id
stringlengths 32
33
| x
stringlengths 41
1.75k
| y
stringlengths 4
39
|
|---|---|---|
2cedb1a0f0c0fbb9bd95d5b54e4967_5 | Table 1 shows the statistical data of the corpus from <cite>Ma, Jurczyk, and Choi [2018]</cite> . Based on the above corpus we created a new data split different from <cite>Ma, Jurczyk, and Choi [2018]</cite> 's data split. | uses |
2cedb1a0f0c0fbb9bd95d5b54e4967_6 | In the previous work of <cite>Ma, Jurczyk, and Choi [2018]</cite> , they used a random data split where 1,187 of 1,349 queries in the development set and 1,207 of 1,353 queries in the test set are generated from the same plot summaries as some queries in the training set with only masking the different character entities which makes the model can see the right answer in the training set. | motivation |
2cedb1a0f0c0fbb9bd95d5b54e4967_7 | We propose three tasks, one is from <cite>Ma, Jurczyk, and Choi [2018]</cite> , and another two tasks are new tasks designed by us. | uses |
2cedb1a0f0c0fbb9bd95d5b54e4967_8 | We propose three tasks, one is from <cite>Ma, Jurczyk, and Choi [2018]</cite> , and another two tasks are new tasks designed by us. The single variable task from <cite>Ma, Jurczyk, and Choi [2018]</cite> consists a dialogue passage p, a query q which is from plot summary of the dialogue passage and an answer a. In this 1 https://github.com/emorynlp/character-mining task, a query q replaces only one character entity with an unknown variable x and the machine is asked to infer the replaced character entity (answer a) from all the possible entities appear in the dialogue passage p. This task is evaluated by computing the accuracy of predictions (see Section ). | uses |
2cedb1a0f0c0fbb9bd95d5b54e4967_9 | The single variable task from <cite>Ma, Jurczyk, and Choi [2018]</cite> consists a dialogue passage p, a query q which is from plot summary of the dialogue passage and an answer a. In this 1 https://github.com/emorynlp/character-mining task, a query q replaces only one character entity with an unknown variable x and the machine is asked to infer the replaced character entity (answer a) from all the possible entities appear in the dialogue passage p. This task is evaluated by computing the accuracy of predictions (see Section ). | background |
2cedb1a0f0c0fbb9bd95d5b54e4967_10 | Based on <cite>Ma, Jurczyk, and Choi [2018]</cite> , we first use CNN to extract the gram-level features of utterances and then use @ent04 asks @ent00 how someone could get a hold of @ent00 's credit card number and @ent00 is surprised at how much was spent . | uses |
2cedb1a0f0c0fbb9bd95d5b54e4967_11 | This method is the SOTA method last year in <cite>Ma, Jurczyk, and Choi [2018]</cite> 's data split which is also selected as one of our experimental methods. | uses |
2cedb1a0f0c0fbb9bd95d5b54e4967_12 | Adding a CNN can achieve even lower accuracy because passing sequences to the CNN only keeps important information after the pooling operation, but for dialogue data, most of the time the replaced entity needs to be decided by <cite>Ma, Jurczyk, and Choi [2018]</cite> are not helpful for these tasks on our data split because dialogues contain so many informal expressions and the size of the corpus is small. | motivation |
2cedb1a0f0c0fbb9bd95d5b54e4967_13 | Results Table 4 shows the results of our experiment. BiL-STM is good at capturing the sequence information of sentences; however, since it only finds some kind of answer distributions on the sequence information, it cannot capture the information of the relation between query and utterance. Adding a CNN can achieve even lower accuracy because passing sequences to the CNN only keeps important information after the pooling operation, but for dialogue data, most of the time the replaced entity needs to be decided by <cite>Ma, Jurczyk, and Choi [2018]</cite> are not helpful for these tasks on our data split because dialogues contain so many informal expressions and the size of the corpus is small. | uses differences |
