Towards Topic Labeling with Phrase Entailment and Aggregation

Proceedings of NAACL-HLT 2013, pages 179–189, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Towards Topic Labeling with Phrase Entailment and Aggregation Yashar Mehdad Giuseppe Carenini Raymond T. Ng Shafiq Joty Department of Computer Science, University of British Columbia Vancouver, BC, V6T 1Z4, Canada {mehdad, carenini, rng, rjoty}@cs.ubc.ca Abstract We propose a novel framework for topic la- beling that assigns the most representative phrases for a given set of sentences cover- ing the same topic. We build an entailment graph over phrases that are extracted from the sentences, and use the entailment relations to identify and select the most relevant phrases. We then aggregate those selected phrases by means of phrase generalization and merging. We motivate our approach by applying over conversational data, and show that our frame- work improves performance significantly over baseline algorithms. 1 Introduction Given text segments about the same topic written in different ways (i.e., language variability), topic la- beling deals with the problem of automatically gen- erating semantically meaningful labels for those text segments. The potential of integrating topic label- ing as a prerequisite for higher-level analysis has been reported in several areas, such as summariza- tion (Harabagiu and Lacatusu, 2010; Kleinbauer et al., 2007; Dias et al., 2007), information extraction (Allan, 2002) and conversation visualization (Liu et al., 2012). Moreover, the huge amount of tex- tual data generated everyday specifically in conver- sations (e.g., emails and blogs) calls for automated methods to analyze and re-organize them into mean- ingful coherent clusters. Table 1 shows an example of two human written topic labels for a topic cluster collected from a blog1, 1http://slashdot.org Text: a: Where do you think the term “Horse laugh” comes from? b: And that rats also giggled when tickled. c: My hypothesis- if an animal can play, it can “laugh” or at least it is familiar with the concept of “laughing”. Many animals play. There are various sorts of humour though. Some involve you laughing because your brain suddenly made a lots of unexpected connections. Possible extracted phrases: animals play, rats have, laugh, Horse laugh, rats also giggle, rats Human-authored topic labels: animals which laugh, animal laughter Table 1: Topic labeling example. and possible phrases that can be extracted from the topic cluster using different approaches. This ex- ample demonstrates that although most approaches (Mei et al., 2007; Lau et al., 2011; Branavan et al., 2007) advocate extracting phrase-level topic labels from the text segments, topically related text seg- ments do not always contain one keyword or key phrase that can capture the meaning of the topic. As shown in this example, such labels do not exist in the original text and cannot be extracted using the exist- ing probabilistic models (e.g., (Mei et al., 2007)). The same problem can be observed with many other examples. This suggests the idea of aggregating and generating topic labels, instead of simply extracting them, as a challenging scenario for this field of re- search. Moreover, to generate a label for a topic we …

Content truncated to 500 words