id
stringlengths 32
33
| x
stringlengths 41
1.75k
| y
stringlengths 4
39
|
|---|---|---|
28038a4fa4182ccdc6134f2138c0da_12 | Perplexity measures for <cite>Noraset et al. (2017)</cite> and Gadetsky et al. (2018) are taken from the authors' respective publications. All our models perform better than previous proposals, by a margin of 4 to 10 points, for a relative improvement of 11-23%. | differences |
28038a4fa4182ccdc6134f2138c0da_13 | A manual analysis of definitions produced by our system reveals issues similar to those discussed by <cite>Noraset et al. (2017)</cite> , namely selfreference, 7 POS-mismatches, over-and underspecificity, antonymy, and incoherence. | similarities |
28038a4fa4182ccdc6134f2138c0da_14 | As for POS-mismatches, we do note that the work of <cite>Noraset et al. (2017)</cite> had a much lower rate of 4.29%: we suggest that this may be due to the fact that they employ a learned character-level convolutional network, which arguably would be able to capture orthography and rudiments of morphology. | differences |
291a6ac3f0c2d27ca69ee8f5f266f5_0 | This paper proposes an expansion of set of primitive constraints available within the Primitive Optimality Theory framework <cite>(Eisner, 1997a)</cite> . | uses |
291a6ac3f0c2d27ca69ee8f5f266f5_1 | This paper proposes an expansion of set of primitive constraints available within the Primitive Optimality Theory framework <cite>(Eisner, 1997a)</cite> . This expansion consists of the addition of a new family of constraints--existential implicational constraints, which allow the specification of faithfulness constraints that can be satisfied at a distance--and the definition of two ways to combine simple constraints into com: plex constraints, that is, constraint disjunction (Crowhurst and Hewitt, 1995) and local constraint conjunction (Smolensky, 1995) . | extends |
291a6ac3f0c2d27ca69ee8f5f266f5_2 | Primitive Optimality Theory (OTP) <cite>(Eisner, 1997a)</cite> , and extensions to it (e.g., Albro (1998) ), can be useful as a formal system in which phonological analyses can be implemented and evaluated. | background |
291a6ac3f0c2d27ca69ee8f5f266f5_3 | Primitive Optimality Theory (OTP) <cite>(Eisner, 1997a)</cite> , and extensions to it (e.g., Albro (1998) ), can be useful as a formal system in which phonological analyses can be implemented and evaluated. However, for certain types of constraints, translation into the primitives of OTP (Eisner (1997b) ) can only be accomplished by adding to the grammar a number of ad hoc phonological tiers. | motivation |
291a6ac3f0c2d27ca69ee8f5f266f5_4 | This paper looks at three types of constraints employed throughout the Optimality Theoretic literature that cannot be translated in to the 1The computation time for an Optimality Theoretic derivation within the implementation of Albro (1998) increases exponentially with the number of tiers. The same is true for the implementation described in<cite> Eisner (1997a)</cite> , although a proposal is given there for a method that might improve the situation. | uses motivation |
291a6ac3f0c2d27ca69ee8f5f266f5_5 | primitives of OTP without reference to ad hoc tiers, and proposes a formalization of these constraints that is compatible with the finite state model described in<cite> Eisner (1997a)</cite> and Albro (1998) . | background |
291a6ac3f0c2d27ca69ee8f5f266f5_6 | 2 Existential Implication 2.1 Motivation OWP as described in<cite> Eisner (1997a)</cite> provides some support for correspondence constraints (input-output only). | background |
291a6ac3f0c2d27ca69ee8f5f266f5_7 | Using the FST notation of<cite> Eisner (1997a)</cite> , the implementation for this constraint would be the following FST: | uses |
29294f2ed3cc2772ca57fd4294274c_0 | <cite>Leuski et al. (2006)</cite> developed algorithms for training such characters using linked questions and responses in the form of unstructured natural language text. | background |
29294f2ed3cc2772ca57fd4294274c_1 | These algorithms have been incorporated into a tool which has been used to create characters for a variety of applications (e.g.<cite> Leuski et al., 2006</cite>; Artstein et al., 2009; Swartout et al., 2010) . | background |
29294f2ed3cc2772ca57fd4294274c_2 | We reimplemented parts of the response ranking algorithms of <cite>Leuski et al. (2006)</cite> , including both the language modeling (LM) and cross-language modeling (CLM) approaches. | extends differences |