2d2da2e9215691bffad74bfb97dbf3_0 | This was the case in SemEval-2013, whose task 2 <cite>(Wilson et al., 2013)</cite> required sentiment analysis of Twitter and SMS text messages. | background |
2d2da2e9215691bffad74bfb97dbf3_1 | And perhaps this is the cause for lower score in the unconstrained mode, something that happened also with many systems in the past edition <cite>(Wilson et al., 2013)</cite> . | similarities |
2d2ec7230a651d1d6786d0f8a71f7e_0 | These two lines of research converge in prior work to show, e.g., the increasing association of the lexical item 'gay' with the meaning dimension of homosexuality<cite> (Kim et al., 2014</cite>; Kulkarni et al., 2015) . | background |
2d2ec7230a651d1d6786d0f8a71f7e_1 | It is thus a continuation of prior work, in which we investigated historical English texts only (Hellrich and Hahn, 2016a) , and also influenced by the design decisions of <cite>Kim et al. (2014)</cite> and Kulkarni et al. (2015) which were the first to use word embeddings in diachronic studies. | uses |
2d2ec7230a651d1d6786d0f8a71f7e_2 | Word embeddings can be used rather directly for tracking semantic changes, namely by measuring the similarity of word representations generated for one word at different points in time-words which underwent semantic shifts will be dissimilar with themselves. These models must either be trained in a continuous manner where the model for each time span is initialized with its predecessor<cite> (Kim et al., 2014</cite>; Hellrich and Hahn, 2016b) , or a mapping between models for different points in time must be calculated (Kulkarni et al., 2015; Hamilton et al., 2016) . The first approach cannot be performed in parallel and is thus rather time-consuming, if texts are not subsampled. | motivation background |
2d2ec7230a651d1d6786d0f8a71f7e_3 | These models must either be trained in a continuous manner where the model for each time span is initialized with its predecessor<cite> (Kim et al., 2014</cite>; Hellrich and Hahn, 2016b) , or a mapping between models for different points in time must be calculated (Kulkarni et al., 2015; Hamilton et al., 2016) . | background |
2d2ec7230a651d1d6786d0f8a71f7e_4 | The averaged cosine values between word embeddings before and after an epoch are used as a convergence measure c<cite> (Kim et al., 2014</cite>; Kulkarni et al., 2015) . | uses |
2d2ec7230a651d1d6786d0f8a71f7e_5 | The convergence criterion proposed by Kulkarni et al. (2015) , i.e., c = 0.9999, was never reached (this observation might be explained by Kulkarni et al.'s decision not to reset the learning rate for each training epoch, as was done by us and <cite>Kim et al. (2014)</cite> ). | similarities |
2d7e98487698b0b6ae85f052402f7c_0 | Prosodic Cues for DA Recognition: It has also been noted that prosodic knowledge plays a major role in DA identification for certain DA types<cite> Stolcke et al., 2000)</cite> . The main reason is that the acoustic signal of the same utterance can be very different in a different DA class. This indicates that if one wants to classify DA classes only from the text, the context must be an important aspect to consider: simply classifying single utterances might not be enough, but considering the preceding utterances as a context is important. | background |
2d7e98487698b0b6ae85f052402f7c_1 | Lexical, Prosodic, and Syntactic Cues: Many studies have been carried out to find out the lexical, prosodic and syntactic cues <cite>(Stolcke et al., 2000</cite>; Surendran and Levow, 2006; O'Shea et al., 2012; Yang et al., 2014) . | background |
