NAACL 2019 论文列表

Recently, the Transformer model that is based solely on attention mechanisms, has advanced the state-of-the-art on various machine translation tasks. However, recent studies reveal that the lack of recurrence modeling hinders its further improvement of translation capacity. In response to this problem, we propose to directly model recurrence for Transformer with an additional recurrence encoder. In addition to the standard recurrent neural network, we introduce a novel attentive recurrent network to leverage the strengths of both attention models and recurrent networks. Experimental results on the widely-used WMT14 English⇒German and WMT17 Chinese⇒English translation tasks demonstrate the effectiveness of the proposed approach. Our studies also reveal that the proposed model benefits from a short-cut that bridges the source and target sequences with a single recurrent layer, which outperforms its deep counterpart.

Generalization and reliability of multilingual translation often highly depend on the amount of available parallel data for each language pair of interest. In this paper, we focus on zero-shot generalization—a challenging setup that tests models on translation directions they have not been optimized for at training time. To solve the problem, we (i) reformulate multilingual translation as probabilistic inference, (ii) define the notion of zero-shot consistency and show why standard training often results in models unsuitable for zero-shot tasks, and (iii) introduce a consistent agreement-based training method that encourages the model to produce equivalent translations of parallel sentences in auxiliary languages. We test our multilingual NMT models on multiple public zero-shot translation benchmarks (IWSLT17, UN corpus, Europarl) and show that agreement-based learning often results in 2-3 BLEU zero-shot improvement over strong baselines without any loss in performance on supervised translation directions.

The overreliance on large parallel corpora significantly limits the applicability of machine translation systems to the majority of language pairs. Back-translation has been dominantly used in previous approaches for unsupervised neural machine translation, where pseudo sentence pairs are generated to train the models with a reconstruction loss. However, the pseudo sentences are usually of low quality as translation errors accumulate during training. To avoid this fundamental issue, we propose an alternative but more effective approach, extract-edit, to extract and then edit real sentences from the target monolingual corpora. Furthermore, we introduce a comparative translation loss to evaluate the translated target sentences and thus train the unsupervised translation systems. Experiments show that the proposed approach consistently outperforms the previous state-of-the-art unsupervised machine translation systems across two benchmarks (English-French and English-German) and two low-resource language pairs (English-Romanian and English-Russian) by more than 2 (up to 3.63) BLEU points.

Current state-of-the-art NMT systems use large neural networks that are not only slow to train, but also often require many heuristics and optimization tricks, such as specialized learning rate schedules and large batch sizes. This is undesirable as it requires extensive hyperparameter tuning. In this paper, we propose a curriculum learning framework for NMT that reduces training time, reduces the need for specialized heuristics or large batch sizes, and results in overall better performance. Our framework consists of a principled way of deciding which training samples are shown to the model at different times during training, based on the estimated difficulty of a sample and the current competence of the model. Filtering training samples in this manner prevents the model from getting stuck in bad local optima, making it converge faster and reach a better solution than the common approach of uniformly sampling training examples. Furthermore, the proposed method can be easily applied to existing NMT models by simply modifying their input data pipelines. We show that our framework can help improve the training time and the performance of both recurrent neural network models and Transformers, achieving up to a 70% decrease in training time, while at the same time obtaining accuracy improvements of up to 2.2 BLEU.

Syntax has been demonstrated highly effective in neural machine translation (NMT). Previous NMT models integrate syntax by representing 1-best tree outputs from a well-trained parsing system, e.g., the representative Tree-RNN and Tree-Linearization methods, which may suffer from error propagation. In this work, we propose a novel method to integrate source-side syntax implicitly for NMT. The basic idea is to use the intermediate hidden representations of a well-trained end-to-end dependency parser, which are referred to as syntax-aware word representations (SAWRs). Then, we simply concatenate such SAWRs with ordinary word embeddings to enhance basic NMT models. The method can be straightforwardly integrated into the widely-used sequence-to-sequence (Seq2Seq) NMT models. We start with a representative RNN-based Seq2Seq baseline system, and test the effectiveness of our proposed method on two benchmark datasets of the Chinese-English and English-Vietnamese translation tasks, respectively. Experimental results show that the proposed approach is able to bring significant BLEU score improvements on the two datasets compared with the baseline, 1.74 points for Chinese-English translation and 0.80 point for English-Vietnamese translation, respectively. In addition, the approach also outperforms the explicit Tree-RNN and Tree-Linearization methods.