29294f2ed3cc2772ca57fd4294274c_3 | We did not implement the parameter learning of <cite>Leuski et al. (2006)</cite> ; instead we use a constant smoothing parameter λ π = λ φ = 0.1. | differences |
29294f2ed3cc2772ca57fd4294274c_4 | We also do not use the response threshold parameter, which <cite>Leuski et al. (2006)</cite> use to determine whether the top-ranked response is good enough. | differences |
29294f2ed3cc2772ca57fd4294274c_5 | This measure does not take into account non-understanding, that is the classifier's determination that the best response is not good enough<cite> (Leuski et al., 2006)</cite> , since this capability was not implemented; however, since all of our test questions are known to have at least one appropriate response, any non-understanding of a question would necessarily count against accuracy anyway. | differences background |
29294f2ed3cc2772ca57fd4294274c_6 | The LM approach almost invariably produced better results than the CLM approach; this is the opposite of the findings of <cite>Leuski et al. (2006)</cite> , where CLM fared consistently better. | differences |
29294f2ed3cc2772ca57fd4294274c_7 | In our experiments the LM approach consistently outperforms the CLM approach, contra <cite>Leuski et al. (2006)</cite> . | differences |
2a01f96893f9c0630a01ecce320184_0 | Several research works have been proposed to detect propaganda on document-level (Rashkin et al., 2017; Barrón-Cedeño et al., 2019b) , sentencelevel and fragment-level <cite>(Da San Martino et al., 2019)</cite> . | background |
2a01f96893f9c0630a01ecce320184_1 | Although Da San<cite> Martino et al. (2019)</cite> indicates that multi-task learning of both the SLC and the FLC could be beneficial for the SLC, in this paper, we only focus on the SLC task so as to better investigate whether context information could improve the performance of our system. | differences |
2a01f96893f9c0630a01ecce320184_2 | A fine-grained propaganda corpus was proposed in Da San<cite> Martino et al. (2019)</cite> which includes both sentencelevel and fragment-level information. | background |
2a01f96893f9c0630a01ecce320184_3 | More details of the dataset could be found in Da San<cite> Martino et al. (2019)</cite> . | background |
2a01f96893f9c0630a01ecce320184_4 | As described in Da San<cite> Martino et al. (2019)</cite> , the source of the dataset that we use is news articles, and since the title is usually the summarization of a news article, we use the title as supplementary information. | uses background |
2a01f96893f9c0630a01ecce320184_5 | In the future, we plan to apply multi-task learning to this context-dependent BERT, similar to the method mentioned in Da San<cite> Martino et al. (2019)</cite> or introducing other kinds of tasks, such as sentiment analysis or domain classification. | similarities future_work |
2a84615479af66bbf875517a3a753b_0 | In our previous work <cite>[7]</cite> , we applied a dual RNN in order to obtain a richer representation by blending the content and acoustic knowledge. | background |
2a84615479af66bbf875517a3a753b_1 | In our previous work <cite>[7]</cite> , we applied a dual RNN in order to obtain a richer representation by blending the content and acoustic knowledge. In this paper, we improve upon our earlier work by incorporating an attention mechanism in the emotion recognition framework. | extends |
2a84615479af66bbf875517a3a753b_2 | Recently,<cite> [7,</cite> 18] combined acoustic information and conversation transcripts using a neural network-based model to improve emotion classification accuracy. | background |
2a84615479af66bbf875517a3a753b_3 | Recently,<cite> [7,</cite> 18] combined acoustic information and conversation transcripts using a neural network-based model to improve emotion classification accuracy. However, none of these studies utilized attention method over audio and text modality in tandem for contextual understanding of the emotion in audio recording. | background motivation |
2a84615479af66bbf875517a3a753b_4 | Motivated by the architecture used in<cite> [7,</cite> 17, 19] , we train a recurrent encoder to predict the categorical class of a given audio signal. | motivation |
2a84615479af66bbf875517a3a753b_5 | To follow previous research <cite>[7]</cite> , we also add another prosodic feature vector, p, with each ot to generate a more informative vector representation of the signal, o A t . | uses |
