ACL 2015 论文列表

We repurpose network security hardware to perform language identification and language modeling tasks. The hardware is a deterministic pushdown transducer since it executes regular expressions and has a stack. One core is 2.4 times as fast at language identification and 1.8 to 6 times as fast at part-of-speech language modeling.

Methods for name matching, an important component to support downstream tasks such as entity linking and entity clustering, have focused on alphabetic languages, primarily English. In contrast, logogram languages such as Chinese remain untested. We evaluate methods for name matching in Chinese, including both string matching and learning approaches. Our approach, based on new representations for Chinese, improves both name matching and a downstream entity clustering task.

The task of event trigger labeling is typically addressed in the standard supervised setting: triggers for each target event type are annotated as training data, based on annotation guidelines. We propose an alternative approach, which takes the example trigger terms mentioned in the guidelines as seeds, and then applies an eventindependent similarity-based classifier for trigger labeling. This way we can skip manual annotation for new event types, while requiring only minimal annotated training data for few example events at system setup. Our method is evaluated on the ACE-2005 dataset, achieving 5.7% F1 improvement over a state-of-the-art supervised system which uses the full training data.

We study the event detection problem using convolutional neural networks (CNNs) that overcome the two fundamental limitations of the traditional feature-based approaches to this task: complicated feature engineering for rich feature sets and error propagation from the preceding stages which generate these features. The experimental results show that the CNNs outperform the best reported feature-based systems in the general setting as well as the domain adaptation setting without resorting to extensive external resources.

This paper investigates the contribution of document level processing of timeanchors for TimeLine event extraction. We developed and tested two different systems. The first one is a baseline system that captures explicit time-anchors. The second one extends the baseline system by also capturing implicit time relations. We have evaluated both approaches in the SemEval 2015 task 4 TimeLine: CrossDocument Event Ordering. We empirically demonstrate that the document-based approach obtains a much more complete time anchoring. Moreover, this approach almost doubles the performance of the systems that participated in the task.

Short texts usually encounter data sparsity and ambiguity problems in representations for their lack of context. In this paper, we propose a novel method to model short texts based on semantic clustering and convolutional neural network. Particularly, we first discover semantic cliques in embedding spaces by a fast clustering algorithm. Then, multi-scale semantic units are detected under the supervision of semantic cliques, which introduce useful external knowledge for short texts. These meaningful semantic units are combined and fed into convolutional layer, followed by max-pooling operation. Experimental results on two open benchmarks validate the effectiveness of the proposed method.

Existing studies have utilized Wikipedia for various knowledge acquisition tasks. However, no attempts have been made to explore multi-level topic knowledge contained in Wikipedia articles’

Many queries in web search are ambiguous or multifaceted. Identifying the major senses or facets of queries is very important for web search. In this paper, we represent the major senses or facets of queries as subtopics and refer to indentifying senses or facets of queries as query subtopic mining, where query subtopic are represented as a number of clusters of queries. Then the challenges of query subtopic mining are how to measure the similarity between queries and group them semantically. This paper proposes an approach for mining subtopics from query log, which jointly learns a similarity measure and groups queries by explicitly modeling the structure among them. Compared with previous approaches using manually defined similarity measures, our approach produces more desirable query subtopics by learning a similarity measure. Experimental results on real queries collected from a search engine log confirm the effectiveness of the proposed approach in mining query subtopics.

In this paper, we present a test collection for mathematical information retrieval composed of real-life, researchlevel mathematical information needs. Topics and relevance judgements have been procured from the on-line collaboration website MathOverflow by delegating domain-specific decisions to experts on-line. With our test collection, we construct a baseline using Lucene’s vectorspace model implementation and conduct an experiment to investigate how prior extraction of technical terms from mathematical text can affect retrieval efficiency. We show that by boosting the importance of technical terms, statistically significant improvements in retrieval performance can be obtained over the baseline.

Co-Simrank is a useful Simrank-like measure of similarity based on graph structure. The existing method iteratively computes each pair of Co-Simrank score from a dot product of two Pagerank vectors, entailing O(log(1/e)*n^3) time to compute all pairs of Co-Simranks in a graph with n nodes, to attain a desired accuracy e. In this study, we devise a model, Co-Simmate, to speed up the retrieval of all pairs of Co-Simranks to O(log2 (log(1/e))*n^3) time. Moreover, we show the optimality of Co-Simmate among other hop-(u^k) variations, and integrate it with a matrix decomposition based method on singular graphs to attain higher efficiency. The viable experiments verify the superiority of Co-Simmate to others.