Traditional language models are unable to efficiently model entity names observed in text. All but the most popular named entities appear infrequently in text providing insufficient context. Recent efforts have recognized that context can be generalized between entity names that share the same type (e.g., person or location) and have equipped language models with an access to external knowledge base (KB). Our Knowledge-Augmented Language Model (KALM) continues this line of work by augmenting a traditional model with a KB. Unlike previous methods, however, we train with an end-to-end predictive objective optimizing the perplexity of text. We do not require any additional information such as named entity tags. In addition to improving language modeling performance, KALM learns to recognize named entities in an entirely unsupervised way by using entity type information latent in the model. On a Named Entity Recognition (NER) task, KALM achieves performance comparable with state-of-the-art supervised models. Our work demonstrates that named entities (and possibly other types of world knowledge) can be modeled successfully using predictive learning and training on large corpora of text without any additional information.

We introduce the deep inside-outside recursive autoencoder (DIORA), a fully-unsupervised method for discovering syntax that simultaneously learns representations for constituents within the induced tree. Our approach predicts each word in an input sentence conditioned on the rest of the sentence. During training we use dynamic programming to consider all possible binary trees over the sentence, and for inference we use the CKY algorithm to extract the highest scoring parse. DIORA outperforms previously reported results for unsupervised binary constituency parsing on the benchmark WSJ dataset.

There has been considerable attention devoted to models that learn to jointly infer an expression’s syntactic structure and its semantics. Yet, Nangia and Bowman (2018) has recently shown that the current best systems fail to learn the correct parsing strategy on mathematical expressions generated from a simple context-free grammar. In this work, we present a recursive model inspired by Choi et al. (2018) that reaches near perfect accuracy on this task. Our model is composed of two separated modules for syntax and semantics. They are cooperatively trained with standard continuous and discrete optimisation schemes. Our model does not require any linguistic structure for supervision, and its recursive nature allows for out-of-domain generalisation. Additionally, our approach performs competitively on several natural language tasks, such as Natural Language Inference and Sentiment Analysis.

Recurrent neural network grammars (RNNG) are generative models of language which jointly model syntax and surface structure by incrementally generating a syntax tree and sentence in a top-down, left-to-right order. Supervised RNNGs achieve strong language modeling and parsing performance, but require an annotated corpus of parse trees. In this work, we experiment with unsupervised learning of RNNGs. Since directly marginalizing over the space of latent trees is intractable, we instead apply amortized variational inference. To maximize the evidence lower bound, we develop an inference network parameterized as a neural CRF constituency parser. On language modeling, unsupervised RNNGs perform as well their supervised counterparts on benchmarks in English and Chinese. On constituency grammar induction, they are competitive with recent neural language models that induce tree structures from words through attention mechanisms.

We address part-of-speech (POS) induction by maximizing the mutual information between the induced label and its context. We focus on two training objectives that are amenable to stochastic gradient descent (SGD): a novel generalization of the classical Brown clustering objective and a recently proposed variational lower bound. While both objectives are subject to noise in gradient updates, we show through analysis and experiments that the variational lower bound is robust whereas the generalized Brown objective is vulnerable. We obtain strong performance on a multitude of datasets and languages with a simple architecture that encodes morphology and context.

