ACL 2015 论文列表

People create morphs, a special type of fake alternative names, to achieve certain communication goals such as expressing strong sentiment or evading censors. For example, “Black Mamba”, the name for a highly venomous snake, is a morph that Kobe Bryant created for himself due to his agility and aggressiveness in playing basketball games. This paper presents the first end-to-end context-aware entity morph decoding system that can automatically identify, disambiguate, verify morph mentions based on specific contexts, and resolve them to target entities. Our approach is based on an absolute “cold-start” it does not require any candidate morph or target entity lists as input, nor any manually constructed morph-target pairs for training. We design a semi-supervised collective inference framework for morph mention extraction, and compare various deep learning based approaches for morph resolution. Our approach achieved significant improvement over the state-of-the-art method (Huang et al., 2013), which used a large amount of training data. 1

An event chronicle provides people with an easy and fast access to learn the past. In this paper, we propose the first novel approach to automatically generate a topically relevant event chronicle during a certain period given a reference chronicle during another period. Our approach consists of two core components – a timeaware hierarchical Bayesian model for event detection, and a learning-to-rank model to select the salient events to construct the final chronicle. Experimental results demonstrate our approach is promising to tackle this new problem.

We study the problem of explaining relationships between pairs of knowledge graph entities with human-readable descriptions. Our method extracts and enriches sentences that refer to an entity pair from a corpus and ranks the sentences according to how well they describe the relationship between the entities. We model this task as a learning to rank problem for sentences and employ a rich set of features. When evaluated on a large set of manually annotated sentences, we find that our method significantly improves over state-of-the-art baseline models.

We study the problem of summarizing DAG-structured topic hierarchies over a given set of documents. Example applications include automatically generating Wikipedia disambiguation pages for a set of articles, and generating candidate multi-labels for preparing machine learning datasets (e.g., for text classification, functional genomics, and image classification). Unlike previous work, which focuses on clustering the set of documents using the topic hierarchy as features, we directly pose the problem as a submodular optimization problem on a topic hierarchy using the documents as features. Desirable properties of the chosen topics include document coverage, specificity, topic diversity, and topic homogeneity, each of which, we show, is naturally modeled by a submodular function. Other information, provided say by unsupervised approaches such as LDA and its variants, can also be utilized by defining a submodular function that expresses coherence between the chosen topics and this information. We use a large-margin framework to learn convex mixtures over the set of submodular components. We empirically evaluate our method on the problem of automatically generating Wikipedia disambiguation pages using human generated clusterings as ground truth. We find that our framework improves upon several baselines according to a variety of standard evaluation metrics including the Jaccard Index, F1 score and NMI, and moreover, can be scaled to extremely large scale problems.

While recent years have seen a surge of interest in automated essay grading, including work on grading essays with respect to particular dimensions such as prompt adherence, coherence, and technical quality, there has been relatively little work on grading the essay dimension of argument strength, which is arguably the most important aspect of argumentative essays. We introduce a new corpus of argumentative student essays annotated with argument strength scores and propose a supervised, feature-rich approach to automatically scoring the essays along this dimension. Our approach significantly outperforms a baseline that relies solely on heuristically applied sentence argument function labels by up to 16.1%.

We characterize a class of indirect answers to yes/no questions, alternative answers, where information is given that is not directly asked about, but which might nonetheless address the underlying motivation for the question. We develop a model rooted in game theory that generates these answers via strategic reasoning about possible unobserved domain-level user requirements. We implement the model within an interactive question answering system simulating real estate dialogue. The system learns a prior probability distribution over possible user requirements by analyzing training dialogues, which it uses to make strategic decisions about answer selection. The system generates pragmatically natural and interpretable answers which make for more efficient interactions compared to a baseline.

