topics pre-determined by the user. To offer a comparison across the methods, We obtained five LDA models corresponding to the five MS levels we considered in detail. The results in Table 2 show that MS and LDA give partitions that are comparably similar to the hand-coded categories (as measured with $NMI$ ), with some differences depending on the scale, whereas the MS clusterings have higher topic coherence (as given by $\widehat{PMI}$ ) across all scales.
To give an indication of the computational cost, we ran both methods on the same servers. Our method takes approximately 13 hours in total to compute both the Doc2Vec model on 13 million records (11 hours) and the full MS scan with 400 partitions across all resolutions (2 hours). The time required to train just the 5 LDA models on the same corpus amounts to 30 hours (with timings ranging from $\sim $ 2 hours for the 3 topic LDA model to 12.5 hours for the 44 topic LDA model).
This comparison also highlights the conceptual difference between our multi-scale methodology and LDA topic modelling. While LDA computes topics at a pre-determined level of resolution, our method obtains partitions at all resolutions in one sweep of the Markov time, from which relevant partitions are chosen based on their robustness. However, the MS partitions at all resolutions are available for further investigation if so needed.
We have used several algorithms readily available in code libraries (i.e., the iGraph module for Python) to cluster/partition the same kNN-MST graph. Figure S5 in the SI shows the comparison against several well-known partitioning methods (Modularity Optimisation BIBREF44 , InfoMap BIBREF4 , Walktrap BIBREF45 , Label Propagation BIBREF46 , and Multi-resolution Louvain BIBREF35 ) which give just one partition (or two in the case of the Louvain implementation in iGraph) into a particular number of clusters, in contrast with our multiscale MS analysis. Our results show that MS provides improved or equal results to other graph partitioning methods for both $NMI$ and $\widehat{PMI}$ across all scales. Only for very fine resolution with more than 50 clusters, Infomap, which partitions graphs into small clique-like subgraphs BIBREF32 , BIBREF47 , provides a slightly improved $NMI$ for that particular scale. Therefore, MS allows us to find relevant, yet high quality clusterings across all scales by sweeping the Markov time parameter.
To quantify the improvement provided by Doc2Vec in our framework, we constructed a MST-kNN graph following the same steps but starting with TF-iDF vectors for each document. | What was constructed to quantify the improvement provided by Doc2Vec? | A MST-kNN graph. |
null | false | 155 | DNN models learn word embeddings over the training data. These learned embeddings across multiple datasets show the difference in nature and style of bullying across cyberbullying topics and SMPs. Here we report results for BLSTM with attention model. Results for other models are similar. We first verify that important words for each topic of cyberbullying form clusters in the learned embeddings. To enable the visualization of grouping, we reduced dimensionality with t-SNE BIBREF16 , a well-known technique for dimensionality reduction particularly well suited for visualization of high dimensional datasets. Please refer to Table TABREF22 . This table shows important clusters observed in t-SNE projection of learned word embeddings. Each cluster shows that words most relevant to a particular topic of bullying form cluster.
We also observed changes in the meanings of the words across topics of cyberbullying. Table TABREF23 shows most similar words for a given query word for two datasets. Twitter dataset which is heavy on sexism and racism, considers word slave as similar to targets of racism and sexism. However, Wikipedia dataset that is about personal attacks does not show such bias.
We first verify that important words for each topic of cyberbullying form clusters in the learned embeddings. To enable the visualization of grouping, we reduced dimensionality with t-SNE , a well-known technique for dimensionality reduction particularly well suited for visualization of high dimensional datasets. | How does the author verify the results? | They first verify that important words for each topic of cyberbullying form clusters in the learned embeddings. To enable the visualization of grouping, they reduced dimensionality with t-SNE , a well-known technique for dimensionality reduction particularly well suited for visualization of high dimensional datasets. |
null | false | 321 | The past decade witnessed rapid growth and widespread usage of social media platforms by generating a significant amount of user-generated text. The user-generated texts contain high information content in the form of news, expression, or knowledge. Automatically mining information from user-generated data is unraveling a new field of research in Natural Language Processing (NLP) and has been a difficult task due to unstructured and noisy nature. In spite of the existing challenges, much research has been conducted on user-generated data in the field of information extraction, sentiment analysis, event extraction, user profiling and many more.
According to Census of India, there are 22 scheduled languages and more than 100 non scheduled languages in India. There are 462 million internet users in India and most people know more than one language. They express their feelings or emotions using more than one languages, thus generating a new code-mixed/code-switched language. The problem of code-mixing and code-switching are well studied in the field of NLP BIBREF0 , BIBREF1 . Information extraction from Indian internet user-generated texts become more difficult due to this multilingual nature. Much research has been conducted in this field such as language identification BIBREF2 , BIBREF3 , part-of-speech tagging BIBREF4 . Joshi et al. JoshiPSV16 have performed sentiment analysis in Hindi-English (HI-EN) code-mixed data and almost no work exists on sentiment analysis of Bengali-English (BN-EN) code-mixed texts. The Sentiment Analysis of Indian Language (Code-Mixed) (SAIL _Code-Mixed) is a shared task at ICON-2017. Two most popular code-mixed languages namely Hindi and Bengali mixed with English were considered for the sentiment identification task. A total of 40 participants registered for the shared task and only nine teams have submitted their predicted outputs. Out of nine unique submitted systems for evaluation, eight teams submitted fourteen runs for HI-EN dataset whereas seven teams submitted nine runs for BN-EN dataset. The training and test dataset were provided after annotating the languages and sentiment (positive, negative, and neutral) tags. The language tags were automatically annotated with the help of different dictionaries whereas the sentiment tags were manually annotated. The submitted systems are ranked using the macro average f-score.
The paper is organized as following manner. Section SECREF2 describes the NLP in Indian languages mainly related to code-mixing and sentiment analysis. The detailed statistics of the dataset and evaluation are described in Section SECREF3 . The baseline systems and participant's system description are described in Section SECREF4 . Finally, conclusion and future research are drawn in Section SECREF5 .
The language tags were automatically annotated with the help of different dictionaries whereas the sentiment tags were manually annotated. | Which one was manually annotated? Language tags or sentiment tags? | The sentiment tags. |
null | false | null | Why do toddlers wake up so early? | A toddlers sleep patterns are a product of how they are trained as a sleeper from early on in life. Toddlers who were removed from a crib regularly in their early months tend to be inconsistent sleepers into the toddler years. The brain development in a toddlers brain is so malleable, they rely heavily on their inherent circadian rhythms to wake and fall asleep. |
|
1704.04521 | false | null | FLOAT SELECTED: Figure 2: NMT training after replacing technical term pairs with technical term tokens “TTi” (i = 1, 2, . . .)
FLOAT SELECTED: Figure 3: NMT decoding with technical term tokens “TTi” (i = 1, 2, . . .) and SMT technical term translation
FLOAT SELECTED: Figure 4: NMT rescoring of 1,000-best SMT translations with technical term tokens “TTi” (i = 1, 2, . . .)
In this paper, we propose a method that enables NMT to translate patent sentences with a large vocabulary of technical terms. We use an NMT model similar to that used by Sutskever et al. Sutskever14, which uses a deep long short-term memories (LSTM) BIBREF7 to encode the input sentence and a separate deep LSTM to output the translation. We train the NMT model on a bilingual corpus in which the technical terms are replaced with technical term tokens; this allows it to translate most of the source sentences except technical terms. Similar to Sutskever et al. Sutskever14, we use it as a decoder to translate source sentences with technical term tokens and replace the tokens with technical term translations using statistical machine translation (SMT). We also use it to rerank the 1,000-best SMT translations on the basis of the average of the SMT and NMT scores of the translated sentences that have been rescored with the technical term tokens. Our experiments on Japanese-Chinese patent sentences show that our proposed NMT system achieves a substantial improvement of up to 3.1 BLEU points and 2.3 RIBES points over a traditional SMT system and an improvement of approximately 0.6 BLEU points and 0.8 RIBES points over an equivalent NMT system without our proposed technique.
One important difference between our NMT model and the one used by Sutskever et al. Sutskever14 is that we added an attention mechanism. Recently, Bahdanau et al. Bahdanau15 proposed an attention mechanism, a form of random access memory, to help NMT cope with long input sequences. Luong et al. Luong15b proposed an attention mechanism for different scoring functions in order to compare the source and target hidden states as well as different strategies for placing the attention. In this paper, we utilize the attention mechanism proposed by Bahdanau et al. Bahdanau15, wherein each output target word is predicted on the basis of not only a recurrent hidden state and the previously predicted word but also a context vector computed as the weighted sum of the hidden states.
According to the approach proposed by Dong et al. Dong15b, we identify Japanese-Chinese technical term pairs using an SMT phrase translation table. Given a parallel sentence pair $\langle S_J, S_C\rangle $ containing a Japanese technical term $t_J$ , the Chinese translation candidates collected from the phrase translation table are matched against the Chinese sentence $S_C$ of the parallel sentence pair. Of those found in $S_C$ , $t_C$ with the largest translation probability $P(t_C\mid t_J)$ is selected, and the bilingual technical term pair $\langle t_J,t_C\rangle $ is identified.
For the Japanese technical terms whose Chinese translations are not included in the results of Step UID11 , we then use an approach based on SMT word alignment. Given a parallel sentence pair $\langle S_J, S_C\rangle $ containing a Japanese technical term $t_J$ , a sequence of Chinese words is selected using SMT word alignment, and we use the Chinese translation $t_C$ for the Japanese technical term $t_J$ .
Figure 3 illustrates the procedure for producing Chinese translations via decoding the Japanese sentence using the method proposed in this paper. In the step 1 of Figure 3 , when given an input Japanese sentence, we first automatically extract the technical terms and replace them with the technical term tokens “ $TT_{i}$ ” ( $i=1,2,\ldots $ ). Consequently, we have an input sentence in which the technical term tokens “ $TT_{i}$ ” ( $i=1,2,\ldots $ ) represent the positions of the technical terms and a list of extracted Japanese technical terms. Next, as shown in the step 2-N of Figure 3 , the source Japanese sentence with technical term tokens is translated using the NMT model trained according to the procedure described in Section "NMT Training after Replacing Technical Term Pairs with Tokens" , whereas the extracted Japanese technical terms are translated using an SMT phrase translation table in the step 2-S of Figure 3 . Finally, in the step 3, we replace the technical term tokens “ $TT_{i}$ ” ( $i=1,2,\ldots $ ) of the sentence translation with SMT the technical term translations.
FLOAT SELECTED: Figure 2: NMT training after replacing technical term pairs with technical term tokens “TTi” (i = 1, 2, . . .)
FLOAT SELECTED: Figure 3: NMT decoding with technical term tokens “TTi” (i = 1, 2, . . .) and SMT technical term translation
FLOAT SELECTED: Figure 4: NMT rescoring of 1,000-best SMT translations with technical term tokens “TTi” (i = 1, 2, . . .)
We use an NMT model similar to that used by Sutskever et al. Sutskever14, which uses a deep long short-term memories (LSTM) BIBREF7 to encode the input sentence and a separate deep LSTM to output the translation. We train the NMT model on a bilingual corpus in which the technical terms are replaced with technical term tokens; this allows it to translate most of the source sentences except technical terms. Similar to Sutskever et al. Sutskever14, we use it as a decoder to translate source sentences with technical term tokens and replace the tokens with technical term translations using statistical machine translation (SMT). We also use it to rerank the 1,000-best SMT translations on the basis of the average of the SMT and NMT scores of the translated sentences that have been rescored with the technical term tokens.
In this paper, we utilize the attention mechanism proposed by Bahdanau et al. Bahdanau15, wherein each output target word is predicted on the basis of not only a recurrent hidden state and the previously predicted word but also a context vector computed as the weighted sum of the hidden states.
According to the approach proposed by Dong et al. Dong15b, we identify Japanese-Chinese technical term pairs using an SMT phrase translation table.
For the Japanese technical terms whose Chinese translations are not included in the results of Step UID11 , we then use an approach based on SMT word alignment
Finally, in the step 3, we replace the technical term tokens “ $TT_{i}$ ” ( $i=1,2,\ldots $ ) of the sentence translation with SMT the technical term translations. | Can the approach be generalized to other technical domains as well? | There is no reason to think that this approach wouldn't also be successful for other technical domains. Technical terms are replaced with tokens, therefore so as long as there is a corresponding process for identifying and replacing technical terms in the new domain this approach could be viable. |
null | false | null | League of Legends (LoL), commonly referred to as League, is a 2009 multiplayer online battle arena video game developed and published by Riot Games. Inspired by Defense of the Ancients, a custom map for Warcraft III, Riot's founders sought to develop a stand-alone game in the same genre. Since its release in October 2009, League has been free-to-play and is monetized through purchasable character customization. The game is available for Microsoft Windows and macOS.
