ACL 2015 论文列表

Neural probabilistic parsers are attractive for their capability of automatic feature combination and small data sizes. A transition-based greedy neural parser has given better accuracies over its linear counterpart. We propose a neural probabilistic structured-prediction model for transition-based dependency parsing, which integrates search and learning. Beam search is used for decoding, and contrastive learning is performed for maximizing the sentence-level log-likelihood. In standard Penn Treebank experiments, the structured neural parser achieves a 1.8% accuracy improvement upon a competitive greedy neural parser baseline, giving performance comparable to the best linear parser.

We present an extension to incremental shift-reduce parsing that handles discontinuous constituents, using a linear classifier and beam search. We achieve very high parsing speeds (up to 640 sent./sec.) and accurate results (up to 79.52 F1 on TiGer).

Incremental parsing is the task of assigning a syntactic structure to an input sentence as it unfolds word by word. Incremental parsing is more difficult than fullsentence parsing, as incomplete input increases ambiguity. Intuitively, an incremental parser that has access to semantic information should be able to reduce ambiguity by ruling out semantically implausible analyses, even for incomplete input. In this paper, we test this hypothesis by combining an incremental TAG parser with an incremental semantic role labeler in a discriminative framework. We show a substantial improvement in parsing performance compared to the baseline parser, both in full-sentence F-score and in incremental F-score.

Given a set of basic binary features, we propose a new L1 norm SVM based feature selection method that explicitly selects the features in their polynomial or tree kernel spaces. The efficiency comes from the anti-monotone property of the subgradients: the subgradient with respect to a combined feature can be bounded by the subgradient with respect to each of its component features, and a feature can be pruned safely without further consideration if its corresponding subgradient is not steep enough. We conduct experiments on the English dependency parsing task with a third order graph-based parser. Benefiting from the rich features selected in the tree kernel space, our model achieved the best reported unlabeled attachment score of 93.72 without using any additional resource.

Constituent parsing is typically modeled by a chart-based algorithm under probabilistic context-free grammars or by a transition-based algorithm with rich features. Previous models rely heavily on richer syntactic information through lexicalizing rules, splitting categories, or memorizing long histories. However enriched models incur numerous parameters and sparsity issues, and are insufficient for capturing various syntactic phenomena. We propose a neural network structure that explicitly models the unbounded history of actions performed on the stack and queue employed in transition-based parsing, in addition to the representations of partially parsed tree structure. Our transition-based neural constituent parsing achieves performance comparable to the state-of-the-art parsers, demonstrating F1 score of 90.68% for English and 84.33% for Chinese, without reranking, feature templates or additional data to train model parameters.

In this work, we address the problem to model all the nodes (words or phrases) in a dependency tree with the dense representations. We propose a recursive convolutional neural network (RCNN) architecture to capture syntactic and compositional-semantic representations of phrases and words in a dependency tree. Different with the original recursive neural network, we introduce the convolution and pooling layers, which can model a variety of compositions by the feature maps and choose the most informative compositions by the pooling layers. Based on RCNN, we use a discriminative model to re-rank a $k$-best list of candidate dependency parsing trees. The experiments show that RCNN is very effective to improve the state-of-the-art dependency parsing on both English and Chinese datasets.

Dependency parsers are usually evaluated on attachment accuracy. Whilst easily interpreted, the metric does not illustrate the cascading impact of errors, where the parser chooses an incorrect arc, and is subsequently forced to choose further incorrect arcs elsewhere in the parse. We apply arc-level constraints to MSTparser and ZPar, enforcing the correct analysis of specific error classes, whilst otherwise continuing with decoding. We investigate the direct and indirect impact of applying constraints to the parser. Erroneous NP and punctuation attachments cause the most cascading errors, while incorrect PP and coordination attachments are frequent but less influential. Punctuation is especially challenging, as it has long been ignored in parsing, and serves a variety of disparate syntactic roles.