2a84615479af66bbf875517a3a753b_6 | Previous research used multi-modal information independently using neural network model by concatenating features from each modality<cite> [7,</cite> 21] . | background |
2a84615479af66bbf875517a3a753b_7 | Previous research used multi-modal information independently using neural network model by concatenating features from each modality<cite> [7,</cite> 21] . As opposed to this approach, we propose a neural network architecture that exploits information in each modality by extracting relevant segments of the speech data using information from the lexical content (and vice-versa). | differences |
2a84615479af66bbf875517a3a753b_8 | For consistent comparison with previous works<cite> [7,</cite> 18] , all utterances labeled "excitement" are merged with those labeled "happiness". | uses |
2a84615479af66bbf875517a3a753b_9 | As this research is extended work from previous research <cite>[7]</cite> , we use the same feature extraction method as done in our previous work. | extends |
2a84615479af66bbf875517a3a753b_10 | We use the same dataset and features as other researchers<cite> [7,</cite> 18] . | uses |
2a84615479af66bbf875517a3a753b_11 | In audio-BRE (Fig. 2(a) ), most of the emotion labels are frequently misclassified as neutral class, supporting the claims of<cite> [7,</cite> 25] . | similarities |
2b10893f03b4f5eaac0fe06b4d6115_0 | In order to compare the performance of our system with others, we also used the dataset of<cite> Tu and Roth (2012)</cite> , which contains 1,348 sentences taken from different parts of the British National Corpus. | uses |
2b10893f03b4f5eaac0fe06b4d6115_1 | One example is<cite> Tu and Roth (2012)</cite> , where the authors examined a verbparticle combination only if the verbal components were formed with one of the previously given six verbs (i.e. make, take, have, give, do, get). | background |
2b10893f03b4f5eaac0fe06b4d6115_2 | As Table 3 shows, the six verbs used by<cite> Tu and Roth (2012)</cite> are responsible for only 50 VPCs on the Wiki50 corpus, so it covers only 11.16% of all gold standard VPCs. | background |
2b10893f03b4f5eaac0fe06b4d6115_3 | Furthermore, 127 different verbal component occurred in Wiki50, but the verbs have and do -which are used by<cite> Tu and Roth (2012)</cite> -do not appear in the corpus as verbal component of VPCs. | background |
2b10893f03b4f5eaac0fe06b4d6115_4 | Moreover, Support Vector Machines (SVM) (Cortes and Vapnik, 1995) results are also reported to compare the performance of our methods with that of<cite> Tu and Roth (2012)</cite> . | uses |
2b10893f03b4f5eaac0fe06b4d6115_5 | As<cite> Tu and Roth (2012)</cite> presented only the accuracy scores on the Tu & Roth dataset, we also employed an accuracy score as an evaluation metric on this dataset, where positive and negative examples were also marked. | similarities |
2b10893f03b4f5eaac0fe06b4d6115_6 | We also compared our results with the rule-based results available for Wiki50 and also with the 5-fold cross validation results of<cite> Tu and Roth (2012)</cite> . | uses |
2b10893f03b4f5eaac0fe06b4d6115_7 | In order to compare the performance of our system with others, we evaluated it on the Tu&Roth dataset <cite>(Tu and Roth, 2012)</cite> . | uses |
2b10893f03b4f5eaac0fe06b4d6115_8 | over, it also lists the results of<cite> Tu and Roth (2012)</cite> and the VPCTagger evaluated in the 5-fold cross validation manner, as<cite> Tu and Roth (2012)</cite> applied this evaluation schema. | uses |
2b10893f03b4f5eaac0fe06b4d6115_9 | Moreover, the results obtained with our machine learning approach on the Tu&Roth dataset outperformed those reported in<cite> Tu and Roth (2012)</cite> . | differences |
2b10893f03b4f5eaac0fe06b4d6115_10 | A striking difference between the Tu & Roth database and Wiki50 is that while<cite> Tu and Roth (2012)</cite> included the verbs do and have in their data, they do not occur at all among the VPCs collected from Wiki50. | background |
2b10893f03b4f5eaac0fe06b4d6115_11 | Our method yielded better results than those got using the dependency parsers on the Wiki50 corpus and the method reported in <cite>(Tu and Roth, 2012)</cite> on the Tu&Roth dataset. | differences |