In the game, two teams of five players battle in player-versus-player combat, each team occupying and defending their half of the map. Each of the ten players controls a character, known as a "champion", with unique abilities and differing styles of play. During a match, champions become more powerful by collecting experience points, earning gold, and purchasing items to defeat the opposing team. In League's main mode, Summoner's Rift, a team wins by pushing through to the enemy base and destroying their "Nexus", a large structure located within. | How do you play the game of LoL? | League of legends (LoL) is a multiplayer online battle game with a team of 5 players on each side of the arena playing against each other. Each player controls a "champion" with unique abilities and styles of the play. During the play, the champion of each player will become stronger by leveling up and collecting gold, which can be used to buy powerful items. The goal of the game is to destroy the "Nexus" of the opposite team. |
null | false | null | What is L-AP4 | L-AP4 (L-2-amino-4-phosphonobutyric acid) is a drug used in scientific research, which acts as a group-selective agonist for the group III metabotropic glutamate receptors (mGluR4/6/7/8). It was the first ligand found to act as an agonist selective for this group of mGlu receptors, but does not show selectivity between the different mGluR Group III subtypes. It is widely used in the study of this receptor family and their various functions. |
|
null | false | null | Swiftwater is an unincorporated community in Pocono Township, Monroe County, Pennsylvania, United States.
The community is located within the Paradise Creek Watershed. Upper Swiftwater creek is designated by the PA DEP as exceptional value waters (EV). The community of Swiftwater is located within and around unique topographic features in the Poconos, which encompass the Pocono Plateau Escarpment. Land use consists of 70% forested land, 14% low density residential, 13.6% agricultural lands and approximately 2.4% wetlands. The acres of forested land are nearby to State Game Lands, an important bird area, which support beaver, raccoon, gray, fox, coyote, and mink, and Snow Shoe Hares.
In 1897 Richard Slee created the Pocono Biological Laboratories in Swiftwater. Swiftwater is home to the biggest flu vaccine plant in the United States.
Swiftwater is also the home of the Pocono Cheesecake Factory, located on SR 611.[citation needed] | How much of the land of Swiftwater community is forested land and how much of it is wetland? | According to the paragraph, 70% of the Swiftwater land is forested land and approximately 2.4% of it is wetlands. |
null | false | 513 | Previous DPPLs, DeepProbLog and NeurASP, introduced the Neural Predicate as an annotated-disjunction or as a propositional atom, respectively, to acquire conditional class probabilities, P (C|X), via the softmax function at the output of an arbitrary DNN. As mentioned in the introduction, this approach has certain limitations concerning inference capabilities. To resolve this issue, we introduce Neural-Probabilisitic Predicates (NPPs).
Formally, we denote with
a Neural-Probabilistic Predicate h. Thereby, (i) npp is a reserved word to label an NPP, (ii) h a symbolic name of either a PC, NN or a joint of a PC and NN (cf. Fig.), e.g., color_attr is the name of an NPP of Fig.. Additionally, (iii) x denotes a "term" and (iv) v 1 , . . . , v n are placeholders for each of the n possible outcomes of h. For example, the placeholders for color_attr are the color attributes of an object (Red, Blue, Green, etc.).
An NPP abbreviates an arithmetic literal of the form c = v with c ∈ {h(x)} and v ∈ {v 1 , . . . , v n }. Furthermore, we denote with Π npp a set of NPPs of the form stated in (Eq. 1) and r npp the set of all rules c = v of one NPP, which denotes the possible outcomes, obtained from an NPP in Π npp , e.g. r color_attr = {c = Red, c = Blue, c = Green, ...} for the example depicted in Fig..
Rules of the form npp (h(x), [v 1 , . . . , v n ]) ← Body are used as an abbreviation for application to multiple entities, e.g. multiple slots for the task of set prediction (cf. Fig.). Hereby, Body of the rule is identified by (tautology, true) or ⊥ (contradiction, false) during grounding. Rules of the form Head ← Body with r npp appearing in Head are prohibited for Π npp .
In this work, we largely make use of NPPs that contain probabilistic circuits (specifically SPNs) which allow for tractable density estimation and modelling of joint probabilities. In this way, it is possible to answer a much richer set of probabilistic queries, i.e. P (X, C), P (X|C) and P (C|X).
In addition to this, we introduce the arguably more interesting type of NPP that combines a neural module with a PC. Hereby, the neural module learns to map the raw input data into an optimal latent representation, e.g. object-based slot representations. The PC, in turn, learns to model the joint distribution of these latent variables and produces the final probability estimates. This type of NPP nicely combines the representational power of neural networks with the advantages of PCs in probability estimation and query flexibility.
For making the different probabilistic queries distinguishable in a SLASH program, we introduce the following notation. We denote a given variable with + and the query variable with −. E.g., within the running example of set prediction (cf. Fig.), with the query color_attr(+X, −C) one is asking for P (C|X). Similarly, with color_attr(−X, +C) one is asking for P (X|C) and, finally, with color_attr(−X, −C) for P (X, C).
To summarize, an NPP can consist of neural and/or probabilistic modules and produces querydependent probability estimates. Due to the flexibility of its definition, the term NPP contains the predicates of previous works, but also more interesting predicates discussed above. The specific "flavor" of an NPP should be chosen depending on what type of probability estimation is required. Lastly, NPPs have the unified loss function of the negative log-likelihood:
(2) whereby we are assuming the data to be i.i.d., ground truth x i to be the all-ones vector, ξ to be the parameters of the NPP and P (X,C) ξ are the predictions xi obtained from the PC encoded in the NPP.
The advantage of SLASH lies in the efficient integration of neural, probabilistic and symbolic computations. To emphasize this, we conduct a variety of experimental evaluations.
Experimental Details. We use two benchmark data sets, namely for the task of MNIST-Addition and a variant of the ShapeWorld data set for For ShapeWorld experiments, we generate a data set we refer to as ShapeWorld4. Images of ShapeWorld4 contain between one and four objects, with each object consisting of four attributes: a color (red, blue, green, gray, brown, magenta, cyan or yellow), a shade (bright, or dark), a shape (circle, triangle or square) and a size (small or big). Thus, each object can be created from 84 different combinations of attributes. Fig. depicts an example image.
We measure performance via classification accuracies in the MNIST-Addition task. In our Shape-World4 experiments, we present the average precision. We refer to appendix B for the SLASH programs and queries of each experiment, and appendix C for a detailed description of hyperparameters and further details.
Evaluation 1: SLASH outperforms SOTA DPPLs in MNIST-Addition. The task of MNIST-Addition is to predict the sum of two MNIST digits, presented only as raw images. During test time, however, a model should classify the images directly. Thus, although a model does not receive explicit information about the depicted digits, it must learn to identify digits via indirect feedback on the sum prediction.
We compare the test accuracy after convergence between the three DPPLs: DeepProbLog, NeurASP and SLASH, using a probabilistic circuit (PC) or a deep neural network (DNN) as NPP. Notably, the DNN used in SLASH (DNN) is the LeNet5 model of DeepProbLog and NeurASP. We note that when using the PC as NPP, we have also extracted conditional class probabilities P (C|X), by marginalizing the class variables C to acquire the normalization constant P (X) from the joint P (X, C), and calculating P (X|C).
The results can be seen in Tab. 1a. We observe that training SLASH with a DNN NPP produces SOTA accuracies compared to DeepProbLog and NeurASP, confirming that SLASH's batch-wise loss computation leads to improved performances. We further observe that the test accuracy of SLASH with a PC NPP is slightly below the other DPPLs, however we argue that this may be since a PC, in comparison to a DNN, is learning a true mixture density rather than just conditional probabilities. The advantages of doing so will be investigated in the next experiments. Note that, optimal architecture search for PCs, e.g. for computer vision, is an open research question.These evaluations show SLASH's advantages on the benchmark MNIST-Addition task. Additional benefits will be made clear in the following experiments.
Evaluation 2: Handling Missing Data with SLASH. SLASH offers the advantage of its flexibility to use various kinds of NPPs. Thus, in comparison to previous DPPLs, one can easily integrate NPPs into SLASH that perform joint probability estimation. For this evaluation, we consider the task of MNIST-Addition with missing data. We trained SLASH (PC) and DeepProbLog with the MNIST-Addition task with images in which a percentage of pixels per image has been removed. It is important to mention here that whereas DeepProbLog handles the missing data simply as background pixels, SLASH (PC) specifically models the missing data as uncertain data by marginalizing the denoted pixels at inference time. We use DeepProbLog here representative of DPPLs without true density estimation. The results can be seen in Tab. 1b for 50%, 80%, 90% and 97% missing pixels per image. We observe that at 50%, DeepProbLog and SLASH produce almost equal accuracies. With 80% percent missing pixels, there is a substantial difference in the ability of the two DPPLs to correctly classify images, with SLASH being very stable. By further increasing the percentage of missing pixels, this difference becomes even more substantial with SLASH still reaching a 82% test accuracy even when 97% of the pixels per image are missing, whereas DeepProbLog degrades to an average of 32% test accuracy. We further note that SLASH, in comparison to DeepProbLog, produces largely reduced standard deviations over runs.
Thus, by utilizing the power of true density estimation SLASH, with an appropriate NPP, can produce more robust results in comparison to other DPPLs. Further, we refer to Appendix D, which contains results of additional experiments where training is performed with the full MNIST data set whereas only the test set entails different rates of missing pixels.
Evaluation 3: Improved Concept Learning via SLASH. We show that SLASH can be very effective for the complex task of set prediction, which previous DPPLs have not tackled. We revert to the ShapeWorld4 data set for this setting. For set prediction, a model is trained to predict the discrete attributes of a set of objects in an image (cf. Fig. for an example ShapeWorld4 image). The difficulty for the model lies therein that it must match an unordered set of corresponding attributes (with varying number of entities over samples) with its internal representations of the image.
The slot attention module introduced by allows for an attractive object-centric approach to this task. Specifically, this module represents a pluggable, differentiable module that can be easily added to any architecture and, through a competitive softmax-based attention mechanism, can enforce the binding of specific parts of a latent representation into permutation-invariant, taskspecific vectors, called slots.
In our experiments, we wish to show that by adding logical constraints to the training setting, one can improve the overall performances and generalization properties of such a model. For this, we train SLASH with NPPs as depicted in Fig. consisting of a shared slot encoder and separate PCs, each modelling the mixture of latent slot variables and the attributes of one category, e.g. color.
For ShapeWorld4, we thereby have altogether four NPPs. SLASH is trained via queries of the kind exemplified in Fig. in the Appendix. We refer to this configuration as SLASH Attention.
We compare SLASH Attention to a baseline slot attention encoder using an MLP and Hungarian loss for predicting the object properties from the slot encodings as in. The results of these experiments can be found in Fig. (top). We observe that the average precision after convergence on the held-out test set with SLASH Attention is greatly improved to that of the baseline model. Additionally, in Fig. we observe that SLASH Attention reaches the average precision value of the baseline model in much fewer number of epochs. Thus, we can summarize that adding logical knowledge in the training procedure via SLASH can greatly improve the capabilities of a neural module for set prediction.
Evaluation 4: Improved Compositional Generalization with SLASH. To test the hypothesis that SLASH Attention possesses improved generalization properties in comparison to the baseline model, we ran experiments on a variant of ShapeWorld4 similar to the CLEVR Compositional Generalization Test (CoGenT). The goal of CoGenT is to investigate a model's ability to handle novel combinations of attributes that were not seen during training.
For this purpose, we established two conditions within a ShapeWorld4 CoGenT data set: Condition (A) -the training and test data set contains squares with the colors gray, blue, brown, or yellow, triangles with the colors red, green, magenta, or cyan and circles of all colors. Condition (B) -the training set is as in Condition (A). However, the test set contains squares with the colors red, green, magenta, or cyan, triangles with the colors gray, blue, brown, or yellow and circles of all colors. The goal is to investigate how well a model can generalize that, e.g., also squares can have the color red, although never having seen evidence for this during training.
The resulting average precision test scores are presented in Fig. (bottom). We observe that, even though the SLASH Program used for this experiment was not explicitly written to handle composition generalization, SLASH Attention shows greatly improved generalization capabilities. This can be seen in the approx. 13% higher average precision scores on the Condition (B) test set in comparison to the baseline model. Importantly, this trend still holds even when subtracting the higher precision scores observed in Condition (A).
To summarize our findings from the experiments on set prediction: we observe that adding prior knowledge in the form of logical constraints via SLASH can greatly improve a neural module in terms of performance and generalizability. On a side note: training neural networks for novel tasks, often involves defining explicit loss functions, e.g. Hungarian loss for set prediction. In contrast with SLASH, no matter the choice of NPP and underlying task, the training loss remains the same. Task-related requirements simply need to be added as lines of code to the SLASH program. This additionally highlights SLASH's versatility and flexibility.