2d7e98487698b0b6ae85f052402f7c_2 | For the SwDA corpus, the state-of-the-art baseline result was 71% for more than a decade using a standard Hidden Markov Model (HMM) with language features such as words and n-grams<cite> (Stolcke et al., 2000)</cite> . The inter-annotator agreement accuracy for the same corpus is 84%, and in this particular case, we are still far from achieving human accuracy. However, words like 'yeah' appear in many classes such as backchannel, yes-answer, agree/accept etc. | motivation background |
2d7e98487698b0b6ae85f052402f7c_3 | We follow the same data split of 1115 training and 19 test conversations as in the baseline approach <cite>(Stolcke et al., 2000</cite>; Kalchbrenner and Blunsom, 2013) . | uses |
2db25254f275303c41f1e7ab15a5e0_0 | However, Sporleder and Lascarides (2008) show that models trained on explicitly marked examples generalize poorly to implicit relation identification. They argued that explicit and implicit examples may be linguistically dissimilar, as writers tend to avoid discourse connectives if the discourse relation could be inferred from context (Grice, 1975) . Similar observations are made by <cite>Rutherford and Xue (2015)</cite> , who attempt to add automatically-labeled instances to improve supervised classification of implicit discourse relations. In this paper, we approach this problem from the perspective of domain adaptation. | motivation background |
2db25254f275303c41f1e7ab15a5e0_1 | <cite>Rutherford and Xue (2015)</cite> explore several selection heuristics for adding automatically-labeled examples from Gigaword to their system for implicit relation detection, obtaining a 2% improvement in Macro-F 1 . Our work differs from these previous efforts in that we focus exclusively on training from automaticallylabeled explicit instances, rather than supplementing a training set of manually-labeled implicit examples. | differences background |
2db25254f275303c41f1e7ab15a5e0_2 | It may also be desirable to ensure that the source and target training instances are similar in terms of their observed features; this is the idea behind the instance weighting approach to domain adaptation (Jiang and Zhai, 2007) . Motivated by this idea, we require that sampled instances from the source domain have a cosine similarity of at least τ with at least one target domain instance<cite> (Rutherford and Xue, 2015)</cite> . | background similarities |
2db25254f275303c41f1e7ab15a5e0_3 | In a pilot study we found that larger amounts of additional training data yielded no further improvements, which is consistent with the recent results of <cite>Rutherford and Xue (2015)</cite> . | similarities |
2db25254f275303c41f1e7ab15a5e0_4 | We have presented two methods -feature representation learning and resampling -from domain adaptation to close the gap of using explicit examples for unsupervised implicit discourse relation identification. Future work will explore the combination of this approach with more sophisticated techniques for instance selection<cite> (Rutherford and Xue, 2015)</cite> and feature selection (Park and Cardie, 2012; Biran and McKeown, 2013) , while also tackling the more difficult problems of multi-class relation classification and fine-grained level-2 discourse relations. | future_work |
2eaa48dbc5e42a5934e905ec2288ac_0 | Although traditional AES methods typically rely on handcrafted features (Larkey, 1998; Foltz et al., 1999; Attali and Burstein, 2006; Dikli, 2006; Wang and Brown, 2008; Chen and He, 2013; Somasundaran et al., 2014; Yannakoudakis et al., 2014; Phandi et al., 2015) , recent results indicate that state-of-the-art deep learning methods reach better performance (Alikaniotis et al., 2016; Dong and Zhang, 2016; Taghipour and Ng, 2016; Song et al., 2017; <cite>Tay et al., 2018</cite>) , perhaps because <cite>these methods</cite> are able to capture subtle and complex information that is relevant to the task (Dong and Zhang, 2016) . | background |
