NAACL 2015 Papers

This study tackles the problem of paraphrase acquisition: achieving high coverage as well as accuracy. Our method first induces paraphrase patterns from given seed paraphrases, exploiting the generality of paraphrases exhibited by pairs of lexical variants, e.g., “amendment” and “amending,” in a fully empirical way. It then searches monolingual corpora for new paraphrases that match the patterns. This can extract paraphrases comprising words that are completely different from those of the given seeds. In experiments, our method expanded seed sets by factors of 42 to 206, gaining 84% to 208% more coverage than a previous method that generalizes only identical word forms. Human evaluation through a paraphrase substitution test demonstrated that the newly acquired paraphrases retained reasonable quality, given substantially high-quality seeds.

A key challenge of designing coherent semantic ontology for spoken language understanding is to consider inter-slot relations. In practice, however, it is difficult for domain experts and professional annotators to define a coherent slot set, while considering various lexical, syntactic, and semantic dependencies. In this paper, we exploit the typed syntactic dependency theory for unsupervised induction and filling of semantics slots in spoken dialogue systems. More specifically, we build two knowledge graphs: a slot-based semantic graph, and a word-based lexical graph. To jointly consider word-to-word, word-toslot, and slot-to-slot relations, we use a random walk inference algorithm to combine the two knowledge graphs, guided by dependency grammars. The experiments show that considering inter-slot relations is crucial for generating a more coherent and compete slot set, resulting in a better spoken language understanding model, while enhancing the interpretability of semantic slots.

We present Model Invertibility Regularization (MIR), a method that jointly trains two directional sequence alignment models, one in each direction, and takes into account the invertibility of the alignment task. By coupling the two models through their parameters (as opposed to through their inferences, as in Liang et al.’s Alignment by Agreement (ABA), and Ganchev et al.’s Posterior Regularization (PostCAT)), our method seamlessly extends to all IBMstyle word alignment models as well as to alignment without parallel data. Our proposed algorithm is mathematically sound and inherits convergence guarantees from EM. We evaluate MIR on two tasks: (1) On word alignment, applying MIR on fertility based models we attain higher F-scores than ABA and PostCAT. (2) On Japanese-to-English backtransliteration without parallel data, applied to the decipherment model of Ravi and Knight, MIR learns sparser models that close the gap in whole-name error rate by 33% relative to a model trained on parallel data, and further, beats a previous approach by Mylonakis et al.

Linguistic borrowing is the phenomenon of transferring linguistic constructions (lexical, phonological, morphological, and syntactic) from a “donor” language to a “recipient” language as a result of contacts between communities speaking different languages. Borrowed words are found in all languages, and—in contrast to cognate relationships—borrowing relationships may exist across unrelated languages (for example, about 40% of Swahili’s vocabulary is borrowed from Arabic). In this paper, we develop a model of morpho-phonological transformations across languages with features based on universal constraints from Optimality Theory (OT). Compared to several standard— but linguistically naive—baselines, our OTinspired model obtains good performance with only a few dozen training examples, making this a cost-effective strategy for sharing lexical information across languages.

Research on ranked retrieval of spoken content has assumed the existence of some automated (word or phonetic) transcription. Recently, however, methods have been demonstrated for matching spoken terms to spoken content without the need for language-tuned transcription. This paper describes the first application of such techniques to ranked retrieval, evaluated using a newly created test collection. Both the queries and the collection to be searched are based on Gujarati produced naturally by native speakers; relevance assessment was performed by other native speakers of Gujarati. Ranked retrieval is based on fast acoustic matching that identifies a deeply nested set of matching speech regions, coupled with ways of combining evidence from those matching regions. Results indicate that the resulting ranked lists may be useful for some practical similarity-based ranking tasks.

We present a novel evaluation method for grammatical error correction that addresses problems with previous approaches and scores systems in terms of improvement on the original text. Our method evaluates corrections at the token level using a globally optimal alignment between the source, a system hypothesis, and a reference. Unlike the M 2 Scorer, our method provides scores for both detection and correction and is sensitive to different types of edit operations.

The semantic representation of individual word senses and concepts is of fundamental importance to several applications in Natural Language Processing. To date, concept modeling techniques have in the main based their representation either on lexicographic resources, such as WordNet, or on encyclopedic resources, such as Wikipedia. We propose a vector representation technique that combines the complementary knowledge of both these types of resource. Thanks to its use of explicit semantics combined with a novel cluster-based dimensionality reduction and an effective weighting scheme, our representation attains state-of-the-art performance on multiple datasets in two standard benchmarks: word similarity and sense clustering. We are releasing our vector representations at http://lcl.uniroma1.it/nasari/.

Multiview LSA (MVLSA) is a generalization of Latent Semantic Analysis (LSA) that supports the fusion of arbitrary views of data and relies on Generalized Canonical Correlation Analysis (GCCA). We present an algorithm for fast approximate computation of GCCA, which when coupled with methods for handling missing values, is general enough to approximate some recent algorithms for inducing vector representations of words. Experiments across a comprehensive collection of test-sets show our approach to be competitive with the state of the art.

Analyzing public opinions towards products, services and social events is an important but challenging task. An accurate sentiment analyzer should take both lexicon-level information and corpus-level information into account. It also needs to exploit the domainspecific knowledge and utilize the common knowledge shared across domains. In addition, we want the algorithm being able to deal with missing labels and learning from incomplete sentiment lexicons. This paper presents a LCCT (Lexicon-based and Corpus-based, Co-Training) model for semi-supervised sentiment classification. The proposed method combines the idea of lexicon-based learning and corpus-based learning in a unified cotraining framework. It is capable of incorporating both domain-specific and domainindependent knowledge. Extensive experiments show that it achieves very competitive classification accuracy, even with a small portion of labeled data. Comparing to state-ofthe-art sentiment classification methods, the LCCT approach exhibits significantly better performances on a variety of datasets in both English and Chinese.

