NAACL 2019 Papers

We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models (Peters et al., 2018a; Radford et al., 2018), BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5 (7.7 point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement).

Current work on multimodal machine translation (MMT) has suggested that the visual modality is either unnecessary or only marginally beneficial. We posit that this is a consequence of the very simple, short and repetitive sentences used in the only available dataset for the task (Multi30K), rendering the source text sufficient as context. In the general case, however, we believe that it is possible to combine visual and textual information in order to ground translations. In this paper we probe the contribution of the visual modality to state-of-the-art MMT models by conducting a systematic analysis where we partially deprive the models from source-side textual context. Our results show that under limited textual context, models are capable of leveraging the visual input to generate better translations. This contradicts the current belief that MMT models disregard the visual modality because of either the quality of the image features or the way they are integrated into the model.

When answering a question, people often draw upon their rich world knowledge in addition to the particular context. Recent work has focused primarily on answering questions given some relevant document or context, and required very little general background. To investigate question answering with prior knowledge, we present CommonsenseQA: a challenging new dataset for commonsense question answering. To capture common sense beyond associations, we extract from ConceptNet (Speer et al., 2017) multiple target concepts that have the same semantic relation to a single source concept. Crowd-workers are asked to author multiple-choice questions that mention the source concept and discriminate in turn between each of the target concepts. This encourages workers to create questions with complex semantics that often require prior knowledge. We create 12,247 questions through this procedure and demonstrate the difficulty of our task with a large number of strong baselines. Our best baseline is based on BERT-large (Devlin et al., 2018) and obtains 56% accuracy, well below human performance, which is 89%.

This paper seeks to model human language by the mathematical framework of quantum physics. With the well-designed mathematical formulations in quantum physics, this framework unifies different linguistic units in a single complex-valued vector space, e.g. words as particles in quantum states and sentences as mixed systems. A complex-valued network is built to implement this framework for semantic matching. With well-constrained complex-valued components, the network admits interpretations to explicit physical meanings. The proposed complex-valued network for matching (CNM) achieves comparable performances to strong CNN and RNN baselines on two benchmarking question answering (QA) datasets.

Recent work has improved our ability to detect linguistic knowledge in word representations. However, current methods for detecting syntactic knowledge do not test whether syntax trees are represented in their entirety. In this work, we propose a structural probe, which evaluates whether syntax trees are embedded in a linear transformation of a neural network’s word representation space. The probe identifies a linear transformation under which squared L2 distance encodes the distance between words in the parse tree, and one in which squared L2 norm encodes depth in the parse tree. Using our probe, we show that such transformations exist for both ELMo and BERT but not in baselines, providing evidence that entire syntax trees are embedded implicitly in deep models’ vector geometry.

Lemmatization aims to reduce the sparse data problem by relating the inflected forms of a word to its dictionary form. Using context can help, both for unseen and ambiguous words. Yet most context-sensitive approaches require full lemma-annotated sentences for training, which may be scarce or unavailable in low-resource languages. In addition (as shown here), in a low-resource setting, a lemmatizer can learn more from n labeled examples of distinct words (types) than from n (contiguous) labeled tokens, since the latter contain far fewer distinct types. To combine the efficiency of type-based learning with the benefits of context, we propose a way to train a context-sensitive lemmatizer with little or no labeled corpus data, using inflection tables from the UniMorph project and raw text examples from Wikipedia that provide sentence contexts for the unambiguous UniMorph examples. Despite these being unambiguous examples, the model successfully generalizes from them, leading to improved results (both overall, and especially on unseen words) in comparison to a baseline that does not use context.

In recent years neural language models (LMs) have set the state-of-the-art performance for several benchmarking datasets. While the reasons for their success and their computational demand are well-documented, a comparison between neural models and more recent developments in n-gram models is neglected. In this paper, we examine the recent progress in n-gram literature, running experiments on 50 languages covering all morphological language families. Experimental results illustrate that a simple extension of Modified Kneser-Ney outperforms an lstm language model on 42 languages while a word-level Bayesian n-gram LM (Shareghi et al., 2017) outperforms the character-aware neural model (Kim et al., 2016) on average across all languages, and its extension which explicitly injects linguistic knowledge (Gerz et al., 2018) on 8 languages. Further experiments on larger Europarl datasets for 3 languages indicate that neural architectures are able to outperform computationally much cheaper n-gram models: n-gram training is up to 15,000x quicker. Our experiments illustrate that standalone n-gram models lend themselves as natural choices for resource-lean or morphologically rich languages, while the recent progress has significantly improved their accuracy.

