NAACL 2019 论文列表

Meaning conflation deficiency is one of the main limiting factors of word representations which, given their widespread use at the core of many NLP systems, can lead to inaccurate semantic understanding of the input text and inevitably hamper the performance. Sense representations target this problem. However, their potential impact has rarely been investigated in downstream NLP applications. Through a set of experiments on a state-of-the-art reverse dictionary system based on neural networks, we show that a simple adjustment aimed at addressing the meaning conflation deficiency can lead to substantial improvements.

Abuse on the Internet represents a significant societal problem of our time. Previous research on automated abusive language detection in Twitter has shown that community-based profiling of users is a promising technique for this task. However, existing approaches only capture shallow properties of online communities by modeling follower–following relationships. In contrast, working with graph convolutional networks (GCNs), we present the first approach that captures not only the structure of online communities but also the linguistic behavior of the users within them. We show that such a heterogeneous graph-structured modeling of communities significantly advances the current state of the art in abusive language detection.

We present a new dataset comprised of 210,532 tokens evenly drawn from 100 different English-language literary texts annotated for ACE entity categories (person, location, geo-political entity, facility, organization, and vehicle). These categories include non-named entities (such as “the boy”, “the kitchen”) and nested structure (such as [[the cook]’s sister]). In contrast to existing datasets built primarily on news (focused on geo-political entities and organizations), literary texts offer strikingly different distributions of entity categories, with much stronger emphasis on people and description of settings. We present empirical results demonstrating the performance of nested entity recognition models in this domain; training natively on in-domain literary data yields an improvement of over 20 absolute points in F-score (from 45.7 to 68.3), and mitigates a disparate impact in performance for male and female entities present in models trained on news data.

Peer-review plays a critical role in the scientific writing and publication ecosystem. To assess the efficiency and efficacy of the reviewing process, one essential element is to understand and evaluate the reviews themselves. In this work, we study the content and structure of peer reviews under the argument mining framework, through automatically detecting (1) the argumentative propositions put forward by reviewers, and (2) their types (e.g., evaluating the work or making suggestions for improvement). We first collect 14.2K reviews from major machine learning and natural language processing venues. 400 reviews are annotated with 10,386 propositions and corresponding types of Evaluation, Request, Fact, Reference, or Quote. We then train state-of-the-art proposition segmentation and classification models on the data to evaluate their utilities and identify new challenges for this new domain, motivating future directions for argument mining. Further experiments show that proposition usage varies across venues in amount, type, and topic.

We explore the challenge of action prediction from textual descriptions of scenes, a testbed to approximate whether text inference can be used to predict upcoming actions. As a case of study, we consider the world of the Harry Potter fantasy novels and inferring what spell will be cast next given a fragment of a story. Spells act as keywords that abstract actions (e.g. ‘Alohomora’ to open a door) and denote a response to the environment. This idea is used to automatically build HPAC, a corpus containing 82,836 samples and 85 actions. We then evaluate different baselines. Among the tested models, an LSTM-based approach obtains the best performance for frequent actions and large scene descriptions, but approaches such as logistic regression behave well on infrequent actions.

A pun is a form of wordplay for an intended humorous or rhetorical effect, where a word suggests two or more meanings by exploiting polysemy (homographic pun) or phonological similarity to another word (heterographic pun). This paper presents an approach that addresses pun detection and pun location jointly from a sequence labeling perspective. We employ a new tagging scheme such that the model is capable of performing such a joint task, where useful structural information can be properly captured. We show that our proposed model is effective in handling both homographic and heterographic puns. Empirical results on the benchmark datasets demonstrate that our approach can achieve new state-of-the-art results.

In the context of fake news, bias, and propaganda, we study two important but relatively under-explored problems: (i) trustworthiness estimation (on a 3-point scale) and (ii) political ideology detection (left/right bias on a 7-point scale) of entire news outlets, as opposed to evaluating individual articles. In particular, we propose a multi-task ordinal regression framework that models the two problems jointly. This is motivated by the observation that hyper-partisanship is often linked to low trustworthiness, e.g., appealing to emotions rather than sticking to the facts, while center media tend to be generally more impartial and trustworthy. We further use several auxiliary tasks, modeling centrality, hyper-partisanship, as well as left-vs.-right bias on a coarse-grained scale. The evaluation results show sizable performance gains by the joint models over models that target the problems in isolation.

There exist biases in individual’s language use; the same word (e.g., cool) is used for expressing different meanings (e.g., temperature range) or different words (e.g., cloudy, hazy) are used for describing the same meaning. In this study, we propose a method of modeling such personal biases in word meanings (hereafter, semantic variations) with personalized word embeddings obtained by solving a task on subjective text while regarding words used by different individuals as different words. To prevent personalized word embeddings from being contaminated by other irrelevant biases, we solve a task of identifying a review-target (objective output) from a given review. To stabilize the training of this extreme multi-class classification, we perform a multi-task learning with metadata identification. Experimental results with reviews retrieved from RateBeer confirmed that the obtained personalized word embeddings improved the accuracy of sentiment analysis as well as the target task. Analysis of the obtained personalized word embeddings revealed trends in semantic variations related to frequent and adjective words.

Tweets are short messages that often include specialized language such as hashtags and emojis. In this paper, we present a simple strategy to process emojis: replace them with their natural language description and use pretrained word embeddings as normally done with standard words. We show that this strategy is more effective than using pretrained emoji embeddings for tweet classification. Specifically, we obtain new state-of-the-art results in irony detection and sentiment analysis despite our neural network is simpler than previous proposals.