In this work, we largely make use of NPPs that contain probabilistic circuits (specifically SPNs) which allow for tractable density estimation and modelling of joint probabilities. In this way, it is possible to answer a much richer set of probabilistic queries, i.e. P(X, C), P(X|C) and P(C|X). In addition to this, we introduce the arguably more interesting type of NPP that combines a neural module with a PC. Hereby, the neural module learns to map the raw input data into an optimal latent representation, e.g. object-based slot representations. The PC, in turn, learns to model the joint distribution of these latent variables and produces the final probability estimates. This type of NPP nicely combines the representational power of neural networks with the advantages of PCs in probability estimation and query flexibility.****We further observe that the test accuracy of SLASH with a PC NPP is slightly below the other DPPLs, however we argue that this may be since a PC, in comparison to a DNN, is learning a true mixture density rather than just conditional probabilities.****For those experiments using NPPs with PC we have used Einsum Networks (EiNets) for implementing the probabilistic circuits. EiNets are a novel implementation design for SPNs introduced by Peharz et al. (2020) that minimize the issue of computational costs that initial SPNs had suffered. This is accomplished by combining several arithmetic operations via a single monolithic einsum-operation.****These results indicate that the conclusions, drawn in the main part of our work, remain true also in this setting of handling missing data. | What are the semantics of SLASH and the actual benefits of embracing PCs into the framework? | We are happy to hear our work to be well written. While the benefits of Neuro-Symbolic approaches are becoming more well known recently (but are not as apparent in all parts of the AI community), we emphasize the ability of SLASH to combine neural, logical and truly tractable probabilistic modules within a unifying framework and via a single optimization procedure. We believe we have presented the benefits of PCs in Probabilistic Logic Programming over several claims. Firstly, via NPPs with PCs we have a much larger set of probabilistic queries, such as P(X, C) and P(C|X) in addition to P(X|C), see pp. 4 - 5. Particularly, we emphasize that neural predicates as used in previous works can in principle only handle queries of the form P(X|C), p. 4. With our results on computation time, we wished to show that this advantage of tractable probabilistic inference comes at the same or fewer costs as previous DPPLs, p. 7. Next, with our experiments on missing data, we explicitly show the practical advantages of PCs, pp. 7-8. Specifically, simply by using the marginalization abilities of PCs, we can handle the case of missing data out of the box. Lastly, with PC we are getting a generative model, which comes in handy in case of small data sets and/or image generation. |
null | false | null | Which is a species of fish? Tarp or Carp | Carp |
|
null | false | 200 | Recent studies have shown the vulnerability of ML models to adversarial attacks, small perturbations which lead to misclassification of inputs. Adversarial example generation in NLP BIBREF0 is more challenging than in common computer vision tasks BIBREF1, BIBREF2, BIBREF3 due to two main reasons: the discrete nature of input space and ensuring semantic coherence with the original sentence. A major bottleneck in applying gradient based BIBREF4 or generator model BIBREF5 based approaches to generate adversarial examples in NLP is the backward propagation of the perturbations from the continuous embedding space to the discrete token space.
Recent works for attacking text models rely on introducing errors at the character level in words BIBREF6, BIBREF7 or adding and deleting words BIBREF8, BIBREF9, BIBREF10, etc. for creating adversarial examples. These techniques often result in adversarial examples which are unnatural looking and lack grammatical correctness, and thus can be easily identified by humans.
TextFooler BIBREF11 is a black-box attack, that uses rule based synonym replacement from a fixed word embedding space to generate adversarial examples. These adversarial examples do not account for the overall semantics of the sentence, and consider only the token level similarity using word embeddings. This can lead to out-of-context and unnaturally complex replacements (see Table ), which can be easily identifiable by humans.
The recent advent of powerful language models BIBREF12, BIBREF13 in NLP has paved the way for using them in various downstream applications. In this paper, we present a simple yet novel technique: BAE (BERT-based Adversarial Examples), which uses a language model (LM) for token replacement to best fit the overall context. We perturb an input sentence by either replacing a token or inserting a new token in the sentence, by means of masking a part of the input and using a LM to fill in the mask (See Figure FIGREF1). BAE relies on the powerful BERT masked LM for ensuring grammatical correctness of the adversarial examples. Our attack beats the previous baselines by a large margin and confirms the inherent vulnerabilities of modern text classification models to adversarial attacks. Moreover, BAE produces more richer and natural looking adversarial examples as it uses the semantics learned by a LM.
To the best of our knowledge, we are the first to use a LM for adversarial example generation. We summarize our major contributions as follows:
We propose BAE, a novel strategy for generating natural looking adversarial examples using a masked language model.
We introduce 4 BAE attack modes, all of which are almost always stronger than previous baselines on 7 text classification datasets.
We show that, surprisingly, just a few replace/insert operations can reduce the accuracy of even a powerful BERT-based classifier by over $80\%$ on some datasets.
Through human evaluation, we show that BAE yields adversarial examples with improved grammaticality and semantic coherence. | Does BAE yield adversarial examples with improved grammaticality and semantic coherence? | Yes, it does. |
null | false | 483 | Using comments during retrieval further improves the results, confirming that comments are a useful modality and that our approach can successfully leverage this auxiliary information. | To be comparable, pre-training datasets should be the same for all models, either on HowTo100M or CLIP, or both. So whether the improvement is coming from the novel model or from more informative pre-training datasets? | We agree that evaluation in the video-text retrieval is currently less than ideal. Many models aren’t directly comparable and pretraining is not yet standardized. As many current methods are very expensive to train, we cannot level the field in this paper. However, our main goal here was not to directly compare to these methods, but to showcase how the noisy “comment-modality” is a valuable resource and can be used to learn better representations (since our models trained with and without comments are directly comparable to each other) while including the other methods to provide context. |
null | false | null | Which is a bird or fish: Red-crested pochard, Redfish | Redfish is a fish, Red-crested pochard is a bird |
|
null | false | null | What is Python? | Python is a programming language that has a design emphasis on code readability and is strongly typed. You can interact with this programming language through writing 'python code' using proper coding syntax and semantics. The code that is written can be used to build tools, applications, websites, and streaming services. You can also automate tasks with python, analyze data, and perform machine learning and data science. |
|
2004.03762 | false | null | We automatically evaluate on four different types of interpolations (where different combinations of sentences are removed and the model is forced to regenerate them), We evaluate the generations with the ROUGE BIBREF29 and METEOR BIBREF30 metrics using the true sentences as targets. Table TABREF33 shows the automatic evaluation results from interpolations using our proposed models and baselines. The #Sent(s) column indicates which sentence(s) were removed, and then regenerated by the model. We gave the baselines a slight edge over SLDS because they pick the best out of 1000 samples while SLDS is only out of 50. The SLDS models see their largest gain over the baseline models when at least the first sentence is given as an input. The baseline models do better when the first and second sentence need to be imputed. This is likely due to the fact that having access to the earlier sentences allows a better initialization for the Gibbs sampler. Surprisingly, the semi-supervised variants of the SLDS models achieve higher scores. The reasons for this is discussed below in the Perplexity section.
We evaluate the generations with the ROUGE BIBREF29 and METEOR BIBREF30 metrics using the true sentences as targets. | What metrics are used for evaluation? | The answers are shown as follows:
* ROUGE BIBREF29 and METEOR BIBREF30
|
null | false | null | how do you rate strength of tornado | There are several scales for rating the strength of tornadoes. The Fujita scale rates tornadoes by damage caused and has been replaced in some countries by the updated Enhanced Fujita Scale. An F0 or EF0 tornado, the weakest category, damages trees, but not substantial structures. An F5 or EF5 tornado, the strongest category, rips buildings off their foundations and can deform large skyscrapers.
The similar TORRO scale ranges from T0 for extremely weak tornadoes to T11 for the most powerful known tornadoes.
Doppler radar data, photogrammetry, and ground swirl patterns (trochoidal marks) may also be analyzed to determine intensity and assign a rating.
source : https://en.wikipedia.org/wiki/Tornado |
|
null | false | null | What are all of the professional baseball teams in the state of Florida? | The professional baseball teams in the state of Florida are the Tampa Bay Rays and Miami Marlins. |
|
1808.03815 | false | null | Typically, to predict and label arguments for a given predicate, a role classifier is employed on top of the BiLSTM encoder. Some work like BIBREF17 shows that incorporating the predicate's hidden state in their role classifier enhances the model performance, while we argue that a more natural way to incorporate the syntactic information carried by the predicate is to employ the attentional mechanism. Our model adopts the recently introduced biaffine attention BIBREF14 to enhance our role scorer. Biaffine attention is a natural extension of bilinear attention BIBREF18 which is widely used in neural machine translation (NMT).
Our model adopts the recently introduced biaffine attention BIBREF14 to enhance our role scorer. | What is the biaffine scorer? | The answers are shown as follows:
* biaffine attention BIBREF14
|
null | false | null | Scarlet Days is a 1919 American silent Western film produced and directed by D. W. Griffith and released through Paramount/Artcraft Pictures, Artcraft being an affiliate of Paramount. Richard Barthelmess stars in a role for which Griffith had screentested Rudolph Valentino. In today's time, this film is considered by many to be one of Griffith's worst films though it might have worked better as a short film. This film was unlike others created by D.W. Griffith. According to an article written for The Cincinnati Enquirer, written on the 16 of November 1919: "Unlike other recent Griffith production, Scarlet Days is a story of the old West, of the gold rush days of 49- Bret Harte transferred to the screen!" The Western film genre was expanding at this time and Scarlet Days fits into this category. Western films were popular for this time. Considered a lost film, a print was found in the State Film Archives of the Soviet Union, which donated it to the Museum of Modern Art in 1969. The film was screened to the public, for the first time since its rediscovery, on March 24 and 25, 1969. Scarlet Days was shown alongside another missing film, A Romance of Happy Valley, at the Museum of Modern Art's auditorium with titles still printed in Russian. English titles were later created by the Department of Film at the Museum of Modern Art. | How is the movie Scarlet Days? | Scarlet Days is a 1919 American silent Western film produced and directed by D.W. Griffith and released through Paramount/Artcraft Pictures. It is considered one of Griffith's worst films, but it is a story of the old West. A print was found in the State Film Archives of the Soviet Union and donated to the Museum of Modern Art in 1969. Titles were later created by the Department of Film at the Museum. |
null | false | null | Identify which instrument is string or percussion: Wind chime, Hardingfele | Hardingfele is string, Wind chime is percussion. |
|
null | false | null | Do I need to make my bed everyday? | Making your bed everyday is very important for a number of reasons. It will help keep bed bugs at bay. When dust mites get exposed to light they also die. When you do not make your bed or spread out your bed linens it creates areas on the bed where warmth and moisture can occur, and this is a great place where bed bugs and dust mites can thrive. |
|
null | false | null | Which city will host the 2024 Summer Olympics? | The 2024 Summer Olympics will be hosted mainly at Paris from 26 July to 11 August 2024, with other 16 cities spread across Metropolitan France and one in Tahiti as subsites. |
|
null | false | null | Associative reinforcement learning
Associative reinforcement learning tasks combine facets of stochastic learning automata tasks and supervised learning pattern classification tasks. In associative reinforcement learning tasks, the learning system interacts in a closed loop with its environment.
Deep reinforcement learning
This approach extends reinforcement learning by using a deep neural network and without explicitly designing the state space. The work on learning ATARI games by Google DeepMind increased attention to deep reinforcement learning or end-to-end reinforcement learning.
Adversarial deep reinforcement learning
Adversarial deep reinforcement learning is an active area of research in reinforcement learning focusing on vulnerabilities of learned policies. In this research area some studies initially showed that reinforcement learning policies are susceptible to imperceptible adversarial manipulations. While some methods have been proposed to overcome these susceptibilities, in the most recent studies it has been shown that these proposed solutions are far from providing an accurate representation of current vulnerabilities of deep reinforcement learning policies.
Fuzzy reinforcement learning
By introducing fuzzy inference in RL, approximating the state-action value function with fuzzy rules in continuous space becomes possible. The IF - THEN form of fuzzy rules make this approach suitable for expressing the results in a form close to natural language. Extending FRL with Fuzzy Rule Interpolation allows the use of reduced size sparse fuzzy rule-bases to emphasize cardinal rules (most important state-action values).
Inverse reinforcement learning
In inverse reinforcement learning (IRL), no reward function is given. Instead, the reward function is inferred given an observed behavior from an expert. The idea is to mimic observed behavior, which is often optimal or close to optimal.
Safe reinforcement learning
Safe reinforcement learning (SRL) can be defined as the process of learning policies that maximize the expectation of the return in problems in which it is important to ensure reasonable system performance and/or respect safety constraints during the learning and/or deployment processes. | Please describe the broad classes of Reinforcement Learning (RL) within Machine Learning and what problems they attempt to address. | Reinforcement Learning is an area of Machine Learning that works in a problem space or environment where the agent or Reinforcement Learning algorithm, uses an understanding of the past test results and the potential paths through the environment to calculate a policy that solves for the goal of a maximized reward.
Given the variety of environments that RL can operate in, there are a range of approaches that the agent can take and are broadly categarized as six different approaches:
1. Associated RL - which combines traditional approaches of Machine Learning classification with the automated learning on random, or stochastically distributed results.
2. Deep RL - which leverages traditional deep neural networks but allows for flexibility and a lack of pre-defined layers and the state space.