2b148e376c39eae7f674610118e588_0 | In this paper, we consider the referential games of <cite>Lazaridou et al. (2017)</cite> , and investigate the representations the agents develop during their evolving interaction. | motivation |
2b148e376c39eae7f674610118e588_1 | Unlike earlier work (e.g., Briscoe, 2002; Cangelosi and Parisi, 2002; Steels, 2012) , many recent simulations consider realistic visual input, for example, by playing referential games with real-life pictures (e.g., Jorge et al., 2016; <cite>Lazaridou et al., 2017</cite>; Havrylov and Titov, 2017; Lee et al., 2018; Evtimova et al., 2018) . This setup allows us to address the exciting issue of whether the needs of goal-directed communication will lead agents to associate visually-grounded conceptual representations to discrete symbols, developing naturallanguage-like word meanings. | motivation background |
2b148e376c39eae7f674610118e588_2 | We study here agent representations following the model and setup of <cite>Lazaridou et al. (2017)</cite> . | motivation |
2b148e376c39eae7f674610118e588_3 | In their first game, <cite>Lazaridou</cite>'s Sender and Receiver are exposed to the same pair of images, one of them being randomly marked as the "target". | background |
2b148e376c39eae7f674610118e588_4 | Since an analysis of vocabulary usage brings inconclusive evidence that the agents are using the symbols to represent natural concepts (such as beaver or bayonet), <cite>Lazaridou and colleagues</cite> next modify the game, by presenting to the Sender and the Receiver different images for each of the two concepts (e.g., the Sender must now signal that the target is a beaver, while seeing a different beaver from the one shown to the Receiver). | background |
2b148e376c39eae7f674610118e588_5 | <cite>Lazaridou and colleagues</cite> present preliminary evidence suggesting that, indeed, agents are now developing conceptual symbol meanings. | background |
2b148e376c39eae7f674610118e588_6 | We replicate <cite>Lazaridou</cite>'s games, and we find that, in both, the agents develop successfully aligned representations that, however, are not capturing conceptual properties at all. | uses motivation |
2b148e376c39eae7f674610118e588_7 | Architecture We re-implement <cite>Lazaridou</cite>'s Sender and Receiver architectures (using their better-behaved "informed" Sender). | uses |
2b148e376c39eae7f674610118e588_8 | See <cite>Lazaridou et al. (2017</cite>) for details. | background |
2b148e376c39eae7f674610118e588_9 | Data Following <cite>Lazaridou et al. (2017)</cite> , for each of the 463 concepts <cite>they</cite> used, we randomly sample 100 images from ImageNet (Deng et al., 2009 ). | uses similarities |
2b148e376c39eae7f674610118e588_10 | Following <cite>Lazaridou</cite>, the images are passed through a pre-trained VGG ConvNet (Simonyan and Zisserman, 2015) . | similarities uses |
2b148e376c39eae7f674610118e588_11 | Games We re-implement both <cite>Lazaridou</cite>'s same-image game, where Sender and Receiver are shown the same two images (always of different concepts), and their different-image game, where the Receiver sees different images than the Sender's. | uses |
2b148e376c39eae7f674610118e588_12 | As we faithfully reproduced the setup of <cite>Lazaridou et al. (2017)</cite> , we refer the reader there for hyper-parameters and training details. | similarities background |
2b148e376c39eae7f674610118e588_13 | <cite>Lazaridou et al. (2017)</cite> designed <cite>their</cite> second game to encourage more general, concept-like referents. Unfortunately, we replicate the anomalies above in the different-image setup, although to a less marked extent. | background similarities |
2b148e376c39eae7f674610118e588_14 | However, the important contribution of <cite>Lazaridou et al. (2017)</cite> is to play a signaling game with real-life images instead of artificial symbols. This raises new empirical questions that are not answered by the general mathematical results, such as: When the agents do succeed at communicating, what are the input features they rely upon? | motivation |
2b6dd9388c43df4416c738b2d1ed5f_0 | In this work, we use the datasets released by <cite>(Davidson et al. 2017 )</cite> and HEOT dataset provided by (Mathur et al. 2018) . | uses |
2b6dd9388c43df4416c738b2d1ed5f_1 | The embeddings were trained on both the datasets provided by <cite>(Davidson et al. 2017 )</cite> and HEOT. | uses |
