ACL 2015 论文列表

Nearly all work in unsupervised grammar induction aims to induce unlabeled dependency trees from gold part-of-speechtagged text. These clean linguistic classes provide a very important, though unrealistic, inductive bias. Conversely, induced clusters are very noisy. We show here, for the first time, that very limited human supervision (three frequent words per cluster) may be required to induce labeled dependencies from automatically induced word clusters.

We propose a neural network model for scalable generative transition-based dependency parsing. A probability distribution over both sentences and transition sequences is parameterised by a feedforward neural network. The model surpasses the accuracy and speed of previous generative dependency parsers, reaching 91.1% UAS. Perplexity results show a strong improvement over n-gram language models, opening the way to the efficient integration of syntax into neural models for language generation.

We report improved AMR parsing results by adding a new action to a transitionbased AMR parser to infer abstract concepts and by incorporating richer features produced by auxiliary analyzers such as a semantic role labeler and a coreference resolver. We report final AMR parsing results that show an improvement of 7% absolute in F1 score over the best previously reported result. Our parser is available at: https://github.com/ Juicechuan/AMRParsing

We propose a method for semantic structure analysis of noun phrases using Abstract Meaning Representation (AMR). AMR is a graph representation for the meaning of a sentence, in which noun phrases (NPs) are manually annotated with internal structure and semantic relations. We extract NPs from the AMR corpus and construct a data set of NP semantic structures. We also propose a transition-based algorithm which jointly identifies both the nodes in a semantic structure tree and semantic relations between them. Compared to the baseline, our method improves the performance of NP semantic structure analysis by 2.7 points, while further incorporating external dictionary boosts the performance by 7.1 points.

Training a high-accuracy dependency parser requires a large treebank. However, these are costly and time-consuming to build. We propose a learning method that needs less data, based on the observation that there are underlying shared structures across languages. We exploit cues from a different source language in order to guide the learning process. Our model saves at least half of the annotation effort to reach the same accuracy compared with using the purely supervised method.

Coverage maximization with bigram concepts is a state-of-the-art approach to unsupervised extractive summarization. It has been argued that such concepts are adequate and, in contrast to more linguistic concepts such as named entities or syntactic dependencies, more robust, since they do not rely on automatic processing. In this paper, we show that while this seems to be the case for a commonly used newswire dataset, use of syntactic and semantic concepts leads to significant improvements in performance in other domains.

Timeline generation is a summarisation task which transforms a narrative, roughly chronological input text into a set of timestamped summary sentences, each expressing an atomic historical event. We present a methodology for evaluating systems which create such timelines, based on a novel corpus consisting of 36 humancreated timelines. Our evaluation relies on deep semantic units which we call historical content units. An advantage of our approach is that it does not require human annotation of new system summaries.

In this paper, we propose the concept of summary prior to define how much a sentence is appropriate to be selected into summary without consideration of its context. Different from previous work using manually compiled documentindependent features, we develop a novel summary system called PriorSum, which applies the enhanced convolutional neural networks to capture the summary prior features derived from length-variable phrases. Under a regression framework, the learned prior features are concatenated with document-dependent features for sentence ranking. Experiments on the DUC generic summarization benchmarks show that PriorSum can discover different aspects supporting the summary prior and outperform state-of-the-art baselines.

Comunicacio presentada a: the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing del 26 al 31 de juliol de 2015 a Beijing, Xina.

The intelligent personal assistant software such as the Apple’s Siri and Samsung’s S-Voice has been issued these days. This paper introduces a novel Spoken Language Understanding (SLU) module to predict user’s intention for determining system actions of the intelligent personal assistant software. The SLU module usually consists of several connected recognition tasks on a pipeline framework, whereas the proposed SLU module simultaneously recognizes four recognition tasks on a recognition framework using Conditional Random Fields (CRF). The four tasks include named entity, speech-act, target and operation recognition. In the experiments, the new simultaneous recognition method achieves the higher performance of 4% and faster speed of about 25% than other method using a pipeline framework. By a significance test, this improvement is considered to be statistically significant as a p-value of smaller than 0.05.

We investigate the impact of listener’s gaze on predicting reference resolution in situated interactions. We extend an existing model that predicts to which entity in the environment listeners will resolve a referring expression (RE). Our model makes use of features that capture which objects were looked at and for how long, reflecting listeners’ visual behavior. We improve a probabilistic model that considers a basic set of features for monitoring listeners’ movements in a virtual environment. Particularly, in complex referential scenes, where more objects next to the target are possible referents, gaze turns out to be beneficial and helps deciphering listeners’ intention. We evaluate performance at several prediction times before the listener performs an action, obtaining a highly significant accuracy gain.