3. Adversarial Deep RL - where end results are compared using vulnerabilities in state stored in Deep Neural Networks and attempts to find weaknesses in these models.
4. Fuzzy RL - which leverages a near natural language definition of if-then fuzzy rules which determine the calculation of the value of a result.
5. Inverse RL - which removes the common reward function evaluation part of the model and instead injects a human to observe and evalutate the results.
6. Safe RL - is an approach to defining policies that also takes into account the behavior of the agent in the envionment, such that those agent steps are reasonaable or respect safety constraints. |
null | false | null | When did Hungary gain its independence? | Hungary declared independence on October 17, 1918 and officially formed its government on November 1 of that year. |
|
null | false | null | Which is a species of fish? Mola mola or Molar | Mola mola |
|
null | false | 209 | We have created the HispaBlogs dataset by collecting posts from Spanish blogs from five different countries: Argentina, Chile, Mexico, Peru and Spain. For each country, there are 450 and 200 blogs respectively for training and test, ensuring that each author appears only in one set. Each blog contains at least 10 posts. The total number of blogs is 2,250 and 1,000 respectively. Statistics of the number of words are shown in Table 3 .
We have created the HispaBlogs dataset7 by collecting posts from Spanish blogs from five different countries: Argentina, Chile, Mexico, Peru and Spain. | How to create the HispaBlogs dataset? | They created the HispaBlogs dataset by collecting posts from Spanish blogs from five different countries: Argentina, Chile, Mexico, Peru and Spain. |
2004.00139 | true | null | To increase the benefits of our data for ASR systems, we also trained a grapheme-to-phoneme (g2p) model: Out-of-vocabulary words can be a problem for ASR system. For those out-of-vocabulary words we need a model that can generate pronunciations from a written form, in real time. This is why we train a grapheme-to-phoneme (g2p) model that generates a sequence of phonemes for a given word. We train the g2p model using our dictionary and compare its performance with a widely used joint-sequence g2p model, Sequitur BIBREF26. For the g2p model we are using the same architecture as for the p2g model. The only difference is input and output vocabulary. The Sequitur and our model are using the dictionary with the same train (19'898 samples), test (2'412 samples) and validation (2'212 samples) split. Additionally, we also test their performance only on the items from the Zurich and Visp dialect, because most of the samples are from this two dialects. In Table TABREF15 we show the result of the comparison of the two models. We compute the edit distance between the predicted and the true pronunciation and report the number of exact matches. In the first columns we have the result using the whole test set with all the dialects, and in the 2nd and 3rd columns we show the number of exact matches only on the samples from the test set that are from the Zurich and Visp dialect. For here we can clearly see that our model performs better than the Sequitur model. The reason why we have less matches in the Visp dialect compared to Zurich is because most of the our data is from the Zurich dialect.
Additionally, we also test their performance only on the items from the Zurich and Visp dialect, because most of the samples are from this two dialects. In Table TABREF15 we show the result of the comparison of the two models. | Is the model evaluated on the graphemes-to-phonemes task? | Yes. |
2003.05377 | false | null | FLOAT SELECTED: Table 1: The number of songs and artists by genre
From the Vagalume's music web page, we collect the song title and lyrics, and the artist name. The genre was collected from the page of styles, which lists all the musical genres and, for each one, all the artists. We selected only 14 genres that we consider as representative Brazilian music, shown in Table TABREF8. Figure FIGREF6 presents an example of the Vagalume's music Web page with the song “Como é grande o meu amor por você”, of the Brazilian singer Roberto Carlos. Green boxes indicate information about music that can be extracted directly from the web page. From this information, the language in which the lyrics are available can be obtained by looking at the icon indicating the flag of Brazil preceded by the “Original” word.
FLOAT SELECTED: Table 1: The number of songs and artists by genre
We selected only 14 genres that we consider as representative Brazilian music, shown in Table TABREF8. | what genres do they songs fall under? | Gospel, Sertanejo, MPB, Forró, Pagode, Rock, Samba, Pop, Axé, Funk-carioca, Infantil, Velha-guarda, Bossa-nova and Jovem-guarda |
null | false | 135 | As a starting point for our pilots, we made use of texts from the InScript corpus BIBREF10 , which provides stories centered around everyday situations (see Section SECREF7 ). We conducted three different pilot studies to determine the best way of collecting questions that require inference over commonsense knowledge:
The most intuitive way of collecting reading comprehension questions is to show texts to workers and let them formulate questions and answers on the texts, which is what we tried internally in a first pilot. Since our focus is to provide an evaluation framework for inference over commonsense knowledge, we manually assessed the number of questions that indeed require common sense knowledge. We found too many questions and answers collected in this manner to be lexically close to the text.
In a second pilot, we investigated the option to take the questions collected for one text and show them as questions for another text of the same scenario. While this method resulted in a larger number of questions that required inference, we found the majority of questions to not make sense at all when paired with another text. Many questions were specific to a text (and not to a scenario), requiring details that could not be answered from other texts. Since the two previous pilot setups resulted in questions that centered around the texts themselves, we decided for a third pilot to not show workers any specific texts at all. Instead, we asked for questions that centered around a specific script scenario (e.g. eating in a restaurant). We found this mode of collection to result in questions that have the right level of specificity for our purposes: namely, questions that are related to a scenario and that can be answered from different texts (about that scenario), but for which a text does not need to provide the answer explicitly.
The next section will describe the mode of collection chosen for the final dataset, based on the third pilot, in more detail.
The most intuitive way of collecting reading comprehension questions is to show texts to workers and let them formulate questions and answers on the texts, which is what we tried internally in a first pilot. | What is the most intuitive way of collecting reading comprehension questions? | The most intuitive way of collecting reading comprehension questions is to show texts to workers and let them formulate questions and answers on the texts, which is what they tried internally in a first pilot. |
null | false | null | Does spark provide date & time when a particular event is received in the structured streaming? | Not directly. The source of the data is responsible for event time, the “arrival” time is typically the timestamp from the message server source you are reading from (Kafka, Kinesis, etc.) and the “processing” time can be inserted when processing the incomming data stream using the functions package current_timestamp function. https://spark.apache.org/docs/latest/api/scala/index.html#org.apache.spark.sql.functions$ This isn’t exactly when Spark received the data from the server but is close and will be grouped by trigger interval. Note that this function can not be used when in continuous mode. https://spark.apache.org/docs/latest/structured-streaming-programming-guide.html#supported-queries |
|
null | false | null | What is Davis, CA known for? | Being the bike capitol of the U.S. as well as one of the top agriculture cities. UC Davis has bike paths everywhere across its campus - more so than streets for cars! Additionally, there is a consistent smell of manure throughout the campus due to the high number of cows. It's a very nice city to bike around as well as grow a garden in! |
|
1911.09845 | false | null | To evaluate the responses generated by all compared methods, we compute the following automatic metrics on our test set:
BLEU: BLEU-n measures the average n-gram precision on a set of reference responses. We report BLEU-n with n=1,2,3,4.
Distinct-1 & distinct-2 BIBREF5: We count the numbers of distinct uni-grams and bi-grams in the generated responses and divide the numbers by the total number of generated uni-grams and bi-grams in the test set. These metrics can be regarded as an automatic metric to evaluate the diversity of the responses.
To evaluate the responses generated by all compared methods, we compute the following automatic metrics on our test set:
BLEU: BLEU-n measures the average n-gram precision on a set of reference responses. We report BLEU-n with n=1,2,3,4.
Distinct-1 & distinct-2 BIBREF5: We count the numbers of distinct uni-grams and bi-grams in the generated responses and divide the numbers by the total number of generated uni-grams and bi-grams in the test set. These metrics can be regarded as an automatic metric to evaluate the diversity of the responses. | What automatic metrics are used? | The answers are shown as follows:
* BLEU
* Distinct-1 & distinct-2
|
null | false | null | Summary
Games Year Events Best Nation
1 1896 4 Hungary
2 1900 7 Great Britain
3 1904 9 Germany
4 1908 6 Great Britain
5 1912 9 Germany
6
7 1920 10 United States
8 1924 11 United States
9 1928 11 United States
10 1932 11 Japan
11 1936 11 Japan
12
13
14 1948 11 United States
15 1952 11 United States
16 1956 13 Australia
Games Year Events Best Nation
17 1960 15 United States
18 1964 18 United States
19 1968 29 United States
20 1972 29 United States
21 1976 26 United States
22 1980 26 East Germany
23 1984 29 United States
24 1988 31 East Germany
25 1992 31 United States
26 1996 32 United States
27 2000 32 United States
28 2004 32 United States
29 2008 34 United States
30 2012 34 United States
31 2016 34 United States
32 2020 37 United States | Which country has topped the swimming medals list in the summer olympics? | United States |
1908.10449 | false | null | We build the iSQuAD and iNewsQA datasets based on SQuAD v1.1 BIBREF0 and NewsQA BIBREF1. Both original datasets share similar properties. Specifically, every data-point consists of a tuple, $\lbrace p, q, a\rbrace $, where $p$ represents a paragraph, $q$ a question, and $a$ is the answer. The answer is a word span defined by head and tail positions in $p$. NewsQA is more difficult than SQuAD because it has a larger vocabulary, more difficult questions, and longer source documents.
iMRC: Making MRC Interactive ::: Evaluation Metric
Since iMRC involves both MRC and RL, we adopt evaluation metrics from both settings. First, as a question answering task, we use $\text{F}_1$ score to compare predicted answers against ground-truth, as in previous works. When there exist multiple ground-truth answers, we report the max $\text{F}_1$ score. Second, mastering multiple games remains quite challenging for RL agents. Therefore, we evaluate an agent's performance during both its training and testing phases. During training, we report training curves averaged over 3 random seeds. During test, we follow common practice in supervised learning tasks where we report the agent's test performance corresponding to its best validation performance .
We build the iSQuAD and iNewsQA datasets based on SQuAD v1.1 BIBREF0 and NewsQA BIBREF1.
iMRC: Making MRC Interactive ::: Evaluation Metric
Since iMRC involves both MRC and RL, we adopt evaluation metrics from both settings. First, as a question answering task, we use $\text{F}_1$ score to compare predicted answers against ground-truth, as in previous works. When there exist multiple ground-truth answers, we report the max $\text{F}_1$ score. Second, mastering multiple games remains quite challenging for RL agents. Therefore, we evaluate an agent's performance during both its training and testing phases. During training, we report training curves averaged over 3 random seeds. During test, we follow common practice in supervised learning tasks where we report the agent's test performance corresponding to its best validation performance . | What are the models evaluated on? | They evaluate F1 score and agent's test performance on their own built interactive datasets (iSQuAD and iNewsQA) |
null | false | null | what is a computer system? | A computer system is a nominally complete computer that includes the hardware, operating system (main software), and peripheral equipment needed and used for full operation. This term may also refer to a group of computers that are linked and function together, such as a computer network or computer cluster. |
|
null | false | null | Scarlett Ingrid Johansson (/dʒoʊˈhænsən/; born November 22, 1984) is an American actress. The world's highest-paid actress in 2018 and 2019, she has featured multiple times on the Forbes Celebrity 100 list. Time named her one of the 100 most influential people in the world in 2021. Her films have grossed over $14.3 billion worldwide, making Johansson the highest-grossing box office star of all time. She has received various accolades, including a Tony Award and a British Academy Film Award, in addition to nominations for two Academy Awards and five Golden Globe Awards. | When was Scarlett Johansson born? | Scarlett Johansson was born November 22, 1984. |
null | false | null | Provide instructions on making a peanut butter and jelly sandwich | Making the perfect peanut butter and jelly sandwich is part art and part science. One must must be judicious in the quantity of ingredients while ensuring proper coverage to ensure each bite is equal to the one before and after it. Here, both measurement and distribution of ingredients serves as the scientific aspect of constructing your delicious and timeless lunchtime meal. Jars of peanut utter suggest a serving size as 2 tablespoons, which, when applied to the critical soft white bread, may prove too much to hold and cause soak through or tearing of the bread. We think 1.5 tablespoons is the proper amount to ensure there is coverage of the bread slice to the crust without compromising the integrity of the bread. Jelly application is even more precarious, as it holds more fluid and can soak through the bread quicker; equally, jelly has a tendency to congeal. To ensure proper application, heat the jar of jelly closer to room temperature if refrigerated, then apply using a clean butter knife or small spatula. Again, one must ensure it reaches to the crust and is evenly distributed. Once you have complete coverage of peanut butter on one slice and jelly on the other, place the peanut butter side of the first slice on top of the upward-facing jelly slice. This ensure none of the jelly slips or drips. Once together, gently compress the sandwich. Slice delicately from one corner to the opposite corner. Gently press the crust edge of each triangle, and plate for serving. |
|
null | false | 104 | Since the CoNLL annotations have 21 semantic roles in total, we use 21 roles in our model as well as the baseline. Following garg2012unsupervised, we set the number of PRs to 2 (excluding INLINEFORM0 , INLINEFORM1 and INLINEFORM2 ), and SRs to 21-2=19. Table TABREF27 shows the results.
