NAACL 2019 Papers

While evaluating an answer choice for Reading Comprehension task, other answer choices available for the question and the answers of related questions about the same paragraph often provide valuable information. In this paper, we propose a method to leverage the natural language relations between the answer choices, such as entailment and contradiction, to improve the performance of machine comprehension. We use a stand-alone question answering (QA) system to perform QA task and a Natural Language Inference (NLI) system to identify the relations between the choice pairs. Then we perform inference using an Integer Linear Programming (ILP)-based relational framework to re-evaluate the decisions made by the standalone QA system in light of the relations identified by the NLI system. We also propose a multitask learning model that learns both the tasks jointly.

To improve the training efficiency of hierarchical recurrent models without compromising their performance, we propose a strategy named as “the lower the simpler”, which is to simplify the baseline models by making the lower layers simpler than the upper layers. We carry out this strategy to simplify two typical hierarchical recurrent models, namely Hierarchical Recurrent Encoder-Decoder (HRED) and R-NET, whose basic building block is GRU. Specifically, we propose Scalar Gated Unit (SGU), which is a simplified variant of GRU, and use it to replace the GRUs at the middle layers of HRED and R-NET. Besides, we also use Fixed-size Ordinally-Forgetting Encoding (FOFE), which is an efficient encoding method without any trainable parameter, to replace the GRUs at the bottom layers of HRED and R-NET. The experimental results show that the simplified HRED and the simplified R-NET contain significantly less trainable parameters, consume significantly less training time, and achieve slightly better performance than their baseline models.

Neural encoder-decoder models have been successful in natural language generation tasks. However, real applications of abstractive summarization must consider an additional constraint that a generated summary should not exceed a desired length. In this paper, we propose a simple but effective extension of a sinusoidal positional encoding (Vaswani et al., 2017) so that a neural encoder-decoder model preserves the length constraint. Unlike previous studies that learn length embeddings, the proposed method can generate a text of any length even if the target length is unseen in training data. The experimental results show that the proposed method is able not only to control generation length but also improve ROUGE scores.

In our everyday chit-chat, there is a conversation initiator, who proactively casts an initial utterance to start chatting. However, most existing conversation systems cannot play this role. Previous studies on conversation systems assume that the user always initiates conversation, and have placed emphasis on how to respond to the given user’s utterance. As a result, existing conversation systems become passive. Namely they continue waiting until being spoken to by the users. In this paper, we consider the system as a conversation initiator and propose a novel task of generating the initial utterance in open-domain non-task-oriented conversation. Here, in order not to make users bored, it is necessary to generate diverse utterances to initiate conversation without relying on boilerplate utterances like greetings. To this end, we propose to generate initial utterance by summarizing and chatting about news articles, which provide fresh and various contents everyday. To address the lack of the training data for this task, we constructed a novel large-scale dataset through crowd-sourcing. We also analyzed the dataset in detail to examine how humans initiate conversations (the dataset will be released to facilitate future research activities). We present several approaches to conversation initiation including information retrieval based and generation based models. Experimental results showed that the proposed models trained on our dataset performed reasonably well and outperformed baselines that utilize automatically collected training data in both automatic and manual evaluation.

Cross-referencing, which links passages of text to other related passages, can be a valuable study aid for facilitating comprehension of a text. However, cross-referencing requires first, a comprehensive thematic knowledge of the entire corpus, and second, a focused search through the corpus specifically to find such useful connections. Due to this, cross-reference resources are prohibitively expensive and exist only for the most well-studied texts (e.g. religious texts). We develop a topic-based system for automatically producing candidate cross-references which can be easily verified by human annotators. Our system utilizes fine-grained topic modeling with thousands of highly nuanced and specific topics to identify verse pairs which are topically related. We demonstrate that our system can be cost effective compared to having annotators acquire the expertise necessary to produce cross-reference resources unaided.