Fine-tuning neural networks is widely used to transfer valuable knowledge from high-resource to low-resource domains. In a standard fine-tuning scheme, source and target problems are trained using the same architecture. Although capable of adapting to new domains, pre-trained units struggle with learning uncommon target-specific patterns. In this paper, we propose to augment the target-network with normalised, weighted and randomly initialised units that beget a better adaptation while maintaining the valuable source knowledge. Our experiments on POS tagging of social media texts (Tweets domain) demonstrate that our method achieves state-of-the-art performances on 3 commonly used datasets.

While idiosyncrasies of the Chinese classifier system have been a richly studied topic among linguists (Adams and Conklin, 1973; Erbaugh, 1986; Lakoff, 1986), not much work has been done to quantify them with statistical methods. In this paper, we introduce an information-theoretic approach to measuring idiosyncrasy; we examine how much the uncertainty in Mandarin Chinese classifiers can be reduced by knowing semantic information about the nouns that the classifiers modify. Using the empirical distribution of classifiers from the parsed Chinese Gigaword corpus (Graff et al., 2005), we compute the mutual information (in bits) between the distribution over classifiers and distributions over other linguistic quantities. We investigate whether semantic classes of nouns and adjectives differ in how much they reduce uncertainty in classifier choice, and find that it is not fully idiosyncratic; while there are no obvious trends for the majority of semantic classes, shape nouns reduce uncertainty in classifier choice the most.

This paper presents three hybrid models that directly combine latent Dirichlet allocation and word embedding for distinguishing between speakers with and without Alzheimer’s disease from transcripts of picture descriptions. Two of our models get F-scores over the current state-of-the-art using automatic methods on the DementiaBank dataset.

A number of psycholinguistic studies have factorially manipulated words’ contextual predictabilities and corpus frequencies and shown separable effects of each on measures of human sentence processing, a pattern which has been used to support distinct mechanisms underlying prediction on the one hand and lexical retrieval on the other. This paper examines the generalizability of this finding to more realistic conditions of sentence processing by studying effects of frequency and predictability in three large-scale naturalistic reading corpora. Results show significant effects of word frequency and predictability in isolation but no effect of frequency over and above predictability, and thus do not provide evidence of distinct mechanisms. The non-replication of separable effects in a naturalistic setting raises doubts about the existence of such a distinction in everyday sentence comprehension. Instead, these results are consistent with previous claims that apparent effects of frequency are underlyingly effects of predictability.

Understanding procedural language requires reasoning about both hierarchical and temporal relations between events. For example, “boiling pasta” is a sub-event of “making a pasta dish”, typically happens before “draining pasta,” and requires the use of omitted tools (e.g. a strainer, sink...). While people are able to choose when and how to use abstract versus concrete instructions, the NLP community lacks corpora and tasks for evaluating if our models can do the same. In this paper, we introduce KidsCook, a parallel script corpus, as well as a cloze task which matches video captions with missing procedural details. Experimental results show that state-of-the-art models struggle at this task, which requires inducing functional commonsense knowledge not explicitly stated in text.

The variational autoencoder (VAE) imposes a probabilistic distribution (typically Gaussian) on the latent space and penalizes the Kullback-Leibler (KL) divergence between the posterior and prior. In NLP, VAEs are extremely difficult to train due to the problem of KL collapsing to zero. One has to implement various heuristics such as KL weight annealing and word dropout in a carefully engineered manner to successfully train a VAE for text. In this paper, we propose to use the Wasserstein autoencoder (WAE) for probabilistic sentence generation, where the encoder could be either stochastic or deterministic. We show theoretically and empirically that, in the original WAE, the stochastically encoded Gaussian distribution tends to become a Dirac-delta function, and we propose a variant of WAE that encourages the stochasticity of the encoder. Experimental results show that the latent space learned by WAE exhibits properties of continuity and smoothness as in VAEs, while simultaneously achieving much higher BLEU scores for sentence reconstruction.