In the first setting (Line 1), we train and test the monolingual model on the CoNLL data. We observe significant improvements in F1 score over the Baseline (Line 0) in both languages. Using the CoNLL 2009 dataset alone, titovcrosslingual report an F1 score of 80.9% (PU=86.8%, CO=75.7%) for German. Thus, our monolingual model outperforms their monolingual model in German. For English, they report an F1 score of 83.6% (PU=87.5%, CO=80.1%), but note that our English results are not directly comparable to theirs due to differences argument identification, as discussed in section SECREF25 . As their argument identification score is lower, perhaps their system is discarding “difficult” arguments which leads to a higher clustering score.
In the second setting (Line 2), we use the additional monolingual Europarl (EP) data for training. We get equivalent results in English and a significant improvement in German compared to our previous setting (Line 1). The German dataset in CoNLL is quite small and benefits from the additional EP training data. In contrast, the English model is already quite good due to a relatively big dataset from CoNLL, and good accuracy syntactic parsers. Unfortunately, titovcrosslingual do not report results with this setting.
The third setting (Line 3) gives the results of our multilingual model, which adds the word alignments in the EP data. Comparing with Line 2, we get non-significant improvements in both languages. titovcrosslingual obtain an F1 score of 82.7% (PU=85.0%, CO=80.6%) for German, and 83.7% (PU=86.8%, CO=80.7%) for English. Thus, for German, our multilingual Bayesian model is able to capture the cross-lingual patterns at least as well as the external penalty term in BIBREF6 . We cannot compare the English results unfortunately due to differences in argument identification.
We also compared monolingual and bilingual training data using a setting that emulates the standard supervised setup of separate training and test data sets. We train only on the EP dataset and test on the CoNLL dataset. Lines 4 and 5 of Table TABREF27 give the results. The multilingual model obtains small improvements in both languages, which confirms the results from the standard unsupervised setup, comparing lines 2 to 3.
These results indicate that little information can be learned about semantic roles from this parallel data setup. One possible explanation for this result is that the setup itself is inadequate. Given the definition of aligned arguments, only 8% of English arguments and 17% of German arguments are aligned. This plus our experiments suggest that improving the alignment model is a necessary step to making effective use of parallel data in multilingual SRI, for example by joint modeling with SRI. We leave this exploration to future work.
We cannot compare the English results unfortunately due to differences in argument identification. | Why do the English results cannot be compared? | Due to differences in argument identification. |
1906.10551 | false | null | As a basis for our experiments, we reimplemented 12 existing AV approaches, which have shown their potentials in the previous PAN-AV competitions BIBREF11 , BIBREF12 as well as in a number of AV studies. The methods are listed in Table TABREF33 together with their classifications regarding the AV characteristics, which we proposed in Section SECREF3 .
FLOAT SELECTED: Table 2: All 12 AVmethods, classified according to their properties.
The methods are listed in Table TABREF33 together with their classifications regarding the AV characteristics, which we proposed in Section SECREF3 .
FLOAT SELECTED: Table 2: All 12 AVmethods, classified according to their properties. | What are the 12 AV approaches which are examined? | MOCC, OCCAV, COAV, AVeer, GLAD, DistAV, Unmasking, Caravel, GenIM, ImpGI, SPATIUM and NNCD |
null | false | 24 | Reading Comprehension (RC) has become a central task in natural language processing, with great practical value in various industries. In recent years, many large-scale RC datasets in English BIBREF0, BIBREF1, BIBREF2, BIBREF3, BIBREF4, BIBREF5, BIBREF6 have nourished the development of numerous powerful and diverse RC models BIBREF7, BIBREF8, BIBREF9, BIBREF10, BIBREF11. The state-of-the-art model BIBREF12 on SQuAD, one of the most widely used RC benchmarks, even surpasses human-level performance. Nonetheless, RC on languages other than English has been limited due to the absence of sufficient training data. Although some efforts have been made to create RC datasets for Chinese BIBREF13, BIBREF14 and Korean BIBREF15, it is not feasible to collect RC datasets for every language since annotation efforts to collect a new RC dataset are often far from trivial. Therefore, the setup of transfer learning, especially zero-shot learning, is of extraordinary importance.
Existing methods BIBREF16 of cross-lingual transfer learning on RC datasets often count on machine translation (MT) to translate data from source language into target language, or vice versa. These methods may not require a well-annotated RC dataset for the target language, whereas a high-quality MT model is needed as a trade-off, which might not be available when it comes to low-resource languages.
In this paper, we leverage pre-trained multilingual language representation, for example, BERT learned from multilingual un-annotated sentences (multi-BERT), in cross-lingual zero-shot RC. We fine-tune multi-BERT on the training set in source language, then test the model in target language, with a number of combinations of source-target language pair to explore the cross-lingual ability of multi-BERT. Surprisingly, we find that the models have the ability to transfer between low lexical similarity language pair, such as English and Chinese. Recent studies BIBREF17, BIBREF12, BIBREF18 show that cross-lingual language models have the ability to enable preliminary zero-shot transfer on simple natural language understanding tasks, but zero-shot transfer of RC has not been studied. To our knowledge, this is the first work systematically exploring the cross-lingual transferring ability of multi-BERT on RC tasks.
Reading Comprehension (RC) has become a central task in natural language processing, with great practical value in various industries. In recent years, many large-scale RC datasets in English BIBREF0, BIBREF1, BIBREF2, BIBREF3, BIBREF4, BIBREF5, BIBREF6 have nourished the development of numerous powerful and diverse RC models BIBREF7, BIBREF8, BIBREF9, BIBREF10, BIBREF11. The state-of-the-art model BIBREF12 on SQuAD, one of the most widely used RC benchmarks, even surpasses human-level performance. Nonetheless, RC on languages other than English has been limited due to the absence of sufficient training data. Although some efforts have been made to create RC datasets for Chinese BIBREF13, BIBREF14 and Korean BIBREF15, it is not feasible to collect RC datasets for every language since annotation efforts to collect a new RC dataset are often far from trivial. Therefore, the setup of transfer learning, especially zero-shot learning, is of extraordinary importance.
Existing methods BIBREF16 of cross-lingual transfer learning on RC datasets often count on machine translation (MT) to translate data from source language into target language, or vice versa. These methods may not require a well-annotated RC dataset for the target language, whereas a high-quality MT model is needed as a trade-off, which might not be available when it comes to low-resource languages.
In this paper, we leverage pre-trained multilingual language representation, for example, BERT learned from multilingual un-annotated sentences (multi-BERT), in cross-lingual zero-shot RC. We fine-tune multi-BERT on the training set in source language, then test the model in target language, with a number of combinations of source-target language pair to explore the cross-lingual ability of multi-BERT. Surprisingly, we find that the models have the ability to transfer between low lexical similarity language pair, such as English and Chinese. Recent studies BIBREF17, BIBREF12, BIBREF18 show that cross-lingual language models have the ability to enable preliminary zero-shot transfer on simple natural language understanding tasks, but zero-shot transfer of RC has not been studied. To our knowledge, this is the first work systematically exploring the cross-lingual transferring ability of multi-BERT on RC tasks.
Reading Comprehension (RC) has become a central task in natural language processing, with great practical value in various industries.****Surprisingly, we find that the models have the ability to transfer between low lexical similarity language pair, such as English and Chinese. Recent studies (Lample and Conneau, 2019; Devlin et al., 2018; Wu and Dredze, 2019) show that cross-lingual language models have the ability to enable preliminary zero-shot transfer on simple natural language understanding tasks, but zero-shot transfer of RC has not been studied. To our knowledge, this is the first work systematically exploring the cross-lingual transferring ability of multi-BERT on RC tasks. | What field do they research in their paper? | Reading Comprehension. |
1811.02076 | false | null | We evaluate a multi-task approach and three algorithms that explicitly model the task dependencies. We perform experiments on document-level variants of the SQuAD dataset BIBREF1 . The contributions for our papers are:
We perform experiments on document-level variants of the SQuAD dataset BIBREF1 . | What datasets have this method been evaluated on? | The answers are shown as follows:
* document-level variants of the SQuAD dataset
|
null | false | null | Name some famous teams in the Premier League | Manchester United, Manchester City, Arsenal, Liverpool, Chelsea, Hotspur, etc |
|
null | false | null | Jeremiah Mutwalante Twa-Twa (born 29 January 1951) is a Ugandan physician, public health specialist, and politician. His last name is sometimes spelled "Twatwa". He is the elected Member of Parliament representing Iki-Iki County, Budaka District. He has served in that capacity since May 2011. Before that, he served as the director of the Division of Child Health at the Uganda Ministry of Health from 2005 until 2010.
Background and education
He was born in Budaka District on 29 January 1951. Twa-Twa attended Sekulo Primary School in Budaka District from 1959 until 1964. He then studied at Mbale College in Mbale District from 1965 until 1966. He undertook his O-Level studies (S1-S4) at Jinja College in Jinja District from 1967 until 1970. In 1971, he entered Namilyango College in Mukono District to pursue his A-Level (S5-S6) education, graduating in 1972.
He graduated from Makerere University Medical School in 1978 with a Bachelor of Medicine and Bachelor of Surgery degree. In 1981, he obtained a Diploma in Public Health from the Makerere University School of Public Health. In 1995, he received a Master of Arts degree in demography from the same university. He also holds a Certificate in Epidemiology, awarded in 1983 by the University of Nairobi.[ | Given a reference text about Jeremiah Mutwalante Twa-Twa, when and where was he born? | Jeremiah Mutwalante Twa-Twa was born in the Budaka District of Uganda on January 29, 1951. |
null | false | null | U.S. News & World Report ranks UVA tied for 25th among national universities overall, 3rd among public universities and tied for 110th among global universities in its 2022 report, 8th best business program and 3rd best business management program. | According to this paragraph, what is UVA's overall ranking among national universities? | UVA tied for 25th among national universities. |
null | false | null | Pets provide their owners, or guardians, both physical and emotional benefits. Walking a dog can provide both the human and the dog with exercise, fresh air, and social interaction. Pets can give companionship to people who are living alone or elderly adults who do not have adequate social interaction with other people. There is a medically approved class of therapy animals that are brought to visit confined humans, such as children in hospitals or elders in nursing homes. Pet therapy utilizes trained animals and handlers to achieve specific physical, social, cognitive, or emotional goals with patients. | Summarise from the given text, Why do people like pets? | 1. Pets benefit their owners/guardians both physically and emotionally.
2. Talking a dog for a stroll helps both human and dog through physical exercise, inhaling fresh air and social interaction.
3. For people who live alone, aged people in nursery homes, children in hospitals etc who don't interact with other people, There is a class of medically approved therapy animals that are brought to visit to gain specific type of physical, social, cognitive, or emotional benefits |
null | false | 103 | As character and word-piece inputs become commonplace in modern NLP pipelines, it is worth highlighting the vulnerability they add. We show that minimally-doctored attacks can bring down accuracy of classifiers to random guessing. We recommend word recognition as a safeguard against this and build upon RNN-based semi-character word recognizers. We discover that when used as a defense mechanism, the most accurate word recognition models are not always the most robust against adversarial attacks. Additionally, we highlight the need to control the sensitivity of these models to achieve high robustness.
We discover that when used as a defense mechanism, the most accurate word recognition models are not always the most robust against adversarial attacks. | Are the most accurate word recognition models always the most robust against adversarial attacks when used as a defense mechanism? | No. |
null | false | 245 | The goal of sentence simplification is to compose complex sentences into simpler ones so that they are more comprehensible and accessible, while still retaining the original information content and meaning. Sentence simplification has a number of practical applications. On one hand, it provides reading aids for people with limited language proficiency BIBREF1 , BIBREF2 , or for patients with linguistic and cognitive disabilities BIBREF3 . On the other hand, it can improve the performance of other NLP tasks BIBREF4 , BIBREF5 , BIBREF6 . Prior work has explored monolingual machine translation (MT) approaches, utilizing corpora of simplified texts, e.g., Simple English Wikipedia (SEW), and making use of statistical MT models, such as phrase-based MT (PBMT) BIBREF7 , BIBREF8 , BIBREF9 , tree-based MT (TBMT) BIBREF10 , BIBREF11 , or syntax-based MT (SBMT) BIBREF12 .
