ACL 2015 论文列表

Comparisons in text, such as in online reviews, serve as useful decision aids. In this paper, we focus on the task of identifying whether a comparison exists between a specific pair of entity mentions in a sentence. This formulation is transformative, as previous work only seeks to determine whether a sentence is comparative, which is presumptuous in the event the sentence mentions multiple entities and is comparing only some, not all, of them. Our approach leverages not only lexical features such as salient words, but also structural features expressing the relationships among words and entity mentions. To model these features seamlessly, we rely on a dependency tree representation, and investigate the applicability of a series of tree kernels. This leads to the development of a new context-sensitive tree kernel: Skip-node Kernel (SNK). We further describe both its exact and approximate computations. Through experiments on real-life datasets, we evaluate the effectiveness of our kernel-based approach for comparison identification, as well as the utility of SNK and its approximations.

Prepositional phrases (PPs) express crucial information that knowledge base construction methods need to extract. However, PPs are a major source of syntactic ambiguity and still pose problems in parsing. We present a method for resolving ambiguities arising from PPs, making extensive use of semantic knowledge from various resources. As training data, we use both labeled and unlabeled data, utilizing an expectation maximization algorithm for parameter estimation. Experiments show that our method yields improvements over existing methods including a state of the art dependency parser.

A standard pipeline for statistical relational learning involves two steps: one first constructs the knowledge base (KB) from text, and then performs the learning and reasoning tasks using probabilistic first-order logics. However, a key issue is that information extraction (IE) errors from text affect the quality of the KB, and propagate to the reasoning task. In this paper, we propose a statistical relational learning model for joint information extraction and reasoning. More specifically, we incorporate context-based entity extraction with structure learning (SL) in a scalable probabilistic logic framework. We then propose a latent context invention (LCI) approach to improve the performance. In experiments, we show that our approach outperforms state-of-the-art baselines over three real-world Wikipedia datasets from multiple domains; that joint learning and inference for IE and SL significantly improve both tasks; that latent context invention further improves the results.

Relation triples produced by open domain information extraction (open IE) systems are useful for question answering, inference, and other IE tasks. Traditionally these are extracted using a large set of patterns; however, this approach is brittle on out-of-domain text and long-range dependencies, and gives no insight into the substructure of the arguments. We replace this large pattern set with a few patterns for canonically structured sentences, and shift the focus to a classifier which learns to extract self-contained clauses from longer sentences. We then run natural logic inference over these short clauses to determine the maximally specific arguments for each candidate triple. We show that our approach outperforms a state-of-the-art open IE system on the end-to-end TAC-KBP 2013 Slot Filling task.

This work was sponsored in part by the U. S. Army Research Laboratory and the U. S. Army Research Office/nunder contract/grant number W911NF-10-1-0533, and in part by NSF CAREER grant IIS-1054319./nMiguel Ballesteros is supported by the European Commission under the contract numbers FP7-ICT-610411 (project MULTISENSOR) and H2020-RIA-645012 (project KRISTINA).

We present structured perceptron training for neural network transition-based dependency parsing. We learn the neural network representation using a gold corpus augmented by a large number of automatically parsed sentences. Given this fixed network representation, we learn a final layer using the structured perceptron with beam-search decoding. On the Penn Treebank, our parser reaches 94.26% unlabeled and 92.41% labeled attachment accuracy, which to our knowledge is the best accuracy on Stanford Dependencies to date. We also provide indepth ablative analysis to determine which aspects of our model provide the largest gains in accuracy.

Most existing graph-based parsing models rely on millions of hand-crafted features, which limits their generalization ability and slows down the parsing speed. In this paper, we propose a general and effective Neural Network model for graph-based dependency parsing. Our model can automatically learn high-order feature combinations using only atomic features by exploiting a novel activation function tanhcube. Moreover, we propose a simple yet effective way to utilize phrase-level information that is expensive to use in conventional graph-based parsers. Experiments on the English Penn Treebank show that parsers based on our model perform better than conventional graph-based parsers.