Contextual word representations derived from large-scale neural language models are successful across a diverse set of NLP tasks, suggesting that they encode useful and transferable features of language. To shed light on the linguistic knowledge they capture, we study the representations produced by several recent pretrained contextualizers (variants of ELMo, the OpenAI transformer language model, and BERT) with a suite of sixteen diverse probing tasks. We find that linear models trained on top of frozen contextual representations are competitive with state-of-the-art task-specific models in many cases, but fail on tasks requiring fine-grained linguistic knowledge (e.g., conjunct identification). To investigate the transferability of contextual word representations, we quantify differences in the transferability of individual layers within contextualizers, especially between recurrent neural networks (RNNs) and transformers. For instance, higher layers of RNNs are more task-specific, while transformer layers do not exhibit the same monotonic trend. In addition, to better understand what makes contextual word representations transferable, we compare language model pretraining with eleven supervised pretraining tasks. For any given task, pretraining on a closely related task yields better performance than language model pretraining (which is better on average) when the pretraining dataset is fixed. However, language model pretraining on more data gives the best results.

Recent work has attempted to enhance vector space representations using information from structured semantic resources. This process, dubbed retrofitting (Faruqui et al., 2015), has yielded improvements in word similarity performance. Research has largely focused on the retrofitting algorithm, or on the kind of structured semantic resources used, but little research has explored why some resources perform better than others. We conducted a fine-grained analysis of the original retrofitting process, and found that the utility of different lexical resources for retrofitting depends on two factors: the coverage of the resource and the evaluation metric. Our assessment suggests that the common practice of using correlation measures to evaluate increases in performance against full word similarity benchmarks 1) obscures the benefits offered by smaller resources, and 2) overlooks incremental gains in word similarity performance. We propose root-mean-square error (RMSE) as an alternative evaluation metric, and demonstrate that correlation measures and RMSE sometimes yield opposite conclusions concerning the efficacy of retrofitting. This point is illustrated by word vectors retrofitted with novel treatments of the FrameNet data (Fillmore and Baker, 2010).

In traditional Distributional Semantic Models (DSMs) the multiple senses of a polysemous word are conflated into a single vector space representation. In this work, we propose a DSM that learns multiple distributional representations of a word based on different topics. First, a separate DSM is trained for each topic and then each of the topic-based DSMs is aligned to a common vector space. Our unsupervised mapping approach is motivated by the hypothesis that words preserving their relative distances in different topic semantic sub-spaces constitute robust semantic anchors that define the mappings between them. Aligned cross-topic representations achieve state-of-the-art results for the task of contextual word similarity. Furthermore, evaluation on NLP downstream tasks shows that multiple topic-based embeddings outperform single-prototype models.

A standard word embedding algorithm, such as word2vec and glove, makes a strong assumption that words are likely to be semantically related only if they co-occur locally within a window of fixed size. However, this strong assumption may not capture the semantic association between words that co-occur frequently but non-locally within documents. In this paper, we propose a graph-based word embedding method, named ‘word-node2vec’. By relaxing the strong constraint of locality, our method is able to capture both the local and non-local co-occurrences. Word-node2vec constructs a graph where every node represents a word and an edge between two nodes represents a combination of both local (e.g. word2vec) and document-level co-occurrences. Our experiments show that word-node2vec outperforms word2vec and glove on a range of different tasks, such as predicting word-pair similarity, word analogy and concept categorization.

Insufficient or even unavailable training data of emerging classes is a big challenge of many classification tasks, including text classification. Recognising text documents of classes that have never been seen in the learning stage, so-called zero-shot text classification, is therefore difficult and only limited previous works tackled this problem. In this paper, we propose a two-phase framework together with data augmentation and feature augmentation to solve this problem. Four kinds of semantic knowledge (word embeddings, class descriptions, class hierarchy, and a general knowledge graph) are incorporated into the proposed framework to deal with instances of unseen classes effectively. Experimental results show that each and the combination of the two phases achieve the best overall accuracy compared with baselines and recent approaches in classifying real-world texts under the zero-shot scenario.