Inspired by the success of neural MT BIBREF13 , BIBREF14 , recent work has started exploring neural simplification with sequence to sequence (Seq2seq) models, also referred to as encoder-decoder models. Nisioi et al. Nisioi:17 implemented a standard LSTM-based Seq2seq model and found that they outperform PBMT, SBMT, and unsupervised lexical simplification approaches. Zhang and Lapata BIBREF15 viewed the encoder-decoder model as an agent and employed a deep reinforcement learning framework in which the reward has three components capturing key aspects of the target output: simplicity, relevance, and fluency. The common practice for Seq2seq models is to use recurrent neural networks (RNNs) with Long Short-Term Memory BIBREF16 or Gated Recurrent Unit BIBREF17 for the encoder and decoder BIBREF18 , BIBREF15 . These architectures were designed to be capable of memorizing long-term dependencies across sequences. Nevertheless, their memory is typically small and might not be enough for the simplification task, where one is confronted with long and complicated sentences. In this study, we go beyond the conventional LSTM/GRU-based Seq2seq models and propose to use a memory-augmented RNN architecture called Neural Semantic Encoders (NSE). This architecture has been shown to be effective in a wide range of NLP tasks BIBREF0 . The contribution of this paper is twofold:
(1) First, we present a novel simplification model which is, to the best of our knowledge, the first model that use memory-augmented RNN for the task. We investigate the effectiveness of neural Seq2seq models when different neural architectures for the encoder are considered. Our experiments reveal that the NseLstm model that uses an NSE as the encoder and an LSTM as the decoder performed the best among these models, improving over strong simplification systems. (2) Second, we perform an extensive evaluation of various approaches proposed in the literature on different datasets. Results of both automatic and human evaluation show that our approach is remarkably effective for the task, significantly reducing the reading difficulty of the input, while preserving grammaticality and the original meaning. We further discuss some advantages and disadvantages of these approaches.
We investigate the effectiveness of neural Seq2seq models when different neural architectures for the encoder are considered. | What models' effectiveness do they investigate when different neural architectures for the encoder are considered? | Neural Seq2seq models. |
null | false | null | Why is Utah a great place to visit? | Utah is a great state to visit especially for nature lovers. There are 5 famous national parks in Utah also known as The Mighty Five: Zion, Bryce Canyon, Arches, Canyonlands and Capitol Reef. Many parks are close to well developed cities which provide great lodging, dining and entertaining options. Park facilities are well maintained. There are numerous recreational activities in the park for all levels and ages and the views are absolutely stunning. |
|
null | false | null | How many continents are there on Earth? | There are 7 continents - Africa, Antarctica, Asia, Australia, Europe, North America, and South America. |
|
null | false | null | Name some popular American wine brands. | Josh, E&J Gallo, Justin, Chateau Ste. Michelle, Little Penguin, Franzia, Sutter Home |
|
null | false | 205 | Poetry generation is an interesting research topic in the field of text generation. As one of the most valuable literary and cultural heritages of China, Chinese classical poetry is very familiar and loved by Chinese people from generation to generation. It has many particular characteristics in its language structure, ranging from form, sound to meaning, thus is regarded as an ideal testing task for text generation. In this paper, we propose a GPT-2 based uniformed framework for generating major types of Chinese classical poems. We define a unified format for formulating all types of training samples by integrating detailed form information, then present a simple form-stressed weighting method in GPT-2 to strengthen the control to the form of the generated poems, with special emphasis on those forms with longer body length. Preliminary experimental results show this enhanced model can generate Chinese classical poems of major types with high quality in both form and content, validating the effectiveness of the proposed strategy. The model has been incorporated into Jiuge, the most influential Chinese classical poetry generation system developed by Tsinghua University (Guo et al., 2019).
Preliminary experimental results show this enhanced model can generate Chinese classical poems of major types with high quality in both form and content, validating the effectiveness of the proposed strategy. | Whether the proposed model can generate Chinese classical poems of major types? | Yes. |
null | false | 373 | The exploding multimedia content over the Internet, has created a new world of spoken content processing, for example the retrieval BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 , browsing BIBREF5 , summarization BIBREF0 , BIBREF5 , BIBREF6 , BIBREF7 , and comprehension BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 of spoken content. On the other hand, we may realize there still exists a huge part of multimedia content not yet taken care of, i.e., the singing content or those with audio including songs. Songs are human voice carrying plenty of semantic information just as speech. It will be highly desired if the huge quantities of singing content can be similarly retrieved, browsed, summarized or comprehended by machine based on the lyrics just as speech. For example, it is highly desired if song retrieval can be achieved based on the lyrics in addition.
Singing voice can be considered as a special type of speech with highly flexible and artistically designed prosody: the rhythm as artistically designed duration, pause and energy patterns, the melody as artistically designed pitch contours with much wider range, the lyrics as artistically authored sentences to be uttered by the singer. So transcribing lyrics from song audio is an extended version of automatic speech recognition (ASR) taking into account these differences.
On the other hand, singing voice and speech differ widely in both acoustic and linguistic characteristics. Singing signals are often accompanied with some extra music and harmony, which are noisy for recognition. The highly flexible pitch contours with much wider range BIBREF12 , BIBREF13 , the significantly changing phone durations in songs, including the prolonged vowels BIBREF14 , BIBREF15 over smoothly varying pitch contours, create much more problems not existing in speech. The falsetto in singing voice may be an extra type of human voice not present in normal speech. Regarding linguistic characteristics BIBREF16 , BIBREF17 , word repetition and meaningless words (e.g.oh) frequently appear in the artistically authored lyrics in singing voice.
Applying ASR technologies to singing voice has been studied for long. However, not too much work has been reported, probably because the recognition accuracy remained to be relatively low compared to the experiences for speech. But such low accuracy is actually natural considering the various difficulties caused by the significant differences between singing voice and speech. An extra major problem is probably the lack of singing voice database, which pushed the researchers to collect their own closed datasets BIBREF12 , BIBREF15 , BIBREF17 , which made it difficult to compare results from different works.
Having the language model learned from a data set of lyrics is definitely helpful BIBREF15 , BIBREF17 . Hosoya et al. BIBREF16 achieved this with finite state automaton. Sasou et al. BIBREF12 actually prepared a language model for each song. In order to cope with the acoustic characteristics of singing voice, Sasou et al. BIBREF12 , BIBREF14 proposed AR-HMM to take care of the high-pitched sounds and prolonged vowels, while recently Kawai et al. BIBREF15 handled the prolonged vowels by extending the vowel parts in the lexicon, both achieving good improvement. Adaptation from models trained with speech was attractive, and various approaches were compared by Mesaros el al. BIBREF18 .
In this paper, we wish our work can be compatible to more available singing content, therefore in the initial effort we collected about five hours of music-removed version of English songs directly from commercial singing content on YouTube. The descriptive term "music-removed" implies the background music have been removed somehow. Because many very impressive works were based on Japanese songs BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 , the comparison is difficult. We analyzed various approaches with HMM, deep learning with data augmentation, and acoustic adaptation on fragment, song, singer, and genre levels, primarily based on fMLLR BIBREF19 . We also trained the language model with a corpus of lyrics, and modify the pronunciation lexicon and increase the transition probability of HMM for prolonged vowels. Initial results are reported.
In this paper, we wish our work can be compatible to more available singing content, therefore in the initial effort we collected about five hours of music-removed version of English songs directly from commercial singing content on YouTube. | Where did they collect about five hours of the music-removed version of English songs? | YouTube. |
null | false | null | When is the best time of year to visit Seattle? | Seattle is a wonderful city with a variety of different tourist activities. The summer is the most popular time to visit Seattle. Seattle receives the most sun in the summer, which allows for outdoor activities like hiking, boating, and sight seeing. The winters are often cold, dark, and overcast, which most tourists typically avoid. However, if you enjoy winter sports like skiing and snow shoeing, it may be the best time to visit. |
|
null | false | null | The English-language print publication has a circulation of 30,000 qualified subscribers, of which 7,000 are outside the United States. | How many subscribers are inside the United States? | 30,000 - 7,000 = 23,000 |
null | false | 462 | Theoretical Analysis of the DG Setting and Algorithms The DG problem setting was first analysed in. Since then there have been some attempts to analyse DG algorithms from a generalisation bound perspective. However these studies have theoretical results that are either restricted to specific model classes, such as kernel machines, or make strong assumptions about how the domains seen during training will resemble those seen at test time-e.g., that all domains are convex combinations of a finite pre-determined set of prototypical domains. In contrast, our Rademacher complexity approach can be applied to a broad range of model classes (including neural networks), and makes comparatively milder assumptions about the relationship between domains-i.e., they are i.i.d. samples from another arbitrary distribution over domains.
The majority of the existing work investigating the theoretical foundations of DG follow the initial formalisation of the domain generalisation problem put forth by, where the goal is to minimise the expected error over unseen domains. However, several recent works have also explored the idea of bounding the error on a single unseen domain with the most pathological distribution shift. This type of analysis is typically rooted in methods from causal inference, rather than statistical learning theory. As a consequence, they are able to make stronger claims for the problems they address, but the scope of their analysis is necessarily limited to the scenarios where their assumptions about the underlying causal structures are valid. For example, Janzing (2019) provides bounds that assume problems conform to a specific class of structural equation models, and the analysis is performed under the assumption that infinite training data is available within each of the observed training domains. Throughout the work we address the stan-dard DG formalisation given by, where one is concerned with the expected performance of a model on domains sampled from some distribution over domains.
Others rely on trying to link between domain adaptation objectives (where target domains are observable for alignment to source domains) and domain generalisation (where target domains are not observable and thus cannot correctly be used in a learning objective). proceed by making assumptions on the structure of the distribution over possible domains (i.e., that it has support determined by the convex hull of a finite set of prototypical domains), which allows them to upper bound the domain alignment metric. provide a bound that depends on an unobservable domain distance quantity, which they then approximate in experiments using kernel density estimates. is another piece of work that theoretically investigates the generalisation of ERM in a DG setting. They deal with online DG, where each time-step corresponds to observing a new domain, and the learner must produce a new model capable of generalising to novel domains. Another point of difference between their work and the standard DG problem setting of is that the domain at each time-step is chosen by an adversary. They analyse this game for a finite number of time-steps, but they assume each domain has an infinite amount of data. They also put some limitations on the adversary: e.g., it must choose a domain that is a convex combination of a finite number of pre-determined domains. In contrast, our theoretical analysis is in the more realistic setting where one has a finite amount of data per domain, and the domains we consider are not limited to convex combinations of a set of prototypical domains. Possibly the most similar work to our theoretical contributions is due to, who also provide learningtheoretic generalisation bounds for DG. However, their analysis only applies to finite hypothesis classes (which does not include, e.g., linear models or neural networks), whereas ours can be applied to any class amenable to analysis with Rademacher complexity.
The main existing empirical analysis on DG is, who compared several state of the art DG methods under a common evaluation and hyper-parameter tuning protocol called DomainBed. They ultimately defend Empirical Risk Minimization (ERM) over more sophisticated alternatives on the grounds that no competitor consistently beats it across the benchmark suite. We also broadly defend ERM, and build on the same benchmark, but differently we provide a much deeper analysis into when and why ERM works. More specifically: (i) We provide a new theoretical analysis of ERM's generalisation quality unlike the prior purely empirical evaluation, (ii) We re-use the DomainBed benchmark to directly corroborate this theory under controlled conditions using linear models where model complexity can be tractably and accurately tuned. (iii) We use our complexity-based analysis to explain the previously erratic results of prior DomainBed competitors in terms of model complexity. (iv) We identify, and empirically validate, the preferred model selection criterion for DG, a point which was inconclusive in.
Theoretical Analysis of the DG Setting and Algorithms The DG problem setting was first analysed in. Since then there have been some attempts to analyse DG algorithms from a generalisation bound perspective. However these studies have theoretical results that are either restricted to specific model classes, such as kernel machines, or make strong assumptions about how the domains seen during training will resemble those seen at test time-e.g., that all domains are convex combinations of a finite pre-determined set of prototypical domains. In contrast, our Rademacher complexity approach can be applied to a broad range of model classes (including neural networks), and makes comparatively milder assumptions about the relationship between domains-i.e., they are i.i.d. samples from another arbitrary distribution over domains.
The majority of the existing work investigating the theoretical foundations of DG follow the initial formalisation of the domain generalisation problem put forth by, where the goal is to minimise the expected error over unseen domains. However, several recent works have also explored the idea of bounding the error on a single unseen domain with the most pathological distribution shift. This type of analysis is typically rooted in methods from causal inference, rather than statistical learning theory. As a consequence, they are able to make stronger claims for the problems they address, but the scope of their analysis is necessarily limited to the scenarios where their assumptions about the underlying causal structures are valid. For example, Janzing (2019) provides bounds that assume problems conform to a specific class of structural equation models, and the analysis is performed under the assumption that infinite training data is available within each of the observed training domains. Throughout the work we address the stan-dard DG formalisation given by, where one is concerned with the expected performance of a model on domains sampled from some distribution over domains.