In this work, we define the task of teaser generation and provide an evaluation benchmark and baseline systems for the process of generating teasers. A teaser is a short reading suggestion for an article that is illustrative and includes curiosity-arousing elements to entice potential readers to read particular news items. Teasers are one of the main vehicles for transmitting news to social media users. We compile a novel dataset of teasers by systematically accumulating tweets and selecting those that conform to the teaser definition. We have compared a number of neural abstractive architectures on the task of teaser generation and the overall best performing system is See et al. seq2seq with pointer network.

We present a new neural model for text summarization that first extracts sentences from a document and then compresses them. The pro-posed model offers a balance that sidesteps thedifficulties in abstractive methods while gener-ating more concise summaries than extractivemethods. In addition, our model dynamically determines the length of the output summary based on the gold summaries it observes during training and does not require length constraints typical to extractive summarization. The model achieves state-of-the-art results on the CNN/DailyMail and Newsroom datasets, improving over current extractive and abstractive methods. Human evaluations demonstratethat our model generates concise and informa-tive summaries. We also make available a new dataset of oracle compressive summaries derived automatically from the CNN/DailyMailreference summaries.

Neural abstractive summarizers generate summary texts using a language model conditioned on the input source text, and have recently achieved high ROUGE scores on benchmark summarization datasets. We investigate how they achieve this performance with respect to human-written gold-standard abstracts, and whether the systems are able to understand deeper syntactic and semantic structures. We generate a set of contrastive summaries which are perturbed, deficient versions of human-written summaries, and test whether existing neural summarizers score them more highly than the human-written summaries. We analyze their performance on different datasets and find that these systems fail to understand the source text, in a majority of the cases.

Recent work in the field of automatic summarization and headline generation focuses on maximizing ROUGE scores for various news datasets. We present an alternative, extrinsic, evaluation metric for this task, Answering Performance for Evaluation of Summaries. APES utilizes recent progress in the field of reading-comprehension to quantify the ability of a summary to answer a set of manually created questions regarding central entities in the source article. We first analyze the strength of this metric by comparing it to known manual evaluation metrics. We then present an end-to-end neural abstractive model that maximizes APES, while increasing ROUGE scores to competitive results.

Source Code Summarization is the task of writing short, natural language descriptions of source code. The main use for these descriptions is in software documentation e.g. the one-sentence Java method descriptions in JavaDocs. Code summarization is rapidly becoming a popular research problem, but progress is restrained due to a lack of suitable datasets. In addition, a lack of community standards for creating datasets leads to confusing and unreproducible research results – we observe swings in performance of more than 33% due only to changes in dataset design. In this paper, we make recommendations for these standards from experimental results. We release a dataset based on prior work of over 2.1m pairs of Java methods and one sentence method descriptions from over 28k Java projects. We describe the dataset and point out key differences from natural language data, to guide and support future researchers.

The existence of universal models to describe the syntax of languages has been debated for decades. The availability of resources such as the Universal Dependencies treebanks and the World Atlas of Language Structures make it possible to study the plausibility of universal grammar from the perspective of dependency parsing. Our work investigates the use of high-level language descriptions in the form of typological features for multilingual dependency parsing. Our experiments on multilingual parsing for 40 languages show that typological information can indeed guide parsers to share information between similar languages beyond simple language identification.

We introduce Rosita, a method to produce multilingual contextual word representations by training a single language model on text from multiple languages. Our method combines the advantages of contextual word representations with those of multilingual representation learning. We produce language models from dissimilar language pairs (English/Arabic and English/Chinese) and use them in dependency parsing, semantic role labeling, and named entity recognition, with comparisons to monolingual and non-contextual variants. Our results provide further evidence for the benefits of polyglot learning, in which representations are shared across multiple languages.

Cross-lingual word vectors are typically obtained by fitting an orthogonal matrix that maps the entries of a bilingual dictionary from a source to a target vector space. Word vectors, however, are most commonly used for sentence or document-level representations that are calculated as the weighted average of word embeddings. In this paper, we propose an alternative to word-level mapping that better reflects sentence-level cross-lingual similarity. We incorporate context in the transformation matrix by directly mapping the averaged embeddings of aligned sentences in a parallel corpus. We also implement cross-lingual mapping of deep contextualized word embeddings using parallel sentences with word alignments. In our experiments, both approaches resulted in cross-lingual sentence embeddings that outperformed context-independent word mapping in sentence translation retrieval. Furthermore, the sentence-level transformation could be used for word-level mapping without loss in word translation quality.