Document enrichment focuses on retrieving relevant knowledge from external resources, which is essential because text is generally replete with gaps. Since conventional work primarily relies on special resources, we instead use triples of Subject, Predicate, Object as knowledge and incorporate distributional semantics to rank them. Our model first extracts these triples automatically from raw text and converts them into real-valued vectors based on the word semantics captured by Latent Dirichlet Allocation. We then represent these triples, together with the source document that is to be enriched, as a graph of triples, and adopt a global iterative algorithm to propagate relevance weight from source document to these triples so as to select the most relevant ones. Evaluated as a ranking problem, our model significantly outperforms multiple strong baselines. Moreover, we conduct a task-based evaluation by incorporating these triples as additional features into document classification and enhances the performance by 3.02%.

We propose that entity queries are generated via a two-step process: users first select entity facts that can distinguish target entities from the others; and then choose words to describe each selected fact. Based on this query generation paradigm, we propose a new entity representation model named as entity factoid hierarchy. An entity factoid hierarchy is a tree structure composed of factoid nodes. A factoid node describes one or more facts about the entity in different information granularities. The entity factoid hierarchy is constructed via a factor graph model, and the inference on the factor graph is achieved by a modified variant of Multiple-try Metropolis algorithm. Entity retrieval is performed by decomposing entity queries and computing the query likelihood on the entity factoid hierarchy. Using an array of benchmark datasets, we demonstrate that our proposed framework significantly improves the retrieval performance over existing models.

Non-linear models recently receive a lot of attention as people are starting to discover the power of statistical and embedding features. However, tree-based models are seldom studied in the context of structured learning despite their recent success on various classification and ranking tasks. In this paper, we propose S-MART, a tree-based structured learning framework based on multiple additive regression trees. S-MART is especially suitable for handling tasks with dense features, and can be used to learn many different structures under various loss functions. We apply S-MART to the task of tweet entity linking — a core component of tweet information extraction, which aims to identify and link name mentions to entities in a knowledge base. A novel inference algorithm is proposed to handle the special structure of the task. The experimental results show that S-MART significantly outperforms state-of-the-art tweet entity linking systems.

Automatic speech recognition (ASR) outputs often contain various disfluencies. It is necessary to remove these disfluencies before processing downstream tasks. In this paper, an efficient disfluency detection approach based on right-to-left transitionbased parsing is proposed, which can efficiently identify disfluencies and keep ASR outputs grammatical. Our method exploits a global view to capture long-range dependencies for disfluency detection by integrating a rich set of syntactic and disfluency features with linear complexity. The experimental results show that our method outperforms state-of-the-art work and achieves a 85.1% f-score on the commonly used English Switchboard test set. We also apply our method to in-house annotated Chinese data and achieve a significantly higher f-score compared to the baseline of CRF-based approach.

Spoken dialogue systems (SDS) typically require a predefined semantic ontology to train a spoken language understanding (SLU) module. In addition to the annotation cost, a key challenge for designing such an ontology is to define a coherent slot set while considering their complex relations. This paper introduces a novel matrix factorization (MF) approach to learn latent feature vectors for utterances and semantic elements without the need of corpus annotations. Specifically, our model learns the semantic slots for a domain-specific SDS in an unsupervised fashion, and carries out semantic parsing using latent MF techniques. To further consider the global semantic structure, such as inter-word and inter-slot relations, we augment the latent MF-based model with a knowledge graph propagation model based on a slot-based semantic graph and a word-based lexical graph. Our experiments show that the proposed MF approaches produce better SLU models that are able to predict semantic slots and word patterns taking into account their relations and domain-specificity in a joint manner.

In natural language understanding (NLU), a user utterance can be labeled differently depending on the domain or application (e.g., weather vs. calendar). Standard domain adaptation techniques are not directly applicable to take advantage of the existing annotations because they assume that the label set is invariant. We propose a solution based on label embeddings induced from canonical correlation analysis (CCA) that reduces the problem to a standard domain adaptation task and allows use of a number of transfer learning techniques. We also introduce a new transfer learning technique based on pretraining of hidden-unit CRFs (HUCRFs). We perform extensive experiments on slot tagging on eight personal digital assistant domains and demonstrate that the proposed methods are superior to strong baselines.