This paper describes a parsing model that combines the exact dynamic programming of CRF parsing with the rich nonlinear featurization of neural net approaches. Our model is structurally a CRF that factors over anchored rule productions, but instead of linear potential functions based on sparse features, we use nonlinear potentials computed via a feedforward neural network. Because potentials are still local to anchored rules, structured inference (CKY) is unchanged from the sparse case. Computing gradients during learning involves backpropagating an error signal formed from standard CRF sufficient statistics (expected rule counts). Using only dense features, our neural CRF already exceeds a strong baseline CRF model (Hall et al., 2014). In combination with sparse features, our system achieves 91.1 F1 on section 23 of the Penn Treebank, and more generally outperforms the best prior single parser results on a range of languages.

An elementary way of using language is to refer to objects. Often, these objects are physically present in the shared environment and reference is done via mention of perceivable properties of the objects. This is a type of language use that is modelled well neither by logical semantics nor by distributional semantics, the former focusing on inferential relations between expressed propositions, the latter on similarity relations between words or phrases. We present an account of word and phrase meaning that is perceptually grounded, trainable, compositional, and ‘dialogueplausible’ in that it computes meanings word-by-word. We show that the approach performs well (with an accuracy of 65% on a 1-out-of-32 reference resolution task) on direct descriptions and target/landmark descriptions, even when trained with less than 800 training examples and automatically transcribed utterances.

Over the last two decades, numerous algorithms have been developed that successfully capture something of the semanticsof single words by looking at their distribution in text and comparing these distributions in a vector space model. However, it is not straightforward to construct meaning representations beyond the levelof individual words–i.e. the combination of words into larger units – using dis-tributional methods. Our contribution is twofold. First of all, we carry out a large-scale evaluation, comparing different composition methods within the distributional framework for the cases of both adjective-noun and noun-noun composition, makinguse of a newly developed dataset. Secondly, we propose a novel method focomposition, which generalises the approach by Baroni and Zamparelli (2010). The performance of our novel method is also evaluated on our new dataset andproves competitive with the best methods.

Zero-shot methods in language, vision and other domains rely on a cross-space mapping function that projects vectors from the relevant feature space (e.g., visualfeature-based image representations) to a large semantic word space (induced in an unsupervised way from corpus data), where the entities of interest (e.g., objects images depict) are labeled with the words associated to the nearest neighbours of the mapped vectors. Zero-shot cross-space mapping methods hold great promise as a way to scale up annotation tasks well beyond the labels in the training data (e.g., recognizing objects that were never seen in training). However, the current performance of cross-space mapping functions is still quite low, so that the strategy is not yet usable in practical applications. In this paper, we explore some general properties, both theoretical and empirical, of the cross-space mapping function, and we build on them to propose better methods to estimate it. In this way, we attain large improvements over the state of the art, both in cross-linguistic (word translation) and cross-modal (image labeling) zero-shot experiments.

Answering natural language questions over a knowledge base is an important and challenging task. Most of existing systems typically rely on hand-crafted features and rules to conduct question understanding and/or answer ranking. In this paper, we introduce multi-column convolutional neural networks (MCCNNs) to understand questions from three different aspects (namely, answer path, answer context, and answer type) and learn their distributed representations. Meanwhile, we jointly learn low-dimensional embeddings of entities and relations in the knowledge base. Question-answer pairs are used to train the model to rank candidate answers. We also leverage question paraphrases to train the column networks in a multi-task learning manner. We use FREEBASE as the knowledge base and conduct extensive experiments on the WEBQUESTIONS dataset. Experimental results show that our method achieves better or comparable performance compared with baseline systems. In addition, we develop a method to compute the salience scores of question words in different column networks. The results help us intuitively understand what MCCNNs learn.