2eaa48dbc5e42a5934e905ec2288ac_1 | The empirical results indicate that our approach yields a better performance than state-of-the-art approaches (Phandi et al., 2015; Dong and Zhang, 2016; <cite>Tay et al., 2018</cite>) . | differences |
2eaa48dbc5e42a5934e905ec2288ac_2 | Since the official test data of the ASAP competition is not released to the public, we, as well as others before us (Phandi et al., 2015; Dong and Zhang, 2016; 1 https://www.kaggle.com/c/asap-aes/data <cite>Tay et al., 2018</cite>) , use only the training data in our experiments. | similarities |
2eaa48dbc5e42a5934e905ec2288ac_3 | We compare our approach with stateof-the-art methods based on handcrafted features (Phandi et al., 2015) , as well as deep features (Dong and Zhang, 2016; <cite>Tay et al., 2018</cite>) . | uses |
2eaa48dbc5e42a5934e905ec2288ac_4 | We used functions from the VLFeat li- Table 2 : In-domain automatic essay scoring results of our approach versus several state-of-the-art methods (Phandi et al., 2015; Dong and Zhang, 2016; <cite>Tay et al., 2018</cite>) . | differences |
2eaa48dbc5e42a5934e905ec2288ac_5 | We first note that the histogram intersection string kernel alone reaches better overall performance (0.780) than all previous works (Phandi et al., 2015; Dong and Zhang, 2016; <cite>Tay et al., 2018</cite>) . | differences |
2eaa48dbc5e42a5934e905ec2288ac_6 | Although the BOSWE model can be regarded as a shallow approach, its overall results are comparable to those of deep learning approaches (Dong and Zhang, 2016; <cite>Tay et al., 2018</cite>) . | similarities |
2eaa48dbc5e42a5934e905ec2288ac_7 | The average QWK score of HISK and BOSWE (0.785) is more than 2% better the average scores of the best-performing state-of-the-art approaches <cite>Tay et al., 2018</cite>) . | differences |
2eaa48dbc5e42a5934e905ec2288ac_8 | We compared our approach on the Automated Student Assessment Prize data set, in both in-domain and crossdomain settings, with several state-of-the-art approaches (Phandi et al., 2015; Dong and Zhang, 2016; <cite>Tay et al., 2018</cite>) . | uses |
2eaa48dbc5e42a5934e905ec2288ac_9 | Using a shallow approach, we report better results compared to recent deep learning approaches (Dong and Zhang, 2016; <cite>Tay et al., 2018</cite>) . | differences |
2ef456a3f6b043350121c4c5cfd404_0 | Hence, an adaptive IS may use a large number of samples to solve this problem whereas NCE is more stable and requires a fixed small number of noise samples (e.g., 100) to achieve a good performance [13, <cite>16]</cite> . | background |
2ef456a3f6b043350121c4c5cfd404_1 | To alleviate this problem, noise samples can be shared across the batch<cite> [16]</cite> . | background |
2ef456a3f6b043350121c4c5cfd404_2 | Furthermore, we can show that this solution optimally approximates the sampling from a unigram distribution, which has been shown to be a good noise distribution choice [13, <cite>16]</cite> . | background |
2ef456a3f6b043350121c4c5cfd404_3 | This can be done by simply drawing an additional K samples form the noise distribution pn, and share them across the batch as it was done in<cite> [16]</cite> . | background |
2ef456a3f6b043350121c4c5cfd404_4 | Each of the models is trained using the proposed B-NCE approach and the shared noise NCE (S-NCE)<cite> [16]</cite> . | uses |
2ef456a3f6b043350121c4c5cfd404_6 | Following the setup proposed in [13, <cite>16]</cite> , S-NCE uses K = 100 noise samples, whereas B-NCE uses only the target words in the batch (K=0). | uses |
2ef456a3f6b043350121c4c5cfd404_7 | Moreover, the performance of the small ReLu-LSTM is comparable to the LSTM models proposed in<cite> [16]</cite> and [18] which use large hidden layers. | background |
2f7b64db6939786a5026fc033c85bd_0 | Until recently, GRE algorithms have focussed on the generation of distinguishing descriptions that are either as short as possible (e.g. (Dale, 1992; Gardent, 2002) ) or almost as short as possible (e.g. <cite>(Dale and Reiter, 1995)</cite> ). | background |