We propose an event-driven model for headline generation. Given an input document, the system identifies a key event chain by extracting a set of structural events that describe them. Then a novel multi-sentence compression algorithm is used to fuse the extracted events, generating a headline for the document. Our model can be viewed as a novel combination of extractive and abstractive headline generation, combining the advantages of both methods using event structures. Standard evaluation shows that our model achieves the best performance compared with previous state-of-the-art systems.

We present a novel syntax-based natural language generation system that is trainable from unaligned pairs of input meaning representations and output sentences. It is divided into sentence planning, which incrementally builds deep-syntactic dependency trees, and surface realization. Sentence planner is based on A* search with a perceptron ranker that uses novel differing subtree updates and a simple future promise estimation; surface realization uses a rule-based pipeline from the Treex NLP toolkit. Our first results show that training from unaligned data is feasible, the outputs of our generator are mostly fluent and relevant.

We present a new factoid-annotated dataset for evaluating content models for scientific survey article generation containing 3,425 sentences from 7 topics in natural language processing. We also introduce a novel HITS-based content model for automated survey article generation called HITSUM that exploits the lexical network structure between sentences from citing and cited papers. Using the factoid-annotated data, we conduct a pyramid evaluation and compare HITSUM with two previous state-of-the-art content models: C-Lexrank, a network based content model, and TOPICSUM, a Bayesian content model. Our experiments show that our new content model captures useful survey-worthy information and outperforms C-Lexrank by 4% and TOPICSUM by 7% in pyramid evaluation.

The sentiment classification performance relies on high-quality sentiment resources. However, these resources are imbalanced in different languages. Cross-language sentiment classification (CLSC) can leverage the rich resources in one language (source language) for sentiment classification in a resource-scarce language (target language). Bilingual embeddings could eliminate the semantic gap between two languages for CLSC, but ignore the sentiment information of text. This paper proposes an approach to learning bilingual sentiment word embeddings (BSWE) for English-Chinese CLSC. The proposed B-SWE incorporate sentiment information of text into bilingual embeddings. Furthermore, we can learn high-quality BSWE by simply employing labeled corpora and their translations, without relying on large-scale parallel corpora. Experiments on NLP&CC 2013 CLSC dataset show that our approach outperforms the state-of-the-art systems.

Cross-lingual sentiment analysis is a task of identifying sentiment polarities of texts in a low-resource language by using sentiment knowledge in a resource-abundant language. While most existing approaches are driven by transfer learning, their performance does not reach to a promising level due to the transferred errors. In this paper, we propose to integrate into knowledge transfer a knowledge validation model, which aims to prevent the negative influence from the wrong knowledge by distinguishing highly credible knowledge. Experiment results demonstrate the necessity and effectiveness of the model.

We propose a cross-lingual framework for fine-grained opinion mining using bitext projection. The only requirements are a running system in a source language and word-aligned parallel data. Our method projects opinion frames from the source to the target language, and then trains a system on the target language using the automatic annotations. Key to our approach is a novel dependency-based model for opinion mining, which we show, as a byproduct, to be on par with the current state of the art for English, while avoiding the need for integer programming or reranking. In cross-lingual mode (English to Portuguese), our approach compares favorably to a supervised system (with scarce labeled data), and to a delexicalized model trained using universal tags and bilingual word embeddings.

Semantic role labeling (SRL) is crucial to natural language understanding as it identifies the predicate-argument structure in text with semantic labels. Unfortunately, resources required to construct SRL models are expensive to obtain and simply do not exist for most languages. In this paper, we present a two-stage method to enable the construction of SRL models for resourcepoor languages by exploiting monolingual SRL and multilingual parallel data. Experimental results show that our method outperforms existing methods. We use our method to generate Proposition Banks with high to reasonable quality for 7 languages in three language families and release these resources to the research community.