Distant supervision (DS) is an important paradigm for automatically extracting relations. It utilizes existing knowledge base to collect examples for the relation we intend to extract, and then uses these examples to automatically generate the training data. However, the examples collected can be very noisy, and pose significant challenge for obtaining high quality labels. Previous work has made remarkable progress in predicting the relation from distant supervision, but typically ignores the temporal relations among those supervising instances. This paper formulates the problem of relation extraction with temporal reasoning and proposes a solution to predict whether two given entities participate in a relation at a given time spot. For this purpose, we construct a dataset called WIKI-TIME which additionally includes the valid period of a certain relation of two entities in the knowledge base. We propose a novel neural model to incorporate both the temporal information encoding and sequential reasoning. The experimental results show that, compared with the best of existing models, our model achieves better performance in both WIKI-TIME dataset and the well-studied NYT-10 dataset.

Extreme classification is a classification task on an extremely large number of labels (tags). User generated labels for any type of online data can be sparing per individual user but intractably large among all users. It would be useful to automatically select a smaller, standard set of labels to represent the whole label set. We can then solve efficiently the problem of multi-label learning with an intractably large number of interdependent labels, such as automatic tagging of Wikipedia pages. We propose a submodular maximization framework with linear cost to find informative labels which are most relevant to other labels yet least redundant with each other. A simple prediction model can then be trained on this label subset. Our framework includes both label-label and label-feature dependencies, which aims to find the labels with the most representation and prediction ability. In addition, to avoid information loss, we extract and predict outlier labels with weak dependency on other labels. We apply our model to four standard natural language data sets including Bibsonomy entries with users assigned tags, web pages with user assigned tags, legal texts with EUROVOC descriptors(A topic hierarchy with almost 4000 categories regarding different aspects of European law) and Wikipedia pages with tags from social bookmarking as well as news videos for automated label detection from a lexicon of semantic concepts. Experimental results show that our proposed approach improves label prediction quality, in terms of precision and nDCG, by 3% to 5% in three of the 5 tasks and is competitive in the others, even with a simple linear prediction model. An ablation study shows how different data sets benefit from different aspects of our model, with all aspects contributing substantially to at least one data set.

Modern weakly supervised methods for event detection (ED) avoid time-consuming human annotation and achieve promising results by learning from auto-labeled data. However, these methods typically rely on sophisticated pre-defined rules as well as existing instances in knowledge bases for automatic annotation and thus suffer from low coverage, topic bias, and data noise. To address these issues, we build a large event-related candidate set with good coverage and then apply an adversarial training mechanism to iteratively identify those informative instances from the candidate set and filter out those noisy ones. The experiments on two real-world datasets show that our candidate selection and adversarial training can cooperate together to obtain more diverse and accurate training data for ED, and significantly outperform the state-of-the-art methods in various weakly supervised scenarios. The datasets and source code can be obtained from https://github.com/thunlp/Adv-ED.

Knowledge graphs have been developed rapidly in recent years and shown their usefulness for many artificial intelligence tasks. However, knowledge graphs often have lots of missing facts. To solve this problem, many knowledge graph embedding models to populate knowledge graphs have been developed and have shown outstanding performance these days. However, knowledge graph embedding models are so called-black box. Hence, we actually does not know how information of a knowledge graph is processed and the models are hard to interpret. In this paper, we utilize graph patterns in a knowledge graph to overcome such problems. Our proposed model, graph pattern entity ranking Model (GRank), constructs an entity ranking system for each graph pattern and evaluate them using a measure for a ranking system. By doing so, we can find helpful graph patterns for predicting facts. Then we conduct the link prediction tasks on standard data sets to evaluate GRank. We show our approach outperforms other state-of-the-art approaches such as ComplEx and TorusE on standard metrics such as HITS@n and MRR. Moreover, This model is easily interpretable because output facts are described by graph patterns.