ACL 2015 Papers

This paper proposes a novel lifelong learning (LL) approach to sentiment classification. LL mimics the human continuous learning process, i.e., retaining the knowledge learned from past tasks and use it to help future learning. In this paper, we first discuss LL in general and then LL for sentiment classification in particular. The proposed LL approach adopts a Bayesian optimization framework based on stochastic gradient descent. Our experimental results show that the proposed method outperforms baseline methods significantly, which demonstrates that lifelong learning is a promising research direction.

A question may be asked not only to elicit information, but also to make a statement. Questions serving the latter purpose, called rhetorical questions, are often lexically and syntactically indistinguishable from other types of questions. Still, it is desirable to be able to identify rhetorical questions, as it is relevant for many NLP tasks, including information extraction and text summarization. In this paper, we explore the largely understudied problem of rhetorical question identification. Specifically, we present a simple n-gram based language model to classify rhetorical questions in the Switchboard Dialogue Act Corpus. We find that a special treatment of rhetorical questions which incorporates contextual information achieves the highest performance.

In this paper, we present a model for improved discriminative semantic parsing. The model addresses an important limitation associated with our previous stateof-the-art discriminative semantic parsing model – the relaxed hybrid tree model by introducing our constrained semantic forests. We show that our model is able to yield new state-of-the-art results on standard datasets even with simpler features. Our system is available for download from http://statnlp.org/research/sp/.

Several compositional distributional semantic methods use tensors to model multi-way interactions between vectors. Unfortunately, the size of the tensors can make their use impractical in large-scale implementations. In this paper, we investigate whether we can match the performance of full tensors with low-rank approximations that use a fraction of the original number of parameters. We investigate the effect of low-rank tensors on the transitive verb construction where the verb is a third-order tensor. The results show that, while the low-rank tensors require about two orders of magnitude fewer parameters per verb, they achieve performance comparable to, and occasionally surpassing, the unconstrained-rank tensors on sentence similarity and verb disam-

In recent years, distributional models (DMs) have shown great success in representing lexical semantics. In this work we show that the extent to which DMs represent semantic knowledge is highly dependent on the type of knowledge. We pose the task of predicting properties of concrete nouns in a supervised setting, and compare between learning taxonomic properties (e.g., animacy) and attributive properties (e.g., size, color). We employ four state-of-the-art DMs as sources of feature representation for this task, and show that they all yield poor results when tested on attributive properties, achieving no more than an average F-score of 0.37 in the binary property prediction task, compared to 0.73 on taxonomic properties. Our results suggest that the distributional hypothesis may not be equally applicable to all types of semantic information.

We propose a simple yet effective approach to learning bilingual word embeddings (BWEs) from non-parallel document-aligned data (based on the omnipresent skip-gram model), and its application to bilingual lexicon induction (BLI). We demonstrate the utility of the induced BWEs in the BLI task by reporting on benchmarking BLI datasets for three language pairs: (1) We show that our BWE-based BLI models significantly outperform the MuPTM-based and context-counting models in this setting, and obtain the best reported BLI results for all three tested language pairs; (2) We also show that our BWE-based BLI models outperform other BLI models based on recently proposed BWEs that require parallel data for bilingual training.

In this paper, the answer selection problem in community question answering (CQA) is regarded as an answer sequence labeling task, and a novel approach is proposed based on the recurrent architecture for this problem. Our approach applies convolution neural networks (CNNs) to learning the joint representation of question-answer pair firstly, and then uses the joint representation as input of the long short-term memory (LSTM) to learn the answer sequence of a question for labeling the matching quality of each answer. Experiments conducted on the SemEval 2015 CQA dataset shows the effectiveness of our approach.

In this paper, we present an approach that address the answer sentence selection problem for question answering. The proposed method uses a stacked bidirectional Long-Short Term Memory (BLSTM) network to sequentially read words from question and answer sentences, and then outputs their relevance scores. Unlike prior work, this approach does not require any syntactic parsing or external knowledge resources such as WordNet which may not be available in some domains or languages. The full system is based on a combination of the stacked BLSTM relevance model and keywords matching. The results of our experiments on a public benchmark dataset from TREC show that our system outperforms previous work which requires syntactic features and external knowledge resources.

We demonstrate significant improvement on the MCTest question answering task (Richardson et al., 2013) by augmenting baseline features with features based on syntax, frame semantics, coreference, and word embeddings, and combining them in a max-margin learning framework. We achieve the best results we are aware of on this dataset, outperforming concurrentlypublished results. These results demonstrate a significant performance gradient for the use of linguistic structure in machine comprehension.

Retrieving similar questions in online QA (2) BOW-CNN is more robust than the pure CNN for long texts.

