NAACL 2016 Papers

Metaphor is pervasive in our communication, which makes it an important problem for nat-ural language processing (NLP). Numerous approaches to metaphor processing have thus been proposed, all of which relied on linguistic features and textual data to construct their models. Human metaphor comprehension is, however, known to rely on both our linguistic and perceptual experience, and vision can play a particularly important role when metaphorically projecting imagery across domains. In this paper, we present the first metaphor identification method that simultaneously draws knowledge from linguistic and visual data. Our results demonstrate that it outperforms linguistic and visual models in isolation, as well as being competitive with the best-performing metaphor identification methods, that rely on hand-crafted knowledge about domains and perception.

,

In this work, we present a novel counter-fitting method which injects antonymy and synonymy constraints into vector space representations in order to improve the vectors' capability for judging semantic similarity. Applying this method to publicly available pre-trained word vectors leads to a new state of the art performance on the SimLex-999 dataset. We also show how the method can be used to tailor the word vector space for the downstream task of dialogue state tracking, resulting in robust improvements across different dialogue domains.

Public debates are a common platform for presenting and juxtaposing diverging views on important issues. In this work we propose a methodology for tracking how ideas flow between participants throughout a debate. We use this approach in a case study of Oxford-style debates---a competitive format where the winner is determined by audience votes---and show how the outcome of a debate depends on aspects of conversational flow. In particular, we find that winners tend to make better use of a debate's interactive component than losers, by actively pursuing their opponents' points rather than promoting their own ideas over the course of the conversation.

We propose a first-ever attempt to employ a Long Short-Term memory based framework to predict humor in dialogues. We analyze data from a popular TV-sitcom, whose canned laughters give an indication of when the audience would react. We model the setup-punchline relation of conversational humor with a Long Short-Term Memory, with utter-ance encodings obtained from a Convolutional Neural Network. Out neural network framework is able to improve the F-score of 8% over a Conditional Random Field baseline. We show how the LSTM effectively models the setup-punchline relation reducing the number of false positives and increasing the recall. We aim to employ our humor prediction model to build effective empathetic machine able to understand jokes.

©2016 Association for Computational Linguistics. Moving from limited-domain natural language generation (NLG) to open domain is difficult because the number of semantic input combinations grows exponentially with the number of domains. Therefore, it is important to leverage existing resources and exploit similarities between domains to facilitate domain adaptation. In this paper, we propose a procedure to train multi-domain, Recurrent Neural Network-based (RNN) language generators via multiple adaptation steps. In this procedure, a model is first trained on counterfeited data synthesised from an out-of-domain dataset, and then fine tuned on a small set of in-domain utterances with a discriminative objective function. Corpus-based evaluation results show that the proposed procedure can achieve competitive performance in terms of BLEU score and slot error rate while significantly reducing the data needed to train generators in new, unseen domains. In subjective testing, human judges confirm that the procedure greatly improves generator performance when only a small amount of data is available in the domain.

Sequence-to-sequence neural network models for generation of conversational responses tend to generate safe, commonplace responses (e.g., "I don't know") regardless of the input. We suggest that the traditional objective function, i.e., the likelihood of output (response) given input (message) is unsuited to response generation tasks. Instead we propose using Maximum Mutual Information (MMI) as the objective function in neural models. Experimental results demonstrate that the proposed MMI models produce more diverse, interesting, and appropriate responses, yielding substantive gains in BLEU scores on two conversational datasets and in human evaluations.

In this paper we present the first, to the best of our knowledge, discourse parser that is able to predict non-tree DAG structures. We use Integer Linear Programming (ILP) to encode both the objective function and the constraints as global decoding over local scores. Our underlying data come from multi-party chat dialogues, which require the prediction of DAGs. We use the dependency parsing paradigm, as has been done in the past (Muller et al., 2012; Li et al., 2014; Afantenos et al., 2015), but we use the underlying formal framework of SDRT and exploit SDRT's notions of left and right distributive relations. We achieve an F-measure of 0.531 for fully labeled structures which beats the previous state of the art.

Abstractive Sentence Summarization generates a shorter version of a given sentence while attempting to preserve its meaning. We introduce a conditional recurrent neural network (RNN) which generates a summary of an input sentence. The conditioning is provided by a novel convolutional attention-based encoder which ensures that the decoder focuses on the appropriate input words at each step of generation. Our model relies only on learned features and is easy to train in an end-to-end fashion on large data sets. Our experiments show that the model significantly outperforms the recently proposed state-of-the-art method on the Giga-word corpus while performing competitively on the DUC-2004 shared task.

A corpus of inference rules between a pair of relation phrases is typically generated using the statistical overlap of argument-pairs associated with the relations (e.g., PATTY, CLEAN). We investigate knowledge-guided linguistic rewrites as a secondary source of evidence and find that they can vastly improve the quality of inference rule corpora, obtaining 27 to 33 point precision improvement while retaining substantial recall. The facts inferred using cleaned inference rules are 29-32 points more accurate.

