NAACL 2019 Papers

We focus on improving name tagging for low-resource languages using annotations from related languages. Previous studies either directly project annotations from a source language to a target language using cross-lingual representations or use a shared encoder in a multitask network to transfer knowledge. These approaches inevitably introduce noise to the target language annotation due to mismatched source-target sentence structures. To effectively transfer the resources, we develop a new neural architecture that leverages multi-level adversarial transfer: (1) word-level adversarial training, which projects source language words into the same semantic space as those of the target language without using any parallel corpora or bilingual gazetteers, and (2) sentence-level adversarial training, which yields language-agnostic sequential features. Our neural architecture outperforms previous approaches on CoNLL data sets. Moreover, on 10 low-resource languages, our approach achieves up to 16% absolute F-score gain over all high-performing baselines on cross-lingual transfer without using any target-language resources.

Recent advances in Question Answering have lead to the development of very complex models which compute rich representations for query and documents by capturing all pairwise interactions between query and document words. This makes these models expensive in space and time, and in practice one has to restrict the length of the documents that can be fed to these models. Such models have also been recently employed for the task of predicting dialog responses from available background documents (e.g., Holl-E dataset). However, here the documents are longer, thereby rendering these complex models infeasible except in select restricted settings. In order to overcome this, we use standard simple models which do not capture all pairwise interactions, but learn to emulate certain characteristics of a complex teacher network. Specifically, we first investigate the conicity of representations learned by a complex model and observe that it is significantly lower than that of simpler models. Based on this insight, we modify the simple architecture to mimic this characteristic. We go further by using knowledge distillation approaches, where the simple model acts as a student and learns to match the output from the complex teacher network. We experiment with the Holl-E dialog data set and show that by mimicking characteristics and matching outputs from a teacher, even a simple network can give improved performance.

Evaluating open-domain dialogue systems is difficult due to the diversity of possible correct answers. Automatic metrics such as BLEU correlate weakly with human annotations, resulting in a significant bias across different models and datasets. Some researchers resort to human judgment experimentation for assessing response quality, which is expensive, time consuming, and not scalable. Moreover, judges tend to evaluate a small number of dialogues, meaning that minor differences in evaluation configuration may lead to dissimilar results. In this paper, we present interpretable metrics for evaluating topic coherence by making use of distributed sentence representations. Furthermore, we introduce calculable approximations of human judgment based on conversational coherence by adopting state-of-the-art entailment techniques. Results show that our metrics can be used as a surrogate for human judgment, making it easy to evaluate dialogue systems on large-scale datasets and allowing an unbiased estimate for the quality of the responses.

One of the first steps in the utterance interpretation pipeline of many task-oriented conversational AI systems is to identify user intents and the corresponding slots. Since data collection for machine learning models for this task is time-consuming, it is desirable to make use of existing data in a high-resource language to train models in low-resource languages. However, development of such models has largely been hindered by the lack of multilingual training data. In this paper, we present a new data set of 57k annotated utterances in English (43k), Spanish (8.6k) and Thai (5k) across the domains weather, alarm, and reminder. We use this data set to evaluate three different cross-lingual transfer methods: (1) translating the training data, (2) using cross-lingual pre-trained embeddings, and (3) a novel method of using a multilingual machine translation encoder as contextual word representations. We find that given several hundred training examples in the the target language, the latter two methods outperform translating the training data. Further, in very low-resource settings, multilingual contextual word representations give better results than using cross-lingual static embeddings. We also compare the cross-lingual methods to using monolingual resources in the form of contextual ELMo representations and find that given just small amounts of target language data, this method outperforms all cross-lingual methods, which highlights the need for more sophisticated cross-lingual methods.

Domain classification is the task to map spoken language utterances to one of the natural language understanding domains in intelligent personal digital assistants (IPDAs). This is observed in mainstream IPDAs in industry and third-party domains are developed to enhance the capability of the IPDAs. As more and more new domains are developed very frequently, how to continuously accommodate the new domains still remains challenging. Moreover, if one wants to use personalized information dynamically for better domain classification, it is infeasible to directly adopt existing continual learning approaches. In this paper, we propose CoNDA, a neural-based approach for continuous domain adaption with normalization and regularization. Unlike existing methods that often conduct full model parameter update, CoNDA only updates the necessary parameters in the model for the new domains. Empirical evaluation shows that CoNDA achieves high accuracy on both the accommodated new domains and the existing known domains for which input samples come with personal information, and outperforms the baselines by a large margin.

Humans use language to refer to entities in the external world. Motivated by this, in recent years several models that incorporate a bias towards learning entity representations have been proposed. Such entity-centric models have shown empirical success, but we still know little about why. In this paper we analyze the behavior of two recently proposed entity-centric models in a referential task, Entity Linking in Multi-party Dialogue (SemEval 2018 Task 4). We show that these models outperform the state of the art on this task, and that they do better on lower frequency entities than a counterpart model that is not entity-centric, with the same model size. We argue that making models entity-centric naturally fosters good architectural decisions. However, we also show that these models do not really build entity representations and that they make poor use of linguistic context. These negative results underscore the need for model analysis, to test whether the motivations for particular architectures are borne out in how models behave when deployed.