2f7b64db6939786a5026fc033c85bd_1 | allow the Full Brevity algorithm (Dale, 1992) to be viewed as minimising cost(S), and the incremental algorithm <cite>(Dale and Reiter, 1995)</cite> as hill-climbing (strictly, hill-descending), guided by the property-ordering which that algorithm requires. | background |
2f7b64db6939786a5026fc033c85bd_2 | Standard GRE algorithms assume that the speaker knows what the hearer knows <cite>(Dale and Reiter, 1995)</cite> . | background |
2fbf5397a8219923d1d9bc0464cb59_0 | Related work on exploring syntactic structured information in pronoun resolution can be typically classified into three categories: parse tree-based search algorithms ( Hobbs 1978) , feature-based (Lappin and Leass 1994; Bergsma and Lin 2006) and tree kernel-based methods<cite> (Yang et al 2006)</cite> . | background |
2fbf5397a8219923d1d9bc0464cb59_1 | As for tree kernel-based methods, <cite>Yang et al (2006)</cite> captured syntactic structured information for pronoun resolution by using the convolution tree kernel (Collins and Duffy 2001) to measure the common sub-trees enumerated from the parse trees and achieved quite success on the ACE 2003 corpus. | background |
2fbf5397a8219923d1d9bc0464cb59_2 | Compared with Collins and Duffy's kernel and its application in pronoun resolution<cite> (Yang et al 2006)</cite> , the context-sensitive convolution tree kernel enumerates not only context-free sub-trees but also context-sensitive sub-trees by taking their ancestor node paths into consideration. | background |
2fbf5397a8219923d1d9bc0464cb59_3 | To deal with the cases that an anaphor and an antecedent candidate do not occur in the same sentence, we construct a pseudo parse tree for an entire text by attaching the parse trees of all its sentences to an upper "S " node, similar to <cite>Yang et al (2006)</cite> . | similarities |
2fbf5397a8219923d1d9bc0464cb59_4 | Figure 2 shows the three tree span schemes explored in <cite>Yang et al (2006)</cite> : MinExpansion (only including the shortest path connecting the anaphor and the antecedent candidate), Simple-Expansion (containing not only all the nodes in Min-Expansion but also the first level children of these nodes) and Full-Expansion (covering the sub-tree between the anaphor and the candidate), such as the sub-trees inside the dash circles of Figures 2(a) , 2(b) and 2(c) respectively. | background |
2fbf5397a8219923d1d9bc0464cb59_5 | It is found<cite> (Yang et al 2006)</cite> that the simpleexpansion tree span scheme performed best on the ACE 2003 corpus in pronoun resolution. | background |
2fbf5397a8219923d1d9bc0464cb59_6 | This convolution tree kernel has been successfully applied by <cite>Yang et al (2006)</cite> in pronoun resolution. | background |
2fbf5397a8219923d1d9bc0464cb59_7 | Table 1 systematically evaluates the impact of different m in our context-sensitive convolution tree kernel and compares our dynamic-expansion tree span scheme with the existing three tree span schemes, min-, simple-and full-expansions as described in <cite>Yang et al (2006)</cite> . | similarities uses |
2fdfa1b36fcf0d77826c96101ac428_0 | To address the model design issue, we discuss several recent solutions (He et al., 2016b; Li et al., 2016; <cite>Xiong et al., 2017)</cite> . | background |
2fdfa1b36fcf0d77826c96101ac428_1 | To address the model design issue, we discuss several recent solutions (He et al., 2016b; Li et al., 2016; <cite>Xiong et al., 2017)</cite> . We then focus on a new case study of hierarchical deep reinforcement learning for video captioning (Wang et al., 2018b) , discussing the techniques of leveraging hierarchies in DRL for NLP generation problems. | differences |
