Identifying Topics by Position

Identifying Topics by Position Chin-Yew Lin and Eduard Hovy Information Sciences Institute of the University of Southern California 4676 Admiralty Way Marina del Rey, CA 90292, USA {cyl ,hovy}@isi. edu Abstract This paper addresses the problem of iden- tifying likely topics of texts by their posi- tion in the text. It describes the automated training and evaluation of an Optimal Posi- tion Policy, a method of locating the likely positions of topic-bearing sentences based on genre-specific regularities of discourse structure. This method can be used in applications such as information retrieval, routing, and text summarization. 1 Introduction: Topic Spotting by Position In an increasingly information-laden world, the problem of automatically finding the major topics of texts acquires new urgency. A module that can suggest likely locations of topics in texts, robustly and with an acceptable degree of correctness, would be useful for a number of important applications, including information retrieval, gisting, and auto- mated summarization. Several methods have been tried to perform Topic Identification. Some involve parsing and semantic analysis of the text, and are therefore less robust over arbitrary input. Others, such as the Cue Phrase and Position methods, are more robust, though gen- erally somewhat less accurate. Of these, the Position Method, identified in the late 1950's, remains among the best; it can outperform newer methods such as those based on word counting (Salton et al., 1994). The Position Method springs from the recognition that texts in a genre generally observe a predictable discourse structure, and that sentences of greater topic centrality tend to occur in certain specifiable locations. The text's title, for example, is a very informative position in most genres, as is the Ab- stract paragraph in scientific articles. Edmundson (Edmundson, 1969) defined the Position Method as follows: "...the machine-readable cues are cer- tain general characteristics of the corpus provided by the skeletons of documents, i.e. headings and format. The Location method is based on the hypothesis that: (1) sentences occurring under certain headings are positively relevant; and (2) topic sen- tences tend to occur very early or very late in a document and its paragraphs." However, since the paradigmatic discourse structure differs significantly over text genres and subject do- mains, the Position Method cannot be defined as straightforwardly as Baxendale's (Baxendale, 1958) title plus first and last sentences of each paragraph; it has to be tailored to genre and domain. Can one develop ways of tailoring this method? Furthermore, since the resolution power of the Position Method is the sentence, while the de- sired output--topics--generally appear at the word or phrase level, the most accurate results of this method may still include too much spurious mate- rial to be really useful. How useful is the method in general? By what measure(s) can one evaluate it? Basic questions about how the Position Method can be tailored for optimality over a genre and how it can be evaluated for effectiveness remain unan- swered. To our knowledge, no systematic study has yet been performed, though some variant of …

Content truncated to 500 words