We consider neural language generation under a novel problem setting: generating the words of a sentence according to the order of their first appearance in its lexicalized PCFG parse tree, in a depth-first, left-to-right manner. Unlike previous tree-based language generation methods, our approach is both (i) top-down and (ii) explicitly generating syntactic structure at the same time. In addition, our method combines neural model with symbolic approach: word choice at each step is constrained by its predicted syntactic function. We applied our model to the task of dialog response generation, and found it significantly improves over sequence-to-sequence baseline, in terms of diversity and relevance. We also investigated the effect of lexicalization on language generation, and found that lexicalization schemes that give priority to content words have certain advantages over those focusing on dependency relations.

Success of deep learning techniques have renewed the interest in development of dialogue systems. However, current systems struggle to have consistent long term conversations with the users and fail to build rapport. Topic spotting, the task of automatically inferring the topic of a conversation, has been shown to be helpful in making dialog system more engaging and efficient. We propose a hierarchical model with self attention for topic spotting. Experiments on the Switchboard corpus show the superior performance of our model over previously proposed techniques for topic spotting and deep models for text classification. Additionally, in contrast to offline processing of dialog, we also analyze the performance of our model in a more realistic setting i.e. in an online setting where the topic is identified in real time as the dialog progresses. Results show that our model is able to generalize even with limited information in the online setting.

Recent end-to-end task oriented dialog systems use memory architectures to incorporate external knowledge in their dialogs. Current work makes simplifying assumptions about the structure of the knowledge base, such as the use of triples to represent knowledge, and combines dialog utterances (context) as well as knowledge base (KB) results as part of the same memory. This causes an explosion in the memory size, and makes the reasoning over memory harder. In addition, such a memory design forces hierarchical properties of the data to be fit into a triple structure of memory. This requires the memory reader to infer relationships across otherwise connected attributes. In this paper we relax the strong assumptions made by existing architectures and separate memories used for modeling dialog context and KB results. Instead of using triples to store KB results, we introduce a novel multi-level memory architecture consisting of cells for each query and their corresponding results. The multi-level memory first addresses queries, followed by results and finally each key-value pair within a result. We conduct detailed experiments on three publicly available task oriented dialog data sets and we find that our method conclusively outperforms current state-of-the-art models. We report a 15-25% increase in both entity F1 and BLEU scores.

The majority of current systems for end-to-end dialog generation focus on response quality without an explicit control over the affective content of the responses. In this paper, we present an affect-driven dialog system, which generates emotional responses in a controlled manner using a continuous representation of emotions. The system achieves this by modeling emotions at a word and sequence level using: (1) a vector representation of the desired emotion, (2) an affect regularizer, which penalizes neutral words, and (3) an affect sampling method, which forces the neural network to generate diverse words that are emotionally relevant. During inference, we use a re-ranking procedure that aims to extract the most emotionally relevant responses using a human-in-the-loop optimization process. We study the performance of our system in terms of both quantitative (BLEU score and response diversity), and qualitative (emotional appropriateness) measures.

Recent work in Dialogue Act classification has treated the task as a sequence labeling problem using hierarchical deep neural networks. We build on this prior work by leveraging the effectiveness of a context-aware self-attention mechanism coupled with a hierarchical recurrent neural network. We conduct extensive evaluations on standard Dialogue Act classification datasets and show significant improvement over state-of-the-art results on the Switchboard Dialogue Act (SwDA) Corpus. We also investigate the impact of different utterance-level representation learning methods and show that our method is effective at capturing utterance-level semantic text representations while maintaining high accuracy.

To capture salient contextual information for spoken language understanding (SLU) of a dialogue, we propose time-aware models that automatically learn the latent time-decay function of the history without a manual time-decay function. We also propose a method to identify and label the current speaker to improve the SLU accuracy. In experiments on the benchmark dataset used in Dialog State Tracking Challenge 4, the proposed models achieved significantly higher F1 scores than the state-of-the-art contextual models. Finally, we analyze the effectiveness of the introduced models in detail. The analysis demonstrates that the proposed methods were effective to improve SLU accuracy individually.

How do we know if a particular medical treatment actually works? Ideally one would consult all available evidence from relevant clinical trials. Unfortunately, such results are primarily disseminated in natural language scientific articles, imposing substantial burden on those trying to make sense of them. In this paper, we present a new task and corpus for making this unstructured published scientific evidence actionable. The task entails inferring reported findings from a full-text article describing randomized controlled trials (RCT) with respect to a given intervention, comparator, and outcome of interest, e.g., inferring if a given article provides evidence supporting the use of aspirin to reduce risk of stroke, as compared to placebo. We present a new corpus for this task comprising 10,000+ prompts coupled with full-text articles describing RCTs. Results using a suite of baseline models — ranging from heuristic (rule-based) approaches to attentive neural architectures — demonstrate the difficulty of the task, which we believe largely owes to the lengthy, technical input texts. To facilitate further work on this important, challenging problem we make the corpus, documentation, a website and leaderboard, and all source code for baselines and evaluation publicly available.