2fdfa1b36fcf0d77826c96101ac428_2 | We outline the applications of deep reinforcement learning in NLP, including dialog (Li et al., 2016) , semi-supervised text classification (Wu et al., 2018) , coreference (Clark and Manning, 2016; Yin et al., 2018) , knowledge graph reasoning<cite> (Xiong et al., 2017</cite> ), text games (Narasimhan et al., 2015; He et al., 2016a) , social media (He et al., 2016b; Zhou and Wang, 2018) , information extraction (Narasimhan et al., 2016; Qin et al., 2018) , language and vision (Pasunuru and Bansal, 2017; Misra et al., 2017; Wang et al., 2018a,b,c; Xiong et al., 2018) , etc. | background |
2fdfa1b36fcf0d77826c96101ac428_3 | To address the model design issue, we discuss several recent solutions (He et al., 2016b; Li et al., 2016; <cite>Xiong et al., 2017)</cite> . | background |
2fdfa1b36fcf0d77826c96101ac428_4 | To address the model design issue, we discuss several recent solutions (He et al., 2016b; Li et al., 2016; <cite>Xiong et al., 2017)</cite> . We then focus on a new case study of hierarchical deep reinforcement learning for video captioning (Wang et al., 2018b) , discussing the techniques of leveraging hierarchies in DRL for NLP generation problems. | differences |
304773c64de1f0906f0246f2aa0d29_0 | To extract opinion targets, pervious approaches usually relied on opinion words which are the words used to express the opinions (Hu and Liu, 2004a; Popescu and Etzioni, 2005; Liu et al., 2005; Wang and Wang, 2008; Qiu et al., 2011;<cite> Liu et al., 2012)</cite> . | background |
304773c64de1f0906f0246f2aa0d29_1 | To resolve these problems,<cite> Liu et al. (2012)</cite> formulated identifying opinion relations between words as an monolingual alignment process. | background |
304773c64de1f0906f0246f2aa0d29_2 | Although <cite>(Liu et al., 2012)</cite> had proved the effectiveness of WAM, they mainly performed experiments on the dataset with medium size. | motivation |
304773c64de1f0906f0246f2aa0d29_3 | <cite>(Liu et al., 2012)</cite> formulated identifying opinion relations between words as an alignment process. | background |
304773c64de1f0906f0246f2aa0d29_4 | We notice these two methods ( <cite>(Liu et al., 2012)</cite> and (Liu et al., 2013) ) only performed experiments on the corpora with a medium size. | motivation |
304773c64de1f0906f0246f2aa0d29_5 | To extract opinion targets from reviews, we adopt the framework proposed by <cite>(Liu et al., 2012)</cite> , which is a graph-based extraction framework and has two main components as follows. | uses |
304773c64de1f0906f0246f2aa0d29_6 | In this paper, we assume opinion targets to be nouns or noun phrases, and opinion words may be adjectives or verbs, which are usually adopted by (Hu and Liu, 2004a; Qiu et al., 2011; Wang and Wang, 2008;<cite> Liu et al., 2012)</cite> . | similarities |
304773c64de1f0906f0246f2aa0d29_7 | Similar to <cite>(Liu et al., 2012)</cite> , every sentence in reviews is replicated to generate a parallel sentence pair, and the word alignment algorithm is applied to the monolingual scenario to align a noun/noun phase with its modifiers. | uses |
304773c64de1f0906f0246f2aa0d29_8 | Then, similar to <cite>(Liu et al., 2012)</cite> , the association between an opinion target candidate and its modifier is estimated as follows. | uses |
304773c64de1f0906f0246f2aa0d29_9 | In the second component, we adopt a graph-based algorithm used in <cite>(Liu et al., 2012)</cite> to compute the confidence of each opinion target candidate, and the candidates with higher confidence than the threshold will be extracted as the opinion targets. | uses |
304773c64de1f0906f0246f2aa0d29_10 | Similar to <cite>(Liu et al., 2012)</cite> , we set each item in , where tf (v) is the term frequency of v in the corpus, and df (v) is computed by using the Google n-gram corpus 2 . | uses |