A growing number of state-of-the-art transfer learning methods employ language models pretrained on large generic corpora. In this paper we present a conceptually simple and effective transfer learning approach that addresses the problem of catastrophic forgetting. Specifically, we combine the task-specific optimization function with an auxiliary language model objective, which is adjusted during the training process. This preserves language regularities captured by language models, while enabling sufficient adaptation for solving the target task. Our method does not require pretraining or finetuning separate components of the network and we train our models end-to-end in a single step. We present results on a variety of challenging affective and text classification tasks, surpassing well established transfer learning methods with greater level of complexity.

Collaborative filtering (CF) is a core technique for recommender systems. Traditional CF approaches exploit user-item relations (e.g., clicks, likes, and views) only and hence they suffer from the data sparsity issue. Items are usually associated with unstructured text such as article abstracts and product reviews. We develop a Personalized Neural Embedding (PNE) framework to exploit both interactions and words seamlessly. We learn such embeddings of users, items, and words jointly, and predict user preferences on items based on these learned representations. PNE estimates the probability that a user will like an item by two terms—behavior factors and semantic factors. On two real-world datasets, PNE shows better performance than four state-of-the-art baselines in terms of three metrics. We also show that PNE learns meaningful word embeddings by visualization.

We examine the new task of detecting derogatory compounds (e.g. “curry muncher”). Derogatory compounds are much more difficult to detect than derogatory unigrams (e.g. “idiot”) since they are more sparsely represented in lexical resources previously found effective for this task (e.g. Wiktionary). We propose an unsupervised classification approach that incorporates linguistic properties of compounds. It mostly depends on a simple distributional representation. We compare our approach against previously established methods proposed for extracting derogatory unigrams.

We present the first exploration of meaning shift over short periods of time in online communities using distributional representations. We create a small annotated dataset and use it to assess the performance of a standard model for meaning shift detection on short-term meaning shift. We find that the model has problems distinguishing meaning shift from referential phenomena, and propose a measure of contextual variability to remedy this.

Continued training is an effective method for domain adaptation in neural machine translation. However, in-domain gains from adaptation come at the expense of general-domain performance. In this work, we interpret the drop in general-domain performance as catastrophic forgetting of general-domain knowledge. To mitigate it, we adapt Elastic Weight Consolidation (EWC)—a machine learning method for learning a new task without forgetting previous tasks. Our method retains the majority of general-domain performance lost in continued training without degrading in-domain performance, outperforming the previous state-of-the-art. We also explore the full range of general-domain performance available when some in-domain degradation is acceptable.

We consider the problem of making efficient use of heterogeneous training data in neural machine translation (NMT). Specifically, given a training dataset with a sentence-level feature such as noise, we seek an optimal curriculum, or order for presenting examples to the system during training. Our curriculum framework allows examples to appear an arbitrary number of times, and thus generalizes data weighting, filtering, and fine-tuning schemes. Rather than relying on prior knowledge to design a curriculum, we use reinforcement learning to learn one automatically, jointly with the NMT system, in the course of a single training run. We show that this approach can beat uniform baselines on Paracrawl and WMT English-to-French datasets by +3.4 and +1.3 BLEU respectively. Additionally, we match the performance of strong filtering baselines and hand-designed, state-of-the-art curricula.

Despite some empirical success at correcting exposure bias in machine translation, scheduled sampling algorithms suffer from a major drawback: they incorrectly assume that words in the reference translations and in sampled sequences are aligned at each time step. Our new differentiable sampling algorithm addresses this issue by optimizing the probability that the reference can be aligned with the sampled output, based on a soft alignment predicted by the model itself. As a result, the output distribution at each time step is evaluated with respect to the whole predicted sequence. Experiments on IWSLT translation tasks show that our approach improves BLEU compared to maximum likelihood and scheduled sampling baselines. In addition, our approach is simpler to train with no need for sampling schedule and yields models that achieve larger improvements with smaller beam sizes.

Incremental domain adaptation, in which a system learns from the correct output for each input immediately after making its prediction for that input, can dramatically improve system performance for interactive machine translation. Users of interactive systems are sensitive to the speed of adaptation and how often a system repeats mistakes, despite being corrected. Adaptation is most commonly assessed using corpus-level BLEU- or TER-derived metrics that do not explicitly take adaptation speed into account. We find that these metrics often do not capture immediate adaptation effects, such as zero-shot and one-shot learning of domain-specific lexical items. To this end, we propose new metrics that directly evaluate immediate adaptation performance for machine translation. We use these metrics to choose the most suitable adaptation method from a range of different adaptation techniques for neural machine translation systems.

The explicit use of syntactic information has been proved useful for neural machine translation (NMT). However, previous methods resort to either tree-structured neural networks or long linearized sequences, both of which are inefficient. Neural syntactic distance (NSD) enables us to represent a constituent tree using a sequence whose length is identical to the number of words in the sentence. NSD has been used for constituent parsing, but not in machine translation. We propose five strategies to improve NMT with NSD. Experiments show that it is not trivial to improve NMT with NSD; however, the proposed strategies are shown to improve translation performance of the baseline model (+2.1 (En–Ja), +1.3 (Ja–En), +1.2 (En–Ch), and +1.0 (Ch–En) BLEU).

We present a robust neural abstractive summarization system for cross-lingual summarization. We construct summarization corpora for documents automatically translated from three low-resource languages, Somali, Swahili, and Tagalog, using machine translation and the New York Times summarization corpus. We train three language-specific abstractive summarizers and evaluate on documents originally written in the source languages, as well as on a fourth, unseen language: Arabic. Our systems achieve significantly higher fluency than a standard copy-attention summarizer on automatically translated input documents, as well as comparable content selection.