Clustering Paraphrases by Word Sense

Proceedings of NAACL-HLT 2016, pages 1463–1472, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Clustering Paraphrases by Word Sense Anne Cocos and Chris Callison-Burch Computer and Information Science Department, University of Pennsylvania Abstract Automatically generated databases of English paraphrases have the drawback that they re- turn a single list of paraphrases for an input word or phrase. This means that all senses of polysemous words are grouped together, un- like WordNet which partitions different senses into separate synsets. We present a new method for clustering paraphrases by word sense, and apply it to the Paraphrase Database (PPDB). We investigate the performance of hi- erarchical and spectral clustering algorithms, and systematically explore different ways of defining the similarity matrix that they use as input. Our method produces sense clusters that are qualitatively and quantitatively good, and that represent a substantial improvement to the PPDB resource. 1 Introduction Many natural language processing tasks rely on the ability to identify words and phrases with equiva- lent meaning but different wording. These alterna- tive ways of expressing the same information are called paraphrases. Several research efforts have produced automatically generated databases of En- glish paraphrases, including DIRT (Lin and Pantel, 2001), the Microsoft Research Paraphrase Phrase Tables (Dolan et al., 2004), and the Paraphrase Database (Ganitkevitch et al., 2013; Pavlick et al., 2015a). A primary benefit of these automatically generated resources is their enormous scale, which provides superior coverage compared to manually compiled resources like WordNet (Miller, 1995). But automatically generated paraphrase resources currently have the drawback that they group all senses of polysemous words together, and do not partition paraphrases into groups like WordNet does bug (n) insect beetle cockroach mosquito pest c1 glitch error malfunction fault mistake failure c2 microbe virus parasite bacterium c3 tracker microphone wire informer snitch c4 Figure 1: Our goal is to partition paraphrases of an input word like bug into clusters representing its dis- tinct senses. with its synsets. Thus a search for paraphrases of the noun bug would yield a single list of para- phrases that includes insect, glitch, beetle, error, mi- crobe, wire, cockroach, malfunction, microphone, mosquito, virus, tracker, pest, informer, snitch, para- site, bacterium, fault, mistake, failure and many oth- ers. The goal of this work is to group these para- phrases into clusters that denote the distinct senses of the input word or phrase, as shown in Figure 1. We develop a method for clustering the para- phrases from the Paraphrase Database (PPDB). PPDB contains over 100 million paraphrases gen- erated using the bilingual pivoting method (Ban- nard and Callison-Burch, 2005), which posits that two English words are potential paraphrases of each other if they share one or more foreign translations. We apply two clustering algorithms, Hierarchical Graph Factorization Clustering (Yu et al., 2005; Sun and Korhonen, 2011) and Self-Tuning Spec- tral Clustering (Ng et al., 2001; Zelnik-Manor and Perona, 2004), and systematically explore different ways of defining the similarity matrix that they use as input. We exploit a variety …

Content truncated to 500 words