304773c64de1f0906f0246f2aa0d29_11 | In this section, to answer the questions mentioned in the first section, we collect a large collection named as LARGE, which includes reviews from three different domains and different languages. This collection was also used in <cite>(Liu et al., 2012)</cite> . | similarities |
304773c64de1f0906f0246f2aa0d29_12 | To further prove the effectiveness of our combination, we compare PSWAM with some state-of-the-art methods, including Hu (Hu and Liu, 2004a) , which extracted frequent opinion target words based on association mining rules, DP (Qiu et al., 2011) , which extracted opinion targets through syntactic patterns, and LIU <cite>(Liu et al., 2012)</cite> , which fulfilled this task by using unsupervised WAM. | uses |
304773c64de1f0906f0246f2aa0d29_13 | To further prove the effectiveness of our combination, we compare PSWAM with some state-of-the-art methods, including Hu (Hu and Liu, 2004a) , which extracted frequent opinion target words based on association mining rules, DP (Qiu et al., 2011) , which extracted opinion targets through syntactic patterns, and LIU <cite>(Liu et al., 2012)</cite> , which fulfilled this task by using unsupervised WAM. The parameter settings in these baselines are the same as the settings in the original papers. | uses |
30718e751f18432c2478442530267e_0 | According to<cite> Jia and Liang (2017)</cite> , the single BiDAF system (Seo et al., 2016) only achieves an F1 score of 4.8 on the ADDANY adversarial dataset. | background |
30718e751f18432c2478442530267e_1 | According to<cite> Jia and Liang (2017)</cite> , the single BiDAF system (Seo et al., 2016) only achieves an F1 score of 4.8 on the ADDANY adversarial dataset. In this paper, we present a method to tackle this problem via answer sentence selection. | motivation |
30718e751f18432c2478442530267e_2 | However,<cite> Jia and Liang (2017)</cite> show that these systems are very vulnerable to paragraphs with adversarial sentences. | background |
30718e751f18432c2478442530267e_3 | Besides the single BiDAF, the single Match LSTM, the ensemble Match LSTM, and the ensemble BiDAF achieve an F1 of 7.6, 11.7, and 2.7 respectively in question answering on ADDANY adversarial dataset<cite> (Jia and Liang, 2017)</cite> . | background |
30718e751f18432c2478442530267e_4 | Besides the single BiDAF, the single Match LSTM, the ensemble Match LSTM, and the ensemble BiDAF achieve an F1 of 7.6, 11.7, and 2.7 respectively in question answering on ADDANY adversarial dataset<cite> (Jia and Liang, 2017)</cite> . Therefore, question answering with adversarial sentences in paragraphs is a prominent issue and is the focus of this study. | background motivation |
30718e751f18432c2478442530267e_5 | Our test set is<cite> Jia and Liang (2017)</cite>'s ADDANY adversarial dataset. | uses |
30718e751f18432c2478442530267e_6 | The performance of question answering is evaluated by the Macro-averaged F1 score (Rajpurkar <cite>Jia and Liang, 2017)</cite> . | uses |
30718e751f18432c2478442530267e_7 | However, following the idea of adversarial examples in image recognition (Goodfellow et al., 2014; Kurakin et al., 2016; Papernot et al., 2016) ,<cite> Jia and Liang (2017)</cite> point out the unreliability of existing question answering models in the presence of adversarial sentences. | background |
30718e751f18432c2478442530267e_8 | However, following the idea of adversarial examples in image recognition (Goodfellow et al., 2014; Kurakin et al., 2016; Papernot et al., 2016) ,<cite> Jia and Liang (2017)</cite> point out the unreliability of existing question answering models in the presence of adversarial sentences. In this study, we propose a method to tackle this problem through answer sentence selection. | background motivation |
30718e751f18432c2478442530267e_9 | However,<cite> Jia and Liang (2017)</cite> also present the deterioration of QA systems on another dataset, ADDSENT adversarial dataset. | similarities |