2b6dd9388c43df4416c738b2d1ed5f_2 | As indicated by the Figure 1 , the model was initially trained on the dataset provided by <cite>(Davidson et al. 2017)</cite> , and then re-trained on the HEOT dataset so as to benefit from the transfer of learned features in the last stage. | uses |
2b6dd9388c43df4416c738b2d1ed5f_3 | For comparison purposes, in Table 4 we have also evaluated our results on the dataset by <cite>(Davidson et al. 2017 )</cite>. | uses |
2b6dd9388c43df4416c738b2d1ed5f_4 | Both the HEOT and <cite>(Davidson et al. 2017 )</cite> datasets contain tweets which are annotated in three categories: offensive, abusive and none (or benign). | background |
2b6dd9388c43df4416c738b2d1ed5f_5 | Both the HEOT and <cite>(Davidson et al. 2017 )</cite> datasets contain tweets which are annotated in three categories: offensive, abusive and none (or benign). We use a LSTM based classifier model for training our model to classify these tweets into these three categories. | uses |
2b6dd9388c43df4416c738b2d1ed5f_6 | Results Table 3 shows the performance of our model (after getting trained on <cite>(Davidson et al. 2017)</cite> ) with two types of embeddings in comparison to the models by (Mathur et al. 2018) and <cite>(Davidson et al. 2017 )</cite> on the HEOT dataset averaged over three runs. | uses similarities differences |
2b7267b7b192aeca15c0d10a5f0a4b_0 | An important work that has relevance here is <cite>[8]</cite> where authors present an even larger movie review dataset of 50,000 movie reviews from IMBD. | background |
2b7267b7b192aeca15c0d10a5f0a4b_1 | In <cite>[8]</cite> for example, authors who created movie review dataset try on it their probabilistic model that is able to capture semantic similarities between words. | background |
2b7267b7b192aeca15c0d10a5f0a4b_2 | Our scores on this task are somehow lower than those reported from various studies that explore advanced deep learning constructs on same dataset. In <cite>[8]</cite> for example, authors who created movie review dataset try on it their probabilistic model that is able to capture semantic similarities between words. | differences |
2bb41cea97a0375f67eab3a77c3a97_0 | Traditional relation-extraction systems rely on manual annotations or domain-specific rules provided by experts, both of which are scarce resources that are not portable across domains. To remedy these problems, recent years have seen interest in the distant supervision approach for relation extraction (Wu and Weld, 2007; <cite>Mintz et al., 2009)</cite> . | motivation |
2bb41cea97a0375f67eab3a77c3a97_1 | While the largest corpus (Wikipedia and New York Times) employed by recent work on distant supervision<cite> (Mintz et al., 2009</cite>; Hoffmann et al., 2011) contain about 2M documents, we run experiments on a 100M-document (50X more) corpus drawn from ClueWeb. | background |
2bb41cea97a0375f67eab3a77c3a97_2 | Since<cite> Mintz et al. (2009)</cite> coined the name "distant supervision," there has been growing interest in this technique. | background |
2bb41cea97a0375f67eab3a77c3a97_3 | At each step of the distant supervision process, we closely follow the recent literature<cite> (Mintz et al., 2009</cite>; . | similarities |
2bb41cea97a0375f67eab3a77c3a97_4 | Following recent work<cite> (Mintz et al., 2009</cite>; Hoffmann et al., 2011) , we use Freebase 5 as the knowledge base for seed facts. | similarities uses |
2bb41cea97a0375f67eab3a77c3a97_5 | As in previous work, we impose the constraint that both mentions (m 1 , m 2 ) ∈ R + i are contained in the same sentence<cite> (Mintz et al., 2009</cite>; Hoffmann et al., 2011) . | similarities uses |
2bb41cea97a0375f67eab3a77c3a97_6 | To generate negative examples for each relation, we follow the assumption in<cite> Mintz et al. (2009)</cite> that relations are disjoint and sample from other relations, i.e., R | similarities uses |
2bb41cea97a0375f67eab3a77c3a97_7 | Following recent work on distant supervision<cite> (Mintz et al., 2009</cite>; Hoffmann et al., 2011) , we use both lexical and syntactic features. | similarities uses |
2bb41cea97a0375f67eab3a77c3a97_8 | Interestingly, the Freebase held-out metric<cite> (Mintz et al., 2009</cite>; Hoffmann et al., 2011 ) turns out to be heavily biased toward distantly labeled data (e.g., increasing human feedback hurts precision; see Section 4.6). | differences |