In spite of the apparent irregularity of the English spelling system, Chomsky and Halle (1968) characterize it as “near optimal”. We investigate this assertion using computational techniques and resources. We design an algorithm to generate word spellings that maximize both phonemic transparency and morphological consistency. Experimental results demonstrate that the constructed system is much closer to optimality than the traditional English orthography.

We present a comparative investigation of word representations for part-of-speech (POS) and morphological tagging, focusing on scenarios with considerable differences between training and test data where a robust approach is necessary. Instead of adapting the model towards a specific domain we aim to build a robust model across domains. We developed a test suite for robust tagging consisting of six languages and different domains. We find that representations similar to Brown clusters perform best for POS tagging and that word representations based on linguistic morphological analyzers perform best for morphological tagging.

Most of previous work in knowledge base (KB) completion has focused on the problem of relation extraction. In this work, we focus on the task of inferring missing entity type instances in a KB, a fundamental task for KB competition yet receives little attention. Due to the novelty of this task, we construct a large-scale dataset and design an automatic evaluation methodology. Our knowledge base completion method uses information within the existing KB and external information from Wikipedia. We show that individual methods trained with a global objective that considers unobserved cells from both the entity and the type side gives consistently higher quality predictions compared to baseline methods. We also perform manual evaluation on a small subset of the data to verify the effectiveness of our knowledge base completion methods and the correctness of our proposed automatic evaluation method.

We present a new approach to harvesting a large-scale, high quality image-caption corpus that makes a better use of already existing web data with no additional human efforts. The key idea is to focus on Deja Image-Captions: naturally existing image descriptions that are repeated almost verbatim – by more than one individual for different images. The resulting corpus provides association structure between 4 million images with 180K unique captions, capturing a rich spectrum of everyday narratives including figurative and pragmatic language. Exploring the use of the new corpus, we also present new conceptual tasks of visually situated paraphrasing, creative image captioning, and creative visual paraphrasing.

We take a novel approach to zero pronoun resolution in Chinese: our model explicitly tracks the flow of focus in a discourse. Our approach, which generalizes to deictic references, is not reliant on the presence of overt noun phrase antecedents to resolve to, and allows us to address the large percentage of “non-anaphoric” pronouns filtered out in other approaches. We furthermore train our model using readily available parallel Chinese/English corpora, allowing for training without hand-annotated data. Our results demonstrate improvements on two test sets, as well as the usefulness of linguistically motivated features.

Gradable terms such as brief, lengthy and extended illustrate varying degrees of a scale and can therefore participate in comparative constructs. Knowing the set of words that can be compared on the same scale and the associated ordering between them (brief < lengthy < extended) is very useful for a variety of lexical semantic tasks. Current techniques to derive such an ordering rely on WordNet to determine which words belong on the same scale and are limited to adjectives. Here we describe an extension to recent work: we investigate a fully automated pipeline to extract gradable terms from a corpus, group them into clusters reflecting the same scale and establish an ordering among them. This methodology reduces the amount of required handcrafted knowledge, and can infer gradability of words independent of their part of speech. Our approach infers an ordering for adjectives with comparable performance to previous work, but also for adverbs with an accuracy of 71%. We find that the technique is useful for inferring such rankings among words across different domains, and present an example using biomedical text.

Context representations are a key element in distributional models of word meaning. In contrast to typical representations based on neighboring words, a recently proposed approach suggests to represent a context of a target word by a substitute vector, comprising the potential fillers for the target word slot in that context. In this work we first propose a variant of substitute vectors, which we find particularly suitable for measuring context similarity. Then, we propose a novel model for representing word meaning in context based on this context representation. Our model outperforms state-of-the-art results on lexical substitution tasks in an unsupervised setting.

Tree trimming is the problem of extracting an optimal subtree from an input tree, and sentence extraction and sentence compression methods can be formulated and solved as tree trimming problems. Previous approaches require integer linear programming (ILP) solvers to obtain exact solutions. The problem of this approach is that ILP solvers are black-boxes and have no theoretical guarantee as to their computation complexity. We propose a dynamic programming (DP) algorithm for tree trimming problems whose running time is O(NLlogN), where N is the number of tree nodes andL is the length limit. Our algorithm exploits the zero-suppressed binary decision diagram (ZDD), a data structure that represents a family of sets as a directed acyclic graph, to represent the set of subtrees in a compact form; the structure of ZDD permits the application of DP to obtain exact solutions, and our algorithm is applicable to different tree trimming problems. Moreover, experiments show that our algorithm is faster than state-of-the-art ILP solvers, and that it scales well to handle large summarization problems.

This paper presents a framework to infer spatial knowledge from verbal semantic role representations. First, we generate potential spatial knowledge deterministically. Second, we determine whether it can be inferred and a degree of certainty. Inferences capture that something is located or is not located somewhere, and temporally anchor this information. An annotation effort shows that inferences are ubiquitous and intuitive to humans.

We present a simple and yet effective approach that can incorporate rationales elicited from annotators into the training of any offthe-shelf classifier. We show that our simple approach is effective for multinomial na¨ Bayes, logistic regression, and support vector machines. We additionally present an active learning method tailored specifically for the learning with rationales framework.