311b238406da4891c09cb9c3c0334d_0 | This makes the task more difficult, compared to the sentiment analysis, but it can often bring complementary information <cite>[3]</cite> . | background |
311b238406da4891c09cb9c3c0334d_1 | We preprocessed the Czech commentaries by the same rules as in the original system <cite>[3]</cite> (for example: all urls were replaced by keyword URL, links to images are replaced by IMGURL, only letters are preserved, the rest of the characters is removed, …). | uses |
311b238406da4891c09cb9c3c0334d_2 | The original system <cite>[3]</cite> used more features, which could not be easily applied on Czech commentaries. | differences |
311b238406da4891c09cb9c3c0334d_3 | We did not identify strong candidates to build a domain specific dictionary as in <cite>[3]</cite> . | differences |
3188ee1583a9c711cf147fc596768d_0 | The techniques examined are Structural Correspondence Learning (SCL)<cite> (Blitzer et al., 2006)</cite> and Self-training (Abney, 2007; McClosky et al., 2006) . | background |
3188ee1583a9c711cf147fc596768d_1 | We examine Structural Correspondence Learning (SCL)<cite> (Blitzer et al., 2006)</cite> for this task, and compare it to several variants of Self-training (Abney, 2007; McClosky et al., 2006) . | similarities |
3188ee1583a9c711cf147fc596768d_2 | So far, Structural Correspondence Learning has been applied successfully to PoS tagging and Sentiment Analysis<cite> (Blitzer et al., 2006</cite>; ). | background |
3188ee1583a9c711cf147fc596768d_3 | Structural Correspondence Learning<cite> (Blitzer et al., 2006)</cite> exploits unlabeled data from both source and target domain to find correspondences among features from different domains. | background |
3188ee1583a9c711cf147fc596768d_4 | Pivots are features occurring frequently and behaving similarly in both domains<cite> (Blitzer et al., 2006)</cite> . | background |
3188ee1583a9c711cf147fc596768d_5 | Intuitively, if we are able to find good correspondences through 'linking' pivots, then the augmented source data should transfer better to a target domain<cite> (Blitzer et al., 2006)</cite> . | similarities |
3188ee1583a9c711cf147fc596768d_6 | So far, pivot features on the word level were used<cite> (Blitzer et al., 2006</cite>; . | background |
3188ee1583a9c711cf147fc596768d_7 | In our empirical setup, we follow<cite> Blitzer et al. (2006)</cite> and balance the size of source and target data. | similarities uses |
3188ee1583a9c711cf147fc596768d_8 | The paper compares Structural Correspondence Learning<cite> (Blitzer et al., 2006)</cite> with (various instances of) self-training (Abney, 2007; McClosky et al., 2006) for the adaptation of a parse selection model to Wikipedia domains. | similarities |
31b06dfc081149e1e436f0bb5e0904_0 | As a global trend, we observe that models that incorporate rich global features are typically more accurate, even if pruning is necessary or decoding needs to be approximate Koo and Collins, 2010; Bohnet and Nivre, 2012; Martins et al., 2009<cite> Martins et al., , 2013</cite> . | motivation background |
31b06dfc081149e1e436f0bb5e0904_1 | The parser was built as an extension of a recent dependency parser, TurboParser (Martins et al., 2010<cite> (Martins et al., , 2013</cite> , with the goal of performing semantic parsing using any of the three formalisms considered in the shared task (DM, PAS, and PSD). | uses |
31b06dfc081149e1e436f0bb5e0904_2 | Most of these features were taken from TurboParser <cite>(Martins et al., 2013)</cite> , and others were inspired by the semantic parser of Johansson and Nugues (2008) . | uses |
31e8c524f05495fdd87bfac6fbecc8_0 | We present a reproduction and extension to the work of <cite>Schulder et al. (2017)</cite> , <cite>which</cite> introduced a lexicon of verbal polarity shifters, as well as methods to increase the size of this lexicon through bootstrapping. | extends uses |