Student course feedback is generated daily in both classrooms and online course discussion forums. Traditionally, instructors manually analyze these responses in a costly manner. In this work, we propose a new approach to summarizing student course feedback based on the integer linear programming (ILP) framework. Our approach allows different student responses to share co-occurrence statistics and alleviates sparsity issues. Experimental results on a student feedback corpus show that our approach outperforms a range of baselines in terms of both ROUGE scores and human evaluation.

Community created content (e.g., product descriptions, reviews) typically discusses one entity at a time and it can be hard as well as time consuming for a user to compare two or more entities. In response, we de-fine a novel task of automatically generating entity comparisons from text. Our output is a table that semantically clusters descriptive phrases about entities. Our clustering algo-rithm is a Gaussian extension of probabilistic latent semantic analysis (pLSA), in which each phrase is represented in word vector embedding space. In addition, our algorithm attempts to balance information about entities in each cluster to generate meaningful comparison tables, where possible. We test our system’s effectiveness on two domains, travel articles and movie reviews, and find that entity-balanced clusters are strongly preferred by users.

A key challenge for timeline summarization is to generate a concise, yet complete storyline from large collections of news stories. Previous studies in extractive timeline generation are limited in two ways: first, most prior work focuses on fully-observable ranking models or clustering models with hand-designed features that may not generalize well. Second, most summarization corpora are text-only, which means that text is the sole source of information considered in timeline summarization, and thus, the rich visual content from news images is ignored. To solve these issues, we leverage the success of matrix factorization techniques from recommender systems, and cast the problem as a sentence recommendation task, using a representation learning approach. To augment text-only corpora, for each candidate sentence in a news article, we take advantage of top-ranked relevant images from the Web and model the image using a convolutional neural network architecture. Finally, we propose a scalable low-rank approximation approach for learning joint embeddings of news stories and images. In experiments, we compare our model to various competitive baselines, and demonstrate the stateof-the-art performance of the proposed textbased and multimodal approaches.

We study the problem of generating abstractive summaries for opinionated text. We propose an attention-based neural network model that is able to absorb information from multiple text units to construct informative, concise, and fluent summaries. An importance-based sampling method is designed to allow the encoder to integrate information from an important subset of input. Automatic evaluation indicates that our system outperforms state-of-the-art abstractive and extractive summarization systems on two newly collected datasets of movie reviews and arguments. Our system summaries are also rated as more informative and grammatical in human evaluation.

The neural network joint model of translation or NNJM (Devlin et al., 2014) combines source and target context to produce a powerful translation feature. However, its softmax layer necessitates a sum over the entire output vocabulary, which results in very slow maximum likelihood (MLE) training. This has led some groups to train using Noise Contrastive Estimation (NCE), which side-steps this sum. We carry out the first direct comparison of MLE and NCE training objectives for the NNJM, showing that NCE is significantly outperformed by MLE on large-scale ArabicEnglish and Chinese-English translation tasks. We also show that this drop can be avoided by using a recently proposed translation noise distribution.

Many languages use honorifics to express politeness, social distance, or the relative social status between the speaker and their ad-dressee(s). In machine translation from a language without honorifics such as English, it is difficult to predict the appropriate honorific, but users may want to control the level of politeness in the output. In this paper, we perform a pilot study to control honorifics in neural machine translation (NMT) via side constraints , focusing on English → German. We show that by marking up the (English) source side of the training data with a feature that en-codes the use of honorifics on the (German) target side, we can control the honorifics produced at test time. Experiments show that the choice of honorifics has a big impact on translation quality as measured by B LEU , and oracle experiments show that substantial im-provements are possible by constraining the translation to the desired level of politeness.

We build a multi-source machine translation model and train it to maximize the probability of a target English string given French and German sources. Using the neural encoder-decoder framework, we explore several combination methods and report up to +4.8 Bleu increases on top of a very strong attention-based neural translation model.

Active learning is a framework that makes it possible to efficiently train statistical models by selecting informative examples from a pool of unlabeled data. Previous work has found this framework effective for machine translation (MT), making it possible to train better translation models with less effort, particularly when annotators translate short phrases instead of full sentences. However, previous methods for phrase-based active learning in MT fail to consider whether the selected units are coherent and easy for human translators to translate, and also have problems with selecting redundant phrases with similar content. In this paper, we tackle these problems by proposing two new methods for selecting more syntactically coherent and less redundant segments in active learning for MT. Experiments using both simulation and extensive manual translation by professional translators find the proposed method effective, achieving both greater gain of BLEU score for the same number of translated words, and allowing translators to be more confident in their translations 1 .

A major benefit of tree-to-tree over tree-to-string translation is that we can use target-side syntax to improve reordering. While this is relatively simple for binarized constituency parses, the reordering problem is considerably harder for dependency parses, in which words can have arbitrarily many children. Previous approaches have tackled this problem by restricting grammar rules, reducing the expressive power of the translation model. In this paper we propose a general model for dependency tree-to-tree reordering based on flexible non-terminals that can compactly encode multiple insertion positions. We explore how insertion positions can be selected even in cases where rules do not entirely cover the children of input sentence words. The proposed method greatly improves the flexibility of translation rules at the cost of only a 30% increase in decoding time, and we demonstrate a 1.2–1.9 BLEU improvement over a strong tree-to-tree baseline.