It is well-known that distributional semantic approaches have difficulty in distinguishing between synonyms and antonyms (Grefenstette, 1992; Padó and Lapata, 2003). Recent work has shown that supervision available in English for this task (e.g., lexical resources) can be transferred to other languages via cross-lingual word embeddings. However, this kind of transfer misses monolingual distributional information available in a target language, such as contrast relations that are indicative of antonymy (e.g. hot ... while ... cold). In this work, we improve the transfer by exploiting monolingual information, expressed in the form of co-occurrences with discourse markers that convey contrast. Our approach makes use of less than a dozen markers, which can easily be obtained for many languages. Compared to a baseline using only cross-lingual embeddings, we show absolute improvements of 4–10% F1-score in Vietnamese and Hindi.

There is no consensus on the state-of-the-art approach to historical text normalization. Many techniques have been proposed, including rule-based methods, distance metrics, character-based statistical machine translation, and neural encoder–decoder models, but studies have used different datasets, different evaluation methods, and have come to different conclusions. This paper presents the largest study of historical text normalization done so far. We critically survey the existing literature and report experiments on eight languages, comparing systems spanning all categories of proposed normalization techniques, analysing the effect of training data quantity, and using different evaluation methods. The datasets and scripts are made publicly available.

Multilingual Neural Machine Translation enables training a single model that supports translation from multiple source languages into multiple target languages. We perform extensive experiments in training massively multilingual NMT models, involving up to 103 distinct languages and 204 translation directions simultaneously. We explore different setups for training such models and analyze the trade-offs between translation quality and various modeling decisions. We report results on the publicly available TED talks multilingual corpus where we show that massively multilingual many-to-many models are effective in low resource settings, outperforming the previous state-of-the-art while supporting up to 59 languages in 116 translation directions in a single model. Our experiments on a large-scale dataset with 103 languages, 204 trained directions and up to one million examples per direction also show promising results, surpassing strong bilingual baselines and encouraging future work on massively multilingual NMT.

Transfer learning approaches for Neural Machine Translation (NMT) train a NMT model on an assisting language-target language pair (parent model) which is later fine-tuned for the source language-target language pair of interest (child model), with the target language being the same. In many cases, the assisting language has a different word order from the source language. We show that divergent word order adversely limits the benefits from transfer learning when little to no parallel corpus between the source and target language is available. To bridge this divergence, we propose to pre-order the assisting language sentences to match the word order of the source language and train the parent model. Our experiments on many language pairs show that bridging the word order gap leads to significant improvement in the translation quality in extremely low-resource scenarios.

Adversarial training has shown impressive success in learning bilingual dictionary without any parallel data by mapping monolingual embeddings to a shared space. However, recent work has shown superior performance for non-adversarial methods in more challenging language pairs. In this work, we revisit adversarial autoencoder for unsupervised word translation and propose two novel extensions to it that yield more stable training and improved results. Our method includes regularization terms to enforce cycle consistency and input reconstruction, and puts the target encoders as an adversary against the corresponding discriminator. Extensive experimentations with European, non-European and low-resource languages show that our method is more robust and achieves better performance than recently proposed adversarial and non-adversarial approaches.

We describe a cross-lingual transfer method for dependency parsing that takes into account the problem of word order differences between source and target languages. Our model only relies on the Bible, a considerably smaller parallel data than the commonly used parallel data in transfer methods. We use the concatenation of projected trees from the Bible corpus, and the gold-standard treebanks in multiple source languages along with cross-lingual word representations. We demonstrate that reordering the source treebanks before training on them for a target language improves the accuracy of languages outside the European language family. Our experiments on 68 treebanks (38 languages) in the Universal Dependencies corpus achieve a high accuracy for all languages. Among them, our experiments on 16 treebanks of 12 non-European languages achieve an average UAS absolute improvement of 3.3% over a state-of-the-art method.