The performance of discriminative constituent parsing relies crucially on feature engineering, and effective features usually have to be carefully selected through a painful manual process. In this paper, we propose to automatically learn a set of effective features via neural networks. Specifically, we build a feedforward neural network model, which takes as input a few primitive units (words, POS tags and certain contextual tokens) from the local context, induces the feature representation in the hidden layer and makes parsing predictions in the output layer. The network simultaneously learns the feature representation and the prediction model parameters using a back propagation algorithm. By pre-training the model on a large amount of automatically parsed data, and then fine-tuning on the manually annotated Treebank data, our parser achieves the highest F1 score at 86.6% on Chinese Treebank 5.1, and a competitive F1 score at 90.7% on English Treebank. More importantly, our parser generalizes well on cross-domain test sets, where we significantly outperform Berkeley parser by 3.4 points on average for Chinese and 2.5 points for English.

Semantic role labeling (SRL) is one of the basic natural language processing (NLP) problems. To this date, most of the successful SRL systems were built on top of some form of parsing results (Koomen et al., 2005; Palmer et al., 2010; Pradhan et al., 2013), where pre-defined feature templates over the syntactic structure are used. The attempts of building an end-to-end SRL learning system without using parsing were less successful (Collobert et al., 2011). In this work, we propose to use deep bi-directional recurrent network as an end-to-end system for SRL. We take only original text information as input feature, without using any syntactic knowledge. The proposed algorithm for semantic role labeling was mainly evaluated on CoNLL-2005 shared task and achieved F1 score of 81.07. This result outperforms the previous state-of-the-art system from the combination of different parsing trees or models. We also obtained the same conclusion with F1 = 81.27 on CoNLL2012 shared task. As a result of simplicity, our model is also computationally efficient that the parsing speed is 6.7k tokens per second. Our analysis shows that our model is better at handling longer sentences than traditional models. And the latent variables of our model implicitly capture the syntactic structure of a sentence.

Complex conjunctions and determiners are often considered as pretokenized units in parsing. This is not always realistic, since they can be ambiguous. We propose a model for joint dependency parsing and multiword expressions identification, in which complex function words are represented as individual tokens linked with morphological dependencies. Our graph-based parser includes standard second-order features and verbal subcategoriza-tion features derived from a syntactic lexicon .We train it on a modified version of the French Treebank enriched with morphological dependencies. It recognizes 81.79% of ADV+que conjunctions with 91.57% precision, and 82.74% of de+DET determiners with 86.70% precision.

Natural language generation of coherent long texts like paragraphs or longer documents is a challenging problem for recurrent networks models. In this paper, we explore an important step toward this generation task: training an LSTM (Longshort term memory) auto-encoder to preserve and reconstruct multi-sentence paragraphs. We introduce an LSTM model that hierarchically builds an embedding for a paragraph from embeddings for sentences and words, then decodes this embedding to reconstruct the original paragraph. We evaluate the reconstructed paragraph using standard metrics like ROUGE and Entity Grid, showing that neural models are able to encode texts in a way that preserve syntactic, semantic, and discourse coherence. While only a first step toward generating coherent text units from neural models, our work has the potential to significantly impact natural language generation and summarization1.

ROVER is a widely used method to combine the output of multiple automatic speech recognition (ASR) systems. Though effective, the basic approach and its variants suffer from potential drawbacks: i) their results depend on the order in which the hypotheses are used to feed the combination process, ii) when applied to combine long hypotheses, they disregard possible differences in transcription quality at local level, iii) they often rely on word confidence information. We address these issues by proposing a segment-based ROVER in which hypothesis ranking is obtained from a confidence-independent ASR quality estimation method. Our results on English data from the IWSLT2012 and IWSLT2013 evaluation campaigns significantly outperform standard ROVER and approximate two strong oracles.

In this paper, we address the problem of evaluating spontaneous speech using a combination of machine learning and crowdsourcing. Machine learning techniques inadequately solve the stated problem because automatic speakerindependent speech transcription is inaccurate. The features derived from it are also inaccurate and so is the machine learning model developed for speech evaluation. To address this, we post the task of speech transcription to a large community of online workers (crowd). We also get spoken English grades from the crowd. We achieve 95% transcription accuracy by combining transcriptions from multiple crowd workers. Speech and prosody features are derived by force aligning the speech samples on these highly accurate transcriptions. Additionally, we derive surface and semantic level features directly from the transcription. To demonstrate the efficacy of our approach we performed experiments on an expert–graded speech sample of 319 adult non–native speakers. Using these features in a regression model, we are able achieve a Pearson correlation of 0.76 with expert grades, an accuracy much higher than any previously reported machine learning approach. Our approach has an accuracy that rivals that of expert agreement. This work is timely given the huge requirement of spoken English training and assessment.