Others rely on trying to link between domain adaptation objectives (where target domains are observable for alignment to source domains) and domain generalisation (where target domains are not observable and thus cannot correctly be used in a learning objective). proceed by making assumptions on the structure of the distribution over possible domains (i.e., that it has support determined by the convex hull of a finite set of prototypical domains), which allows them to upper bound the domain alignment metric. provide a bound that depends on an unobservable domain distance quantity, which they then approximate in experiments using kernel density estimates. is another piece of work that theoretically investigates the generalisation of ERM in a DG setting. They deal with online DG, where each time-step corresponds to observing a new domain, and the learner must produce a new model capable of generalising to novel domains. Another point of difference between their work and the standard DG problem setting of is that the domain at each time-step is chosen by an adversary. They analyse this game for a finite number of time-steps, but they assume each domain has an infinite amount of data. They also put some limitations on the adversary: e.g., it must choose a domain that is a convex combination of a finite number of pre-determined domains. In contrast, our theoretical analysis is in the more realistic setting where one has a finite amount of data per domain, and the domains we consider are not limited to convex combinations of a set of prototypical domains. Possibly the most similar work to our theoretical contributions is due to, who also provide learningtheoretic generalisation bounds for DG. However, their analysis only applies to finite hypothesis classes (which does not include, e.g., linear models or neural networks), whereas ours can be applied to any class amenable to analysis with Rademacher complexity.
The main existing empirical analysis on DG is, who compared several state of the art DG methods under a common evaluation and hyper-parameter tuning protocol called DomainBed. They ultimately defend Empirical Risk Minimization (ERM) over more sophisticated alternatives on the grounds that no competitor consistently beats it across the benchmark suite. We also broadly defend ERM, and build on the same benchmark, but differently we provide a much deeper analysis into when and why ERM works. More specifically: (i) We provide a new theoretical analysis of ERM's generalisation quality unlike the prior purely empirical evaluation, (ii) We re-use the DomainBed benchmark to directly corroborate this theory under controlled conditions using linear models where model complexity can be tractably and accurately tuned. (iii) We use our complexity-based analysis to explain the previously erratic results of prior DomainBed competitors in terms of model complexity. (iv) We identify, and empirically validate, the preferred model selection criterion for DG, a point which was inconclusive in.
Ye et al. (2021) provide a bound that depends on an unobservable domain distance quantity, which they then approximate in experiments using kernel density estimates.****Possibly the most similar work to our theoretical contributions is due to Ahuja et al. (2021), who also provide learningtheoretic generalisation bounds for DG. However, their analysis only applies to finite hypothesis classes (which does not include, e.g., linear models or neural networks), whereas ours can be applied to any class amenable to analysis with Rademacher complexity. | What is the relationship of this work to prior theory [1,2]? [1] Towards a Theoretical Framework of Out-of-Distribution Generalization [2] Empirical or Invariant Risk Minimization? A Sample Complexity Perspective | [1] Does not have a sample complexity generalisation bound which we provide here. They analyse generalisation in terms of distribution distance, but this is not observable. In their experiments they approximate it with kernel density estimation, but the quality of their approximation is not well-justified. (Note also that our theory does not depend on the distance between the source and target distributions.) [2] Provides sample complexity bounds but they only apply to finite hypothesis classes. This means that it is not applicable to real-world model classes. In contrast, we are able to analyse models such as linear classifiers and neural networks with our Rademacher complexity based theory. We have added discussion of these papers to our related work section. |
null | false | null | Give me some tips for increase my savings | To increase the savings, one needs to always strike a mindful balance between wants and needs and more importantly, it also depends on individual circumstances. Here are some things that one can start doing to increase the savings
1. Try to pay off the debt as soon as possible.
2. The principal of "Pay oneself first". It doesn't mean spend this money on wishes, but on the contrary, first put the fixed amount in the savings accounts and then try to go by your expenses.
3. Track your expenses. Sometimes we just don't know that we have so many useless subscriptions that are squeezing your hard earned money under your own eyes.
4. Try to cook at home rather than going to the restaurants or ordering food online on frequent basis. |
|
null | false | null | If you find yourself being pursued by a hungry bear in the woods, label each of these items in your backpack as either 'useful', 'not useful', or 'an unfortunate choice to pack': a fish, a gun, a book, a hamburger. | In such a situation, being pursued by a hungry bear in the woods, a fish would be 'not useful', a gun would be 'useful', a book would be 'not useful', and a hamburger would be 'an unfortunate choice to pack'. |
|
null | false | null | The Crucifixion with the Virgin and St John is an oil painting by the Dutch artist Hendrick ter Brugghen, probably painted around 1625 as an altarpiece for a Catholic clandestine church likely located in Utrecht in the Calvinist Dutch United Provinces. The scene is taken from the account of the crucifixion of Jesus in Gospel of John, and depicts Jesus being mourned by his mother Mary and John the Apostle. At the base of the cross are bones, traditionally identified as those of Adam. The scene is set in an atmosphere of deep dusk with stars visible in the background. The painting was not known until it was discovered in a bombed-out church in South Hackney, London, in 1956. Initially unrecognized, it was identified before an appearance at Sotheby's later that year. It was acquired by the Metropolitan Museum of Art in New York, where it now hangs. | Who was crucified in the crucifixion painting | Jesus Christ.
The oil painting is of a scene from the account of the Gospel of John |
null | false | null | what are the main components of a computer? | A computer consists of a motherboard, a central processing unit (CPU), memory (often called RAM) and a hard disk or hard drive or Solid State Drive for storage. |
|
null | false | null | How to get a good Survey for a product | Firstly you need to respond to the customer with in SLA and try to engage in all the triaging things you are doing to debug the case. secondly schedule 1-1 with customer to see if this can reproduced in the call. if it is a bug, please raise it with the developer to see if they can fix it. finally build the trust and confidence in the product and try to meet customer expectations. |
|
null | false | null | How much money did George Lucas get for selling Lucasfilm to Disney? | $4 billion |
|
null | false | null | How big is the universe? | While the entire size of the universe is unknown, scientists have estimated that the universe is approximately 93 billion light-years in diameter. |
|
null | false | 114 | We compared our model with MLE, RL and GAN baselines. Since COCO and EMNLP2017 WMT don't have input while WeiboDial regards posts as input, we chose the following baselines respectively:
MLE: a RNN model trained with MLE objective BIBREF4 . Its extension, Seq2Seq, can work on the dialogue dataset BIBREF2 .
SeqGAN: The first text GAN model that updates the generator with policy gradient based on the rewards from the discriminator BIBREF7 .
LeakGAN: A variant of SeqGAN that provides rewards based on the leaked information of the discriminator for the generator BIBREF11 .
MaliGAN: A variant of SeqGAN that optimizes the generator with a normalized maximum likelihood objective BIBREF8 .
IRL: This inverse reinforcement learning method replaces the discriminator with a reward approximator to provide dense rewards BIBREF12 .
RAML: A RL approach to incorporate MLE objective into RL training framework, which regards BLEU as rewards BIBREF17 .
DialogGAN: An extension of SeqGAN tuned to dialogue generation task with MLE objective added to the adversarial objective BIBREF16 .
DPGAN: A variant of DialogGAN which uses a language model based discriminator and regards cross-entropy as rewards BIBREF13 .
Note that MLE, SeqGAN, LeakGAN, MaliGAN and IRL are the baselines on COCO and EMNLP2017 WMT, while MLE, RAML, DialogGAN, and DPGAN on WeiboDial. The original codes are used to test the baselines.
We compared our model with MLE, RL and GAN baselines. | What baselines did the author compare their model with? | MLE, RL and GAN baselines. |
null | false | null | The comic strip Mickey Mouse debuted on January 13, 1930, in New York Daily Mirror and by 1931, the strip was published in 60 newspapers in the U.S., and in twenty other countries. After realizing releasing merchandise based on the characters would generate more revenue for the company, in New York, a man asked Walt for the license to put Mickey Mouse on writing tablets he was manufacturing for $300. Walt agreed and Mickey Mouse became the first licensed character. In 1933, Walt asked Kay Karmen, the owner of an Kansas City advertising firm, to run Disney's merchandising; Karmen agreed and transformed Disney's merchandising. Within a year, Kamen had 40 licenses for Mickey Mouse and within two years, had made $35 million worth of sales. In 1934, Walt said he made more money from the merchandising of Mickey Mouse than from the character's films. | Please summarize Micky Mouse's initial licensing and merchandising history | The Mickey Mouse comic was first published on January 1930 in 60 newspapers in the U.S. and in 20 other countries. A man realized he could make more money by selling merchandise based on the characters from the comic and asked Walt for a license to put Mickey Mouse on his product. By 1934, Disney had 40 licenses for Mickey Mouse and by 1936, it had $35 million worth of sales, making more money from merchandising Mickey Mouse than from the character's films. |
null | false | null | Why do people like cats? | People like cats because they do their own thing, they are very independent and have strong personalities, according to cat lovers. |
|
null | false | null | Sanderson rose to prominence in late 2007 when Harriet McDougal, the wife and editor of author Robert Jordan, chose Sanderson to complete the final books in Jordan's epic fantasy series The Wheel of Time after Jordan's death. McDougal asked Sanderson to finish the series after being deeply impressed by his first Mistborn novel, The Final Empire. Tor Books made the announcement on December 7, 2007. After reviewing what was necessary to complete the series, Sanderson and Tor announced on March 30, 2009, that a final three books would be published instead of just one.
The first of these, The Gathering Storm, was published on October 27, 2009, and reached the number-one spot on the New York Times bestseller list for hardcover fiction.Towers of Midnight, the second-to-last Wheel of Time book, was published just over a year after The Gathering Storm on November 2, 2010, debuting at number one on the bestseller list. In early 2013, the series was completed with the publication of A Memory of Light. | Who finished the Wheel of Time series? | The death of Robert Jordan, the original author of the Wheel of Time series, prompted his wife and editor, Harriet McDougal, to find someone to complete her late husband's work. She chose Brandon Sanderson, who had recently released the popular first novel of the Mistborn series, The Final Empire. Sanderson agreed after reviewing Jordan's notes. |
null | false | 35 | The BLEU metric parameters (modified 1-gram, 2-gram, 3-gram and 4-gram precision values) and human evaluation metrics of adequacy, fluency and relative ranking values were used to evaluate the performance of the models.
The BLEU metric parameters (modified 1-gram, 2-gram, 3-gram and 4-gram precision values) and human evaluation metrics of adequacy, fluency and relative ranking values were used to evaluate the performance of the models. | How to evaluate the performance of the models? | The BLEU metric parameters (modified 1-gram, 2-gram, 3-gram and 4-gram precision values) and human evaluation metrics of adequacy, fluency and relative ranking values were used to evaluate the performance of the models. |
null | false | null | is IT a great career | yes |
|
null | false | null | The Sinti are a subgroup of Romani people mostly found in Germany. They arrived in Austria and Germany in the Late Middle Ages as part of the Romani emigration from the Indian Subcontinent, eventually splitting into two groups: Eftavagarja ("the Seven Caravans") and Estraxarja ("from Austria"). They arrived in Germany before 1540. The two groups expanded, the Eftavagarja into France, Portugal and Brazil, where they are called "Manouches", and the Estraxarja into Italy and Central Europe, mainly what are now Croatia, Slovenia, Hungary, Romania, the Czech Republic and Slovakia, eventually adopting various regional names. | Who are Sinti people? | Sinti are a group of people mostly settled in Germany but also settled in other European countries. |
null | false | null | What is the largest living thing? | The Honey mushroom |
|
null | false | 105 | Since humans amass more and more generally available data in the form of unstructured text it would be very useful to teach machines to read and comprehend such data and then use this understanding to answer our questions. A significant amount of research has recently focused on answering one particular kind of questions the answer to which depends on understanding a context document. These are cloze-style questions BIBREF0 which require the reader to fill in a missing word in a sentence. An important advantage of such questions is that they can be generated automatically from a suitable text corpus which allows us to produce a practically unlimited amount of them. That opens the task to notoriously data-hungry deep-learning techniques which now seem to outperform all alternative approaches.
Two such large-scale datasets have recently been proposed by researchers from Google DeepMind and Facebook AI: the CNN/Daily Mail dataset BIBREF1 and the Children's Book Test (CBT) BIBREF2 respectively. These have attracted a lot of attention from the research community BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 with a new state-of-the-art model coming out every few weeks.
However if our goal is a production-level system actually capable of helping humans, we want the model to use all available resources as efficiently as possible. Given that
we believe that if the community is striving to bring the performance as far as possible, it should move its work to larger data.
This thinking goes in line with recent developments in the area of language modelling. For a long time models were being compared on several "standard" datasets with publications often presenting minuscule improvements in performance. Then the large-scale One Billion Word corpus dataset appeared BIBREF15 and it allowed Jozefowicz et al. to train much larger LSTM models BIBREF16 that almost halved the state-of-the-art perplexity on this dataset.
We think it is time to make a similar step in the area of text comprehension. Hence we are introducing the BookTest, a new dataset very similar to the Children's Book test but more than 60 times larger to enable training larger models even in the domain of text comprehension. Furthermore the methodology used to create our data can later be used to create even larger datasets when the need arises thanks to further technological progress.
We show that if we evaluate a model trained on the new dataset on the now standard Children's Book Test dataset, we see an improvement in accuracy much larger than other research groups achieved by enhancing the model architecture itself (while still using the original CBT training data). By training on the new dataset, we reduce the prediction error by almost one third. On the named-entity version of CBT this brings the ensemble of our models to the level of human baseline as reported by Facebook BIBREF2 . However in the final section we show in our own human study that there is still room for improvement on the CBT beyond the performance of our model.