Community Question Answering (cQA) is a new application of QA in social contexts (e.g., fora). It presents new interesting challenges and research directions, e.g., exploiting the dependencies between the different comments of a thread to select the best answer for a given question. In this paper, we explored two ways of modeling such dependencies: (i) by designing specific features looking globally at the thread; and (ii) by applying structure prediction models. We trained and evaluated our models on data from SemEval-2015 Task 3 on Answer Selection in cQA. Our experiments show that: (i) the thread-level features consistently improve the performance for a variety of machine learning models, yielding state-of-the-art results; and (ii) sequential dependencies between the answer labels captured by structured prediction models are not enough to improve the results, indicating that more information is needed in the joint model.

Given a discourse tree for a text as a candidate answer to a compound query, we propose a rule system for valid and invalid occurrence of the query keywords in this tree. To be a valid answer to a query, its keywords need to occur in a chain of elementary discourse unit of this answer so that these units are fully ordered and connected by nucleus – satellite relations. An answer might be invalid if the queries’ keywords occur in the answer's satellite discourse units only. We build the rhetoric map of an answer to prevent it from firing by queries whose keywords occur in non-adjacent areas of the Answer Map. We evaluate the improvement of search relevance by filtering out search results not satisfying the proposed rule system, demonstrating a 4% increase of accuracy with respect to the nearest neighbor learning approach which does not use the discourse tree structure.

This paper presents a universal morphological feature schema that represents the finest distinctions in meaning that are expressed by overt, affixal inflectional morphology across languages. This schema is used to universalize data extracted from Wiktionary via a robust multidimensional table parsing algorithm and feature mapping algorithms, yielding 883,965 instantiated paradigms in 352 languages. These data are shown to be effective for training morphological analyzers, yielding significant accuracy gains when applied to Durrett and DeNero’s (2013) paradigm learning framework.

We study the problem of predicting tense in Chinese conversations. The unique challenges include: (1) Chinese verbs do not have explicit lexical or grammatical forms to indicate tense; (2) Tense information is often implicitly hidden outside of the target sentence. To tackle these challenges, we first propose a set of novel sentence-level (local) features using rich linguistic resources and then propose a new hypothesis of “One tense per scene” to incorporate scene-level (global) evidence to enhance the performance. Experimental results demonstrate the power of this hybrid approach, which can serve as a new and promising benchmark.

Human labeled corpus is indispensable for the training of supervised word segmenters. However, it is time-consuming and laborintensive to label corpus manually. During the process of typing Chinese text by Pingyin, people usually need to type "space" or numeric keys to choose the words due to homophones, which can be viewed as a cue for segmentation. We argue that such a process can be used to build a labeled corpus in a more natural way. Thus, in this paper, we investigate Natural Typing Annotations (NTAs) that are potential word delimiters produced by users while typing Chinese. A detailed analysis on over three hundred user-produced texts containing NTAs reveals that highquality NTAs mostly agree with gold segmentation and, consequently, can be used for improving the performance of supervised word segmentation model in out-of-domain. Experiments show that a classification model combined with a voting mechanism can reliably identify the high-quality NTAs texts that are more readily available labeled corpus. Furthermore, the NTAs might be particularly useful to deal with out-of-vocabulary (OOV) words such as proper names and neo-logisms.

Compared with carefully edited prose, the language of social media is informal in the extreme. The application of NLP techniques in this context may require a better understanding of word usage within social media. In this paper, we compute a word embedding for a corpus of tweets, comparing it to a word embedding for Wikipedia. After learning a transformation of one vector space to the other, and adjusting similarity values according to term frequency, we identify words whose usage differs greatly between the two corpora. For any given word, the set of words closest to it in a particular embedding provides a characterization for that word’s usage within the corresponding corpora.

This paper presents an email importance corpus annotated through Amazon Mechanical Turk (AMT). Annotators annotate the email content type and email importance for three levels of hierarchy (senior manager, middle manager and employee). Each email is annotated by 5 turkers. Agreement study shows that the agreed AMT annotations are close to the expert annotations. The annotated dataset demonstrates difference in proportions of content type between different levels. An email importance prediction system is trained on the dataset and identifies the unimportant emails at minimum 0.55 precision with only text-based features.

This paper proposes an approach to capture the pragmatic context needed to infer irony in tweets. We aim to test the validity of two main hypotheses: (1) the presence of negations, as an internal propriety of an utterance, can help to detect the disparity between the literal and the intended meaning of an utterance, (2) a tweet containing an asserted fact of the form Not(P1) is ironic if and only if one can assess the absurdity of P1. Our first results are encouraging and show that deriving a pragmatic contextual model is feasible.

In this paper, we build a corpus of tweets from Twitter annotated with keywords using crowdsourcing methods. We identify key differences between this domain and the work performed on other domains, such as news, which makes existing approaches for automatic keyword extraction not generalize well on Twitter datasets. These datasets include the small amount of content in each tweet, the frequent usage of lexical variants and the high variance of the cardinality of keywords present in each tweet. We propose methods for addressing these issues, which leads to solid improvements on this dataset for this task.