We investigate a technique to adapt unsupervised word embeddings to specific applications, when only small and noisy labeled datasets are available. Current methods use pre-trained embeddings to initialize model parameters, and then use the labeled data to tailor them for the intended task. However, this approach is prone to overfitting when the training is performed with scarce and noisy data. To overcome this issue, we use the supervised data to find an embedding subspace that fits the task complexity. All the word representations are adapted through a projection into this task-specific subspace, even if they do not occur on the labeled dataset. This approach was recently used in the SemEval 2015 Twitter sentiment analysis challenge, attaining state-of-the-art results. Here we show results improving those of the challenge, as well as additional experiments in a Twitter Part-Of-Speech tagging task.

This paper tackles the issue of the detection of user’s verbal expressions of likes and dislikes in a human-agent interaction. We present a system grounded on the theoretical framework provided by (Martin and White, 2005) that integrates the interaction context by jointly processing agent’s and user’s utterances. It is designed as a rule-based and bottom-up process based on a symbolic representation of the structure of the sentence. This article also describes the annotation campaign – carried out through Amazon Mechanical Turk – for the creation of the evaluation dataset. Finally, we present all measures for rating agreement between our system and the human reference and obtain agreement scores that are equal or higher than substantial agreements.

A review text is normally represented as a bag-of-words (BOW) in sentiment classification. Such a simplified BOW model has fundamental deficiencies in modeling some complex linguistic phenomena such as negation. In this work, we propose a dual-view co-training algorithm based on dual-view BOW representation for semisupervised sentiment classification. In dual-view BOW, we automatically construct antonymous reviews and model a review text by a pair of bags-of-words with opposite views. We make use of the original and antonymous views in pairs, in the training, bootstrapping and testing process, all based on a joint observation of two views. The experimental results demonstrate the advantages of our approach, in meeting the two co-training requirements, addressing the negation problem, and enhancing the semi-supervised sentiment classification efficiency.

Predicting emotion categories, such as anger, joy, and anxiety, expressed by a sentence is challenging due to its inherent multi-label classification difficulty and data sparseness. In this paper, we address above two challenges by incorporating the label dependence among the emotion labels and the context dependence among the contextual instances into a factor graph model. Specifically, we recast sentence-level emotion classification as a factor graph inferring problem in which the label and context dependence are modeled as various factor functions. Empirical evaluation demonstrates the great potential and effectiveness of our proposed approach to sentencelevel emotion classification. 1

Automatically detecting verbal irony (roughly, sarcasm) in online content is important for many practical applications (e.g., sentiment detection), but it is difficult. Previous approaches have relied predominantly on signal gleaned from word counts and grammatical cues. But such approaches fail to exploit the context in which comments are embedded. We thus propose a novel strategy for verbal irony classification that exploits contextual features, specifically by combining noun phrases and sentiment extracted from comments with the forum type (e.g., conservative or liberal) to which they were posted. We show that this approach improves verbal irony classification performance. Furthermore, because this method generates a very large feature space (and we expect predictive contextual features to be strong but few), we propose a mixed regularization strategy that places a sparsity-inducing `1 penalty on the contextual feature weights on top of the `2 penalty applied to all model coefficients. This increases model sparsity and reduces the variance of model performance.

Central to many sentiment analysis tasks are sentiment lexicons (SLs). SLs exhibit polarity inconsistencies. Previous work studied the problem of checking the consistency of an SL for the case when the entries have categorical labels (positive, negative or neutral) and showed that it is NPhard. In this paper, we address the more general problem, in which polarity tags take the form of a continuous distribution in the interval [0, 1]. We show that this problem is polynomial. We develop a general framework for addressing the consistency problem using linear programming (LP) theory. LP tools allow us to uncover inconsistencies efficiently, paving the way to building SL debugging tools. We show that previous work corresponds to 0-1 integer programming, a particular case of LP. Our experimental studies show a strong correlation between polarity consistency in SLs and the accuracy of sentiment tagging in practice.