2bb41cea97a0375f67eab3a77c3a97_9 | In addition to the TAC-KBP benchmark, we also follow prior work<cite> (Mintz et al., 2009</cite>; Hoffmann et al., 2011) and measure the quality using held-out data from Freebase. | differences |
2c3a2999390b82f4e29b00d59f90f2_0 | The most frequently applied technique in the CoNLL-2003 shared task is the Maximum Entropy Model. Three systems used Maximum Entropy Models in isolation (Bender et al., 2003; Chieu and Ng, 2003; Curran and Clark, 2003) . Two more systems used them in combination with other techniques<cite> (Florian et al., 2003</cite>; Klein et al., 2003) . | background |
2c3a2999390b82f4e29b00d59f90f2_1 | Hidden Markov Models were employed by four of the systems that took part in the shared task<cite> (Florian et al., 2003</cite>; Klein et al., 2003; Mayfield et al., 2003; Whitelaw and Patrick, 2003) . | background |
2c3a2999390b82f4e29b00d59f90f2_2 | Zhang and Johnson (2003) used robust risk minimization, which is a Winnow technique. <cite>Florian et al. (2003)</cite> employed the same technique in a combination of learners. | background |
2c3a2999390b82f4e29b00d59f90f2_3 | Transformation-based learning<cite> (Florian et al., 2003)</cite> , Support Vector Machines (Mayfield et al., 2003) and Conditional Random Fields (McCallum and Li, 2003) were applied by one system each. | background |
2c3a2999390b82f4e29b00d59f90f2_4 | <cite>Florian et al. (2003)</cite> tested different methods for combining the results of four systems and found that robust risk minimization worked best. | background |
2c3a2999390b82f4e29b00d59f90f2_5 | One participating team has used externally trained named entity recognition systems for English as a part in a combined system<cite> (Florian et al., 2003)</cite> . with extra information compared to while using only the available training data. | background |
2c3a2999390b82f4e29b00d59f90f2_6 | The inclusion of extra named entity recognition systems seems to have worked well<cite> (Florian et al., 2003)</cite> . | background |
2c3a2999390b82f4e29b00d59f90f2_7 | For English, the combined classifier of <cite>Florian et al. (2003)</cite> achieved the highest overall F β=1 rate. | background |
2c3a2999390b82f4e29b00d59f90f2_8 | <cite>Florian et al. (2003)</cite> have also obtained the highest F β=1 rate for the German data. | background |
2c3a2999390b82f4e29b00d59f90f2_9 | A majority vote of five systems (Chieu and Ng, 2003;<cite> Florian et al., 2003</cite>; Klein et al., 2003; McCallum and Li, 2003; Whitelaw and Patrick, 2003) performed best on the English development data. | background |
2c3a2999390b82f4e29b00d59f90f2_10 | The best performance for both languages has been obtained by a combined learning system that used Maximum Entropy Models, transformation-based learning, Hidden Markov Models as well as robust risk minimization<cite> (Florian et al., 2003)</cite> . | background |
2cedb1a0f0c0fbb9bd95d5b54e4967_0 | Only few approaches have attempted comprehension on multiparty dialogue <cite>Ma, Jurczyk, and Choi [2018]</cite> . | motivation background |
2cedb1a0f0c0fbb9bd95d5b54e4967_1 | Inspired by various options of analytic models and the potential of the dialogue processing market, we extend the corpus presented by <cite>Ma, Jurczyk, and Choi [2018]</cite> for comprehensive predictions of personal entities in multiparty dialogue and develop deep learning models to make robust inference on their contexts. | uses |
2cedb1a0f0c0fbb9bd95d5b54e4967_2 | Distinguished from the previous work that only focused on a single variable per passage <cite>Ma, Jurczyk, and Choi [2018]</cite> , we propose two new passage completion tasks on multiparty dialogue which increase the task complexity by replacing more character mentions with variables with a better motivated data split. | extends |
2cedb1a0f0c0fbb9bd95d5b54e4967_3 | Unlike the above tasks where documents and queries are written in a similar writing style, the multiparty dialogue reading comprehension task introduced by <cite>Ma, Jurczyk, and Choi [2018]</cite> has a very different writing style between dialogues and queries. | background |
2cedb1a0f0c0fbb9bd95d5b54e4967_4 | Plot summaries of all episodes for the first eight seasons were collected by Jurczyk and Choi [2017] to evaluate a document retrieval task. The rest of the plot summaries were collected by <cite>Ma, Jurczyk, and Choi [2018]</cite> . | background |