Most information extraction methods focus on binary relations expressed within single sentences. In high-value domains, however, n-ary relations are of great demand (e.g., drug-gene-mutation interactions in precision oncology). Such relations often involve entity mentions that are far apart in the document, yet existing work on cross-sentence relation extraction is generally confined to small text spans (e.g., three consecutive sentences), which severely limits recall. In this paper, we propose a novel multiscale neural architecture for document-level n-ary relation extraction. Our system combines representations learned over various text spans throughout the document and across the subrelation hierarchy. Widening the system’s purview to the entire document maximizes potential recall. Moreover, by integrating weak signals across the document, multiscale modeling increases precision, even in the presence of noisy labels from distant supervision. Experiments on biomedical machine reading show that our approach substantially outperforms previous n-ary relation extraction methods.

Incorporating domain knowledge is vital in building successful natural language processing (NLP) applications. Many times, cross-domain application of a tool results in poor performance as the tool does not account for domain-specific attributes. The clinical domain is challenging in this aspect due to specialized medical terms and nomenclature, shorthand notation, fragmented text, and a variety of writing styles used by different medical units. Temporal resolution is an NLP task that, in general, is domain-agnostic because temporal information is represented using a limited lexicon. However, domain-specific aspects of temporal resolution are present in clinical texts. Here we explore parsing issues that arose when running our system, a tool built on Newswire text, on clinical notes in the THYME corpus. Many parsing issues were straightforward to correct; however, a few code changes resulted in a cascading series of parsing errors that had to be resolved before an improvement in performance was observed, revealing the complexity temporal resolution and rule-based parsing. Our system now outperforms current state-of-the-art systems on the THYME corpus with little change in its performance on Newswire texts.

To make machines better understand sentiments, research needs to move from polarity identification to understanding the reasons that underlie the expression of sentiment. Categorizing the goals or needs of humans is one way to explain the expression of sentiment in text. Humans are good at understanding situations described in natural language and can easily connect them to the character’s psychological needs using commonsense knowledge. We present a novel method to extract, rank, filter and select multi-hop relation paths from a commonsense knowledge resource to interpret the expression of sentiment in terms of their underlying human needs. We efficiently integrate the acquired knowledge paths in a neural model that interfaces context representations with knowledge using a gated attention mechanism. We assess the model’s performance on a recently published dataset for categorizing human needs. Selectively integrating knowledge paths boosts performance and establishes a new state-of-the-art. Our model offers interpretability through the learned attention map over commonsense knowledge paths. Human evaluation highlights the relevance of the encoded knowledge.

There is growing evidence that changes in speech and language may be early markers of dementia, but much of the previous NLP work in this area has been limited by the size of the available datasets. Here, we compare several methods of domain adaptation to augment a small French dataset of picture descriptions (n = 57) with a much larger English dataset (n = 550), for the task of automatically distinguishing participants with dementia from controls. The first challenge is to identify a set of features that transfer across languages; in addition to previously used features based on information units, we introduce a new set of features to model the order in which information units are produced by dementia patients and controls. These concept-based language model features improve classification performance in both English and French separately, and the best result (AUC = 0.89) is achieved using the multilingual training set with a combination of information and language model features.

AMR-to-text generation is a problem recently introduced to the NLP community, in which the goal is to generate sentences from Abstract Meaning Representation (AMR) graphs. Sequence-to-sequence models can be used to this end by converting the AMR graphs to strings. Approaching the problem while working directly with graphs requires the use of graph-to-sequence models that encode the AMR graph into a vector representation. Such encoding has been shown to be beneficial in the past, and unlike sequential encoding, it allows us to explicitly capture reentrant structures in the AMR graphs. We investigate the extent to which reentrancies (nodes with multiple parents) have an impact on AMR-to-text generation by comparing graph encoders to tree encoders, where reentrancies are not preserved. We show that improvements in the treatment of reentrancies and long-range dependencies contribute to higher overall scores for graph encoders. Our best model achieves 24.40 BLEU on LDC2015E86, outperforming the state of the art by 1.1 points and 24.54 BLEU on LDC2017T10, outperforming the state of the art by 1.24 points.

We introduce Recursive Routing Networks (RRNs), which are modular, adaptable models that learn effectively in diverse environments. RRNs consist of a set of functions, typically organized into a grid, and a meta-learner decision-making component called the router. The model jointly optimizes the parameters of the functions and the meta-learner’s policy for routing inputs through those functions. RRNs can be incorporated into existing architectures in a number of ways; we explore adding them to word representation layers, recurrent network hidden layers, and classifier layers. Our evaluation task is natural language inference (NLI). Using the MultiNLI corpus, we show that an RRN’s routing decisions reflect the high-level genre structure of that corpus. To show that RRNs can learn to specialize to more fine-grained semantic distinctions, we introduce a new corpus of NLI examples involving implicative predicates, and show that the model components become fine-tuned to the inferential signatures that are characteristic of these predicates.