We propose an unsupervised probabilistic model for zero pronoun resolution. To our knowledge, this is the first such model that (1) is trained on zero pronouns in an unsupervised manner; (2) jointly identifies and resolves anaphoric zero pronouns; and (3) exploits discourse information provided by a salience model. Experiments demonstrate that our unsupervised model significantly outperforms its state-of-the-art unsupervised counterpart when resolving the Chinese zero pronouns in the OntoNotes corpus.

We give an algorithm for disambiguating generic versus referential uses of secondperson pronouns in restaurant reviews in Chinese. Reviews in this domain use the ‘you’ pronoun 你 either generically or to refer to shopkeepers, readers, or for selfreference in reported conversation. We first show that linguistic features of the local context (drawn from prior literature) help in disambigation. We then show that document-level features (n-grams and document-level embeddings)— not previously used in the referentiality literature— actually give the largest gain in performance, and suggest this is because pronouns in this domain exhibit ‘one-senseper-discourse’. Our work highlights an important case of discourse effects on pronoun use, and may suggest practical implications for audience extraction and other sentiment tasks in online reviews.

Pronouns are frequently dropped in Chinese sentences, especially in informal data such as text messages. In this work we propose a solution to recover dropped pronouns in SMS data. We manually annotate dropped pronouns in 684 SMS files and apply machine learning algorithms to recover them, leveraging lexical, contextual and syntactic information as features. We believe this is the first work on recovering dropped pronouns in Chinese text messages.

Semantic applications typically extract information from intermediate structures derived from sentences, such as dependency parse or semantic role labeling. In this paper, we study Open Information Extraction’s (Open IE) output as an additional intermediate structure and find that for tasks such as text comprehension, word similarity and word analogy it can be very effective. Specifically, for word analogy, Open IE-based embeddings surpass the state of the art. We suggest that semantic applications will likely benefit from adding Open IE format to their set of potential sentencelevel structures.

We examine a key task in biomedical text processing, normalization of disorder mentions. We present a multi-pass sieve approach to this task, which has the advantage of simplicity and modularity. Our approach is evaluated on two datasets, one comprising clinical reports and the other comprising biomedical abstracts, achieving state-of-the-art results.

We propose a new approach to the task of fine grained entity type classifications based on label embeddings that allows for information sharing among related labels. Specifically, we learn an embedding for each label and each feature such that labels which frequently co-occur are close in the embedded space. We show that it outperforms state-of-the-art methods on two fine grained entity-classification benchmarks and that the model can exploit the finer-grained labels to improve classification of standard coarse types.

Previous research on relation classification has verified the effectiveness of using dependency shortest paths or subtrees. In this paper, we further explore how to make full use of the combination of these dependency information. We first propose a new structure, termed augmented dependency path (ADP), which is composed of the shortest dependency path between two entities and the subtrees attached to the shortest path. To exploit the semantic representation behind the ADP structure, we develop dependency-based neural networks (DepNN): a recursive neural network designed to model the subtrees, and a convolutional neural network to capture the most important features on the shortest path. Experiments on the SemEval-2010 dataset show that our proposed method achieves state-of-art results.

In contrast with traditional relation extraction, which only considers a fixed set of relations, Open Information Extraction (Open IE) aims at extracting all types of relations from text. Because of data sparseness, Open IE systems typically ignore lexical information, and instead employ parse trees and Part-of-Speech (POS) tags. However, the same syntactic structure may correspond to different relations. In this paper, we propose to use a lexicalized tree kernel based on the word embeddings created by a neural network model. We show that the lexicalized tree kernel model surpasses the unlexicalized model. Experiments on three datasets indicate that our Open IE system performs better on the task of relation extraction than the stateof-the-art Open IE systems of Xu et al. (2013) and Mesquita et al. (2013).

Distant supervision is a widely applied approach to automatic training of relation extraction systems and has the advantage that it can generate large amounts of labelled data with minimal effort. However, this data may contain errors and consequently systems trained using distant supervision tend not to perform as well as those based on manually labelled data. This work proposes a novel method for detecting potential false negative training examples using a knowledge inference method. Results show that our approach improves the performance of relation extraction systems trained using distantly supervised data.