Hence we are introducing the BookTest, a new dataset very similar to the Children's Book test but more than 60 times larger to enable training larger models even in the domain of text comprehension. | What's the difference between the BookTest and the Children's Book test? | The BookTest is a new dataset very similar to the Children's Book test but more than 60 times larger to enable training larger models even in the domain of text comprehension. |
null | false | null | What are different types of optical illusions? | There are three types of optical illusions: Physical, physiological and cognitive.
Physical optical illusion occurs when the light bends or changes direction before it reaches human eye. When light travels through different mediums such as air , water or glass, its speed changes and change in speed causes the light to bend. This phenomenon is called refraction. Rainbow is an example of physical optical illusion caused by refraction.
Physiological optical illusion occurs when a pattern repeats over and over again and it is too much information for the brain. This causes the brain to think the images are moving. Kinetoscope uses images to create a video using physiological optical illusion. Flip books are another example of physiological optical illusion.
Cognitive optical illusion occurs when multiple people view the same image but they interpret it differently. Interpretation of the image depends of people's past experience and personalities. |
|
null | false | null | A large language model (LLM) is a language model consisting of a neural network with many parameters (typically billions of weights or more), trained on large quantities of unlabelled text using self-supervised learning. LLMs emerged around 2018 and perform well at a wide variety of tasks. This has shifted the focus of natural language processing research away from the previous paradigm of training specialized supervised models for specific tasks.
Properties
Though the term large language model has no formal definition, it often refers to deep learning models having a parameter count on the order of billions or more. LLMs are general purpose models which excel at a wide range of tasks, as opposed to being trained for one specific task (such as sentiment analysis, named entity recognition, or mathematical reasoning). The skill with which they accomplish tasks, and the range of tasks at which they are capable, seems to be a function of the amount of resources (data, parameter-size, computing power) devoted to them, in a way that is not dependent on additional breakthroughs in design.
Though trained on simple tasks along the lines of predicting the next word in a sentence, neural language models with sufficient training and parameter counts are found to capture much of the syntax and semantics of human language. In addition, large language models demonstrate considerable general knowledge about the world, and are able to "memorize" a great quantity of facts during training.
Hallucinations
Main article: Hallucination (artificial intelligence)
In artificial intelligence in general, and in large language models in particular, a "hallucination" is a confident response that does not seem to be justified by the model's training data.
Emergent abilities
On a number of natural language benchmarks involving tasks such as question answering, models perform no better than random chance until they reach a certain scale (in this case, measured by training computation), at which point their performance sharply increases. These are examples of emergent abilities.
Unpredictable abilities that have been observed in large language models but that were not present in simpler models (and that were not explicitly designed into the model) are usually called "emergent abilities". Researchers note that such abilities "cannot be predicted simply by extrapolating the performance of smaller models". These abilities are discovered rather than programmed-in or designed, in some cases only after the LLM has been publicly deployed. Hundreds of emergent abilities have been described. Examples include multi-step arithmetic, taking college-level exams, identifying the intended meaning of a word, chain-of-thought prompting, decoding the International Phonetic Alphabet, unscrambling a word’s letters, identifying offensive content in paragraphs of Hinglish (a combination of Hindi and English), and generating a similar English equivalent of Kiswahili proverbs.
Architecture and training
Large language models have most commonly used the transformer architecture, which, since 2018, has become the standard deep learning technique for sequential data (previously, recurrent architectures such as the LSTM were most common). LLMs are trained in an unsupervised manner on unannotated text. A left-to-right transformer is trained to maximize the probability assigned to the next word in the training data, given the previous context. Alternatively, an LLM may use a bidirectional transformer (as in the example of BERT), which assigns a probability distribution over words given access to both preceding and following context. In addition to the task of predicting the next word or "filling in the blanks", LLMs may be trained on auxiliary tasks which test their understanding of the data distribution such as Next Sentence Prediction (NSP), in which pairs of sentences are presented and the model must predict whether they appear side-by-side in the training corpus.
The earliest LLMs were trained on corpora having on the order of billions of words. The first model in OpenAI's GPT series was trained in 2018 on BookCorpus, consisting of 985 million words. In the same year, BERT was trained on a combination of BookCorpus and English Wikipedia, totalling 3.3 billion words. In the years since then, training corpora for LLMs have increased by orders of magnitude, reaching up to hundreds of billions or trillions of tokens.
LLMs are computationally expensive to train. A 2020 study estimated the cost of training a 1.5 billion parameter model (1-2 orders of magnitude smaller than the state of the art at the time) at $1.6 million.
A 2020 analysis found that neural language models' capability (as measured by training loss) increased smoothly in a power law relationship with number of parameters, quantity of training data, and computation used for training. These relationships were tested over a wide range of values (up to seven orders of magnitude) and no attenuation of the relationship was observed at the highest end of the range (including for network sizes up to trillions of parameters).
Application to downstream tasks
Between 2018 and 2020, the standard method for harnessing an LLM for a specific natural language processing (NLP) task was to fine tune the model with additional task-specific training. It has subsequently been found that more powerful LLMs such as GPT-3 can solve tasks without additional training via "prompting" techniques, in which the problem to be solved is presented to the model as a text prompt, possibly with some textual examples of similar problems and their solutions.
Fine-tuning
Main article: Fine-tuning (machine learning)
Fine-tuning is the practice of modifying an existing pretrained language model by training it (in a supervised fashion) on a specific task (e.g. sentiment analysis, named entity recognition, or part-of-speech tagging). It is a form of transfer learning. It generally involves the introduction of a new set of weights connecting the final layer of the language model to the output of the downstream task. The original weights of the language model may be "frozen", such that only the new layer of weights connecting them to the output are learned during training. Alternatively, the original weights may receive small updates (possibly with earlier layers frozen).
Prompting
See also: Prompt engineering and Few-shot learning (natural language processing)
In the prompting paradigm, popularized by GPT-3, the problem to be solved is formulated via a text prompt, which the model must solve by providing a completion (via inference). In "few-shot prompting", the prompt includes a small number of examples of similar (problem, solution) pairs. For example, a sentiment analysis task of labelling the sentiment of a movie review could be prompted as follows:
Review: This movie stinks.
Sentiment: negative
Review: This movie is fantastic!
Sentiment:
If the model outputs "positive", then it has correctly solved the task. In zero-shot prompting, no solve examples are provided. An example of a zero-shot prompt for the same sentiment analysis task would be "The sentiment associated with the movie review 'This movie is fantastic!' is".
Few-shot performance of LLMs has been shown to achieve competitive results on NLP tasks, sometimes surpassing prior state-of-the-art fine-tuning approaches. Examples of such NLP tasks are translation, question answering, cloze tasks, unscrambling words, and using a novel word in a sentence. The creation and optimisation of such prompts is called prompt engineering.
Instruction tuning
Instruction tuning is a form of fine-tuning designed to facilitate more natural and accurate zero-shot prompting interactions. Given a text input, a pretrained language model will generate a completion which matches the distribution of text on which it was trained. A naive language model given the prompt "Write an essay about the main themes of Hamlet." might provide a completion such as "A late penalty of 10% per day will be applied to submissions received after March 17." In instruction tuning, the language model is trained on many examples of tasks formulated as natural language instructions, along with appropriate responses. Various techniques for instruction tuning have been applied in practice. OpenAI's InstructGPT protocol involves supervised fine-tuning on a dataset of human-generated (prompt, response) pairs, followed by reinforcement learning from human feedback (RLHF), in which a reward function was learned based on a dataset of human preferences. Another technique, "self-instruct", fine-tunes the language model on a training set of examples which are themselves generated by an LLM (bootstrapped from a small initial set of human-generated examples).
https://en.wikipedia.org/wiki/Large_language_model | Given these paragraphs about Large language models, what is an LLM? | A large language model (LLM) is a language model consisting of a neural network with many parameters (typically billions of weights or more), trained on large quantities of unlabelled text using self-supervised learning. |
null | false | null | How many different Star Trek television series and movies were there in total? | In total, there were 21 when you add both the eight television series and 13 movies. |
|
null | false | 340 | * Equal contribution. Listing order is random.
Lemmatization is a core NLP task that involves a string-to-string transduction from an inflected word form to its citation form, known as the lemma. More concretely, consider the English sentence: The bulls are running in Pamplona. A lemmatizer will seek to map each word to a form you may find in a dictionary—for instance, mapping running to run. This linguistic normalization is important in several downstream NLP applications, especially for highly inflected languages. Lemmatization has previously been shown to improve recall for information retrieval BIBREF0 , BIBREF1 , to aid machine translation BIBREF2 , BIBREF3 and is a core part of modern parsing systems BIBREF4 , BIBREF5 .
However, the task is quite nuanced as the proper choice of the lemma is context dependent. For instance, in the sentence A running of the bulls took place in Pamplona, the word running is its own lemma, since, here, running is a noun rather than an inflected verb. Several counter-examples exist to this trend, as discussed in depth in haspelmath2013understanding. Thus, a good lemmatizer must make use of some representation of each word's sentential context. The research question in this work is, then, how do we design a lemmatization model that best extracts the morpho-syntax from the sentential context?
Recent work BIBREF7 has presented a system that directly summarizes the sentential context using a recurrent neural network to decide how to lemmatize. As N18-1126's system currently achieves state-of-the-art results, it must implicitly learn a contextual representation that encodes the necessary morpho-syntax, as such knowledge is requisite for the task. We contend, however, that rather than expecting the network to implicitly learn some notion of morpho-syntax, it is better to explicitly train a joint model to morphologically disambiguate and lemmatize. Indeed, to this end, we introduce a joint model for the introduction of morphology into a neural lemmatizer. A key feature of our model is its simplicity: Our contribution is to show how to stitch existing models together into a joint model, explaining how to train and decode the model. However, despite the model's simplicity, it still achieves a significant improvement over the state of the art on our target task: lemmatization.
Experimentally, our contributions are threefold. First, we show that our joint model achieves state-of-the-art results, outperforming (on average) all competing approaches on a 20-language subset of the Universal Dependencies (UD) corpora BIBREF8 . Second, by providing the joint model with gold morphological tags, we demonstrate that we are far from achieving the upper bound on performance—improvements on morphological tagging could lead to substantially better lemmatization. Finally, we provide a detailed error analysis indicating when and why morphological analysis helps lemmatization. We offer two tangible recommendations: one is better off using a joint model (i) for languages with fewer training data available and (ii) languages that have richer morphology.
Our system and pre-trained models on all languages in the latest version of the UD corpora are released at https://sigmorphon.github.io/sharedtasks/2019/task2/.
First, we show that our joint model achieves state-of-the-art results, outperforming (on average) all competing approaches on a 20-language subset of the Universal Dependencies (UD) corpora. | Does the authors' joint model achieve state-of-the-art results? | Yes, it does. |
null | false | null | Selena Marie Gomez was born on July 22, 1992, in Grand Prairie, Texas, to Ricardo Joel Gomez and Texas-born former stage actress Mandy Teefey. She was named after Tejano singer Selena Quintanilla, who died in 1995. Her father is of Mexican descent, while her mother, who was adopted, has Italian ancestry. Gomez's paternal grandparents emigrated to Texas from Monterrey in the 1970s. Of her heritage, Gomez has said she is "a proud third-generation American-Mexican" and "My family does have quinceañeras, and we go to the communion church. We do everything that's Catholic, but we don't really have anything traditional except go to the park and have barbecues on Sundays after church." Gomez was fluent in Spanish until age seven. Her parents divorced when she was five years old, and she remained with her mother. Gomez has two younger half-sisters and a younger stepbrother: Gracie Elliot Teefey, through Mandy and her second husband, Brian Teefey, and Victoria "Tori" and Marcus Gomez, through Ricardo and his second wife, Sara. She earned her high-school diploma through homeschooling in May 2010. | Given this paragraph about Selena Gomez, how many siblings does she have? | Selena has 2 half-sisters and 1 stepbrother. |
null | false | null | The Purdue Boilermakers basketball team is a men's college basketball program that competes in NCAA Division I and is a member of the Big Ten Conference.
Purdue basketball holds the most Big Ten regular season championships, with 25. Purdue also holds a winning record against all other Big Ten schools in head-to-head match ups.
The Boilermakers have reached two NCAA Tournament Final Fours and one championship game, but have not won an NCAA Championship. The 1931–32 team was retroactively named a national champion by the Helms Athletic Foundation and the Premo-Porretta Power Poll.
Purdue has sent more than 30 players to the NBA, including two overall No. 1 picks in the NBA draft. Purdue has one main rivalry against the Indiana Hoosiers (see Indiana–Purdue Rivalry). | Tell me how many final fours purdue has been to and whether or not they have won any championships | The Purdue Boilermakers men's basketball team has reached two NCAA Tournament Final Fours.
While they have not won any national championships, they were retroactively named a national champion in 1932. They have also won 25 Big Ten regular season championships. |