In this paper, we introduce the task of selecting compact lexicon from large, noisy gazetteers. This scenario arises often in practice, in particular spoken language understanding (SLU). We propose a simple and effective solution based on matrix decomposition techniques: canonical correlation analysis (CCA) and rank-revealing QR (RRQR) factorization. CCA is first used to derive low-dimensional gazetteer embeddings from domain-specific search logs. Then RRQR is used to find a subset of these embeddings whose span approximates the entire lexicon space. Experiments on slot tagging show that our method yields a small set of lexicon entities with average relative error reduction of > 50% over randomly selected lexicon.

Dialogue Management (DM) is a key issue in Spoken Dialogue System (SDS). Most of the existing studies on DM use Dialogue Act (DA) to represent semantic information of sentence, which might not represent the nuanced meaning sometimes. In this paper, we model DM based on sentence clusters which have more powerful semantic representation ability than DAs. Firstly, sentences are clustered not only based on the internal information such as words and sentence structures, but also based on the external information such as context in dialogue via Recurrent Neural Networks. Additionally, the DM problem is modeled as a Partially Observable Markov Decision Processes (POMDP) with sentence clusters. Finally, experimental results illustrate that the proposed DM scheme is superior to the existing one.

Dialog state tracking is a key component of many modern dialog systems, most of which are designed with a single, well-defined domain in mind. This paper shows that dialog data drawn from different dialog domains can be used to train a general belief tracking model which can operate across all of these domains, exhibiting superior performance to each of the domain-specific models. We propose a training procedure which uses out-of-domain data to initialise belief tracking models for entirely new domains. This procedure leads to improvements in belief tracking performance regardless of the amount of in-domain data available for training the model.

Compared to the categorical approach that represents affective states as several discrete classes (e.g., positive and negative), the dimensional approach represents affective states as continuous numerical values on multiple dimensions, such as the valence-arousal (VA) space, thus allowing for more fine-grained sentiment analysis. In building dimensional sentiment applications, affective lexicons with valence-arousal ratings are useful resources but are still very rare. Therefore, this study proposes a weighted graph model that considers both the relations of multiple nodes and their similarities as weights to automatically determine the VA ratings of affective words. Experiments on both English and Chinese affective lexicons show that the proposed method yielded a smaller error rate on VA prediction than the linear regression, kernel method, and pagerank algorithm used in previous studies.

Most cross-lingual sentiment classification (CLSC) research so far has been performed at sentence or document level. Aspect-level CLSC, which is more appropriate for many applications, presents the additional difficulty that we consider subsentential opinionated units which have to be mapped across languages. In this paper, we extend the possible cross-lingual sentiment analysis settings to aspect-level specific use cases. We propose a method, based on constrained SMT, to transfer opinionated units across languages by preserving their boundaries. We show that cross-language sentiment classifiers built with this method achieve comparable results to monolingual ones, and we compare different cross-lingual settings.

In this paper, we propose a flexible principle-based approach (PBA) for reader-emotion classification and writing assistance. PBA is a highly automated process that learns emotion templates from raw texts to characterize an emotion and is comprehensible for humans. These templates are adopted to predict reader-emotion, and may further assist in emotional resonance writing. Results demonstrate that PBA can effectively detect reader-emotions by exploiting the syntactic structures and semantic associations in the context, thus outperforming wellknown statistical text classification methods and the state-of-the-art reader-emotion classification method. Moreover, writers are able to create more emotional resonance in articles under the assistance of the generated emotion templates. These templates have been proven to be highly interpretable, which is an attribute that is difficult to accomplish in traditional statistical methods.

Humans are idiosyncratic and variable: towards the same topic, they might hold different opinions or express the same opinion in various ways. It is hence important to model opinions at the level of individual users; however it is impractical to estimate independent sentiment classification models for each user with limited data. In this paper, we adopt a modelbased transfer learning solution – using linear transformations over the parameters of a generic model – for personalized opinion analysis. Extensive experimental results on a large collection of Amazon reviews confirm our method significantly outperformed a user-independent generic opinion model as well as several state-ofthe-art transfer learning algorithms.

Code-switching is commonly used in the free-form text environment, such as social media, and it is especially favored in emotion expressions. Emotions in codeswitching texts differ from monolingual texts in that they can be expressed in either monolingual or bilingual forms. In this paper, we first utilize two kinds of knowledge, i.e. bilingual and sentimental information to bridge the gap between different languages. Moreover, we use a term-document bipartite graph to incorporate both bilingual and sentimental information, and propose a label propagation based approach to learn and predict in the bipartite graph. Empirical studies demonstrate the effectiveness of our proposed approach in detecting emotion in code-switching texts.