We improve the informativeness of models for conditional text generation using techniques from computational pragmatics. These techniques formulate language production as a game between speakers and listeners, in which a speaker should generate output text that a listener can use to correctly identify the original input that the text describes. While such approaches are widely used in cognitive science and grounded language learning, they have received less attention for more standard language generation tasks. We consider two pragmatic modeling methods for text generation: one where pragmatics is imposed by information preservation, and another where pragmatics is imposed by explicit modeling of distractors. We find that these methods improve the performance of strong existing systems for abstractive summarization and generation from structured meaning representations.

Pre-trained language model representations have been successful in a wide range of language understanding tasks. In this paper, we examine different strategies to integrate pre-trained representations into sequence to sequence models and apply it to neural machine translation and abstractive summarization. We find that pre-trained representations are most effective when added to the encoder network which slows inference by only 14%. Our experiments in machine translation show gains of up to 5.3 BLEU in a simulated resource-poor setup. While returns diminish with more labeled data, we still observe improvements when millions of sentence-pairs are available. Finally, on abstractive summarization we achieve a new state of the art on the full text version of CNN/DailyMail.

Neural network models for many NLP tasks have grown increasingly complex in recent years, making training and deployment more difficult. A number of recent papers have questioned the necessity of such architectures and found that well-executed, simpler models are quite effective. We show that this is also the case for document classification: in a large-scale reproducibility study of several recent neural models, we find that a simple BiLSTM architecture with appropriate regularization yields accuracy and F1 that are either competitive or exceed the state of the art on four standard benchmark datasets. Surprisingly, our simple model is able to achieve these results without attention mechanisms. While these regularization techniques, borrowed from language modeling, are not novel, to our knowledge we are the first to apply them in this context. Our work provides an open-source platform and the foundation for future work in document classification.

Self-attention networks (SANs) have drawn increasing interest due to their high parallelization in computation and flexibility in modeling dependencies. SANs can be further enhanced with multi-head attention by allowing the model to attend to information from different representation subspaces. In this work, we propose novel convolutional self-attention networks, which offer SANs the abilities to 1) strengthen dependencies among neighboring elements, and 2) model the interaction between features extracted by multiple attention heads. Experimental results of machine translation on different language pairs and model settings show that our approach outperforms both the strong Transformer baseline and other existing models on enhancing the locality of SANs. Comparing with prior studies, the proposed model is parameter free in terms of introducing no more parameters.

Grammatical error correction (GEC) is one of the areas in natural language processing in which purely neural models have not yet superseded more traditional symbolic models. Hybrid systems combining phrase-based statistical machine translation (SMT) and neural sequence models are currently among the most effective approaches to GEC. However, both SMT and neural sequence-to-sequence models require large amounts of annotated data. Language model based GEC (LM-GEC) is a promising alternative which does not rely on annotated training data. We show how to improve LM-GEC by applying modelling techniques based on finite state transducers. We report further gains by rescoring with neural language models. We show that our methods developed for LM-GEC can also be used with SMT systems if annotated training data is available. Our best system outperforms the best published result on the CoNLL-2014 test set, and achieves far better relative improvements over the SMT baselines than previous hybrid systems.

Learning multi-hop reasoning has been a key challenge for reading comprehension models, leading to the design of datasets that explicitly focus on it. Ideally, a model should not be able to perform well on a multi-hop question answering task without doing multi-hop reasoning. In this paper, we investigate two recently proposed datasets, WikiHop and HotpotQA. First, we explore sentence-factored models for these tasks; by design, these models cannot do multi-hop reasoning, but they are still able to solve a large number of examples in both datasets. Furthermore, we find spurious correlations in the unmasked version of WikiHop, which make it easy to achieve high performance considering only the questions and answers. Finally, we investigate one key difference between these datasets, namely span-based vs. multiple-choice formulations of the QA task. Multiple-choice versions of both datasets can be easily gamed, and two models we examine only marginally exceed a baseline in this setting. Overall, while these datasets are useful testbeds, high-performing models may not be learning as much multi-hop reasoning as previously thought.