Community question answering (cQA) has become an important issue due to the popularity of cQA archives on the web. This paper is concerned with the problem of question retrieval. Question retrieval in cQA archives aims to find the existing questions that are semantically equivalent or relevant to the queried questions. However, the lexical gap problem brings about new challenge for question retrieval in cQA. In this paper, we propose to learn continuous word embeddings with metadata of category information within cQA pages for question retrieval. To deal with the variable size of word embedding vectors, we employ the framework of fisher kernel to aggregated them into the fixedlength vectors. Experimental results on large-scale real world cQA data set show that our approach can significantly outperform state-of-the-art translation models and topic-based models for question re-

Understanding open-domain text is one of the primary challenges in NLP. Machine comprehension evaluates the system’s ability to understand text through a series of question-answering tasks on short pieces of text such that the correct answer can be found only in the given text. For this task, we posit that there is a hidden (latent) structure that explains the relation between the question, correct answer, and text. We call this the answer-entailing structure; given the structure, the correctness of the answer is evident. Since the structure is latent, it must be inferred. We present a unified max-margin framework that learns to find these hidden structures (given a corpus of question-answer pairs), and uses what it learns to answer machine comprehension questions on novel texts. We extend this framework to incorporate multi-task learning on the different subtasks that are required to perform machine comprehension. Evaluation on a publicly available dataset shows that our framework outperforms various IR and neuralnetwork baselines, achieving an overall accuracy of 67.8% (vs. 59.9%, the best previously-published result.)

Lexical selection is crucial for statistical machine translation. Previous studies separately exploit sentence-level contexts and documentlevel topics for lexical selection, neglecting their correlations. In this paper, we propose a context-aware topic model for lexical selection, which not only models local contexts and global topics but also captures their correlations. The model uses target-side translations as hidden variables to connect document topics and source-side local contextual words. In order to learn hidden variables and distributions from data, we introduce a Gibbs sampling algorithm for statistical estimation and inference. A new translation probability based on distributions learned by the model is integrated into a translation system for lexical selection. Experiment results on NIST ChineseEnglish test sets demonstrate that 1) our model significantly outperforms previous lexical selection methods and 2) modeling correlations between local words and global topics can further improve translation quality.

We present a method for predicting machine translation output quality geared to the needs of computer-assisted translation. These include the capability to: i) continuously learn and self-adapt to a stream of data coming from multiple translation jobs, ii) react to data diversity by exploiting human feedback, and iii) leverage data similarity by learning and transferring knowledge across domains. To achieve these goals, we combine two supervised machine learning paradigms, online and multitask learning, adapting and unifying them in a single framework. We show the effectiveness of our approach in a regression task (HTER prediction), in which online multitask learning outperforms the competitive online single-task and pooling methods used for comparison. This indicates the feasibility of integrating in a CAT tool a single QE component capable to simultaneously serve (and continuously learn from) multiple translation jobs involving different domains and users.

We present an efficient incremental topdown parsing method for preordering based on Bracketing Transduction Grammar (BTG). The BTG-based preordering framework (Neubig et al., 2012) can be applied to any language using only parallel text, but has the problem of computational efficiency. Our top-down parsing algorithm allows us to use the early update technique easily for the latent variable structured Perceptron algorithm with beam search, and solves the problem. Experimental results showed that the topdown method is more than 10 times faster than a method using the CYK algorithm. A phrase-based machine translation system with the top-down method had statistically significantly higher BLEU scores for 7 language pairs without relying on supervised syntactic parsers, compared to baseline systems using existing preordering methods.

Simultaneous translation is a method to reduce the latency of communication through machine translation (MT) by dividing the input into short segments before performing translation. However, short segments pose problems for syntaxbased translation methods, as it is difficult to generate accurate parse trees for sub-sentential segments. In this paper, we perform the first experiments applying syntax-based SMT to simultaneous translation, and propose two methods to prevent degradations in accuracy: a method to predict unseen syntactic constituents that help generate complete parse trees, and a method that waits for more input when the current utterance is not enough to generate a fluent translation. Experiments on English-Japanese translation show that the proposed methods allow for improvements in accuracy, particularly with regards to word order of the target sentences.

This paper presents a generative model to event schema induction. Previous methods in the literature only use head words to represent entities. However, elements other than head words contain useful information. For instance, an armed man is more discriminative than man. Our model takes into account this information and precisely represents it using probabilistic topic distributions. We illustrate that such information plays an important role in parameter estimation. Mostly, it makes topic distributions more coherent and more discriminative. Experimental results on benchmark dataset empirically confirm this enhancement.