Semantic Similarity Applied to Spoken Dialogue Summarization

Semantic Similarity Applied to Spoken Dialogue Summarization Iryna Gurevych and Michael Strube EML Research gGmbH Schloss-Wolfsbrunnenweg 33 69118 Heidelberg, Germany http://www.eml-research.de/english/homes/{gurevych|strube} Abstract We present a novel approach to spoken dialogue summarization. Our system employs a set of semantic similarity metrics using the noun por- tion of WordNet as a knowledge source. So far, the noun senses have been disambiguated man- ually. The algorithm aims to extract utterances carrying the essential content of dialogues. We evaluate the system on 20 Switchboard dia- logues. The results show that our system out- performs LEAD, RANDOM and TF*IDF base- lines. 1 Introduction Research in automatic text summarization began in the late 1950s and has been receiving more atten- tion again over the last decade. The maturity of this research area is indicated by recent large-scale eval- uation efforts (Radev et al., 2003). In comparison, speech summarization is a rather new research area which emerged only a few years ago. However, the demand for speech summarization is growing be- cause of the increasing availability of (digitally en- coded) speech databases (e.g. spoken news, politi- cal speeches). Our research is concerned with the development of a system for automatically generating summaries of conversational speech. As a potential applica- tion we envision the automatic generation of meet- ing minutes. The approach to spoken dialogue summarization presented herein unifies corpus- and knowledge-based approaches to summarization, i.e. we develop a shallow knowledge-based approach. Our system employs a set of semantic similar- ity metrics which utilize WordNet as a knowledge source. We claim that semantic similarity between a given utterance and the dialogue as a whole is an appropriate criterion for the selection of utterances which carry the essential content of the dialogue, i.e. relevant utterances. – In order to study the perfor- mance of semantic similarity methods, we remove the noise from the pre-processing modules by man- ually disambiguating lexical noun senses. In Section 2, we briefly describe research on sum- marization and how spoken dialogue summarization differs from text summarization. Section 3 gives the semantic similarity metrics we use and describes how they are applied to the summarization problem. Section 4 provides information about the data used in our experiments, while Section 5 describes the experiments and the results together with their sta- tistical significance. 2 Text, Speech and Dialogue Summarization Most research on automatic summarization dealt with written text. This work was based either on corpus-based, statistical methods or on knowledge- based techniques (for an overview over both strands of research see Mani & Maybury (1999)). Re- cent advances in text summarization are mostly due to statistical techniques with some additional us- age of linguistic knowledge, e.g. (Marcu, 2000; Teufel & Moens, 2002), which can be applied to un- restricted input. Research on speech summarization focused mainly on single-speaker, written-to-be-spoken text (e.g. spoken news, political speeches, etc.). The methods were mostly derived from work on text summarization, but extended it by exploiting partic- ular characteristics of spoken language, e.g. acous- tic confidence scores or intonation. Difficulties arise …

Content truncated to 500 words