Contextual Salience in Query-based Summarization

International Conference RANLP 2009 - Borovets, Bulgaria, pages 39–44 Contextual salience in query-based summarization Wauter Bosma CLTL, VU University Amsterdam Boelelaan 1105, 1081HV Amsterdam w.bosma@let.vu.nl Abstract Discourse theories claim that text gets meaning in context. Most summarization systems do not take advantage of this. They assess the rele- vance of each passage individually rather than modeling the way context affects the relevance of passages. This paper presents a framework for graph-based summarization in order to model relations in text, so that the passages can be viewed in a broader context. The result is a summarization system which is more in line with discourse theory but still fully automatic. evaluated the content selection performance ofI an implementation of the framework in differ- ent configurations. The system significantly out- performs a competitive baseline (and participant systems) on the DUC 2005 evaluation set. Keywords Query-based summarization, content selection, semantic net- works, discourse structure, graph theory. 1 Introduction One of the challenges in automatic summarization is content selection – deciding what should be in a sum- mary, and what shouldn’t. Summarization systems typically do this by determining the relevance of each passage independently, and then composing a sum- mary of the top passages. Classical features for scor- ing sentences include the presence of cue phrases, term frequency, stop word lists, etc. [4, 7]. Systems which assess the relevance of each sentence individually violate insights in discourse organization (e.g., [9]), which claim that meaning is tightly related to discourse organization. The meaning in a text is not merely the sum of the meaning in its passages, but a passage should be interpreted in the context shaped by other passages. For example, given the two passages in Fig. 1, the second passage had little meaning if the context provided by the first would be omitted. Hence, a generic summarization system should include the second sentence in a summary only if the first (or simi- lar) is also included. Recently, summarization systems have broadened their scope from generic single docu- ment summarization to multi-document summariza- tion, query-based summarization and update summa- rization [11]. These summarization tasks have made the need for dealing with issues like redundancy and coherence even more critical. For instance, in case of query-based summarization, the query is part of the summary’s context. Update summarization extends the context to prior knowledge, represented by a num- ber of documents which are assumed to be read by the user. A number of ad-hoc solutions to redundancy and co- herence emerged in response to the increasingly com- plex summarization tasks. For instance, [3] intro- duced the concept of marginal relevance: i.e., that the salience of a content unit is reduced by its redundancy with respect to the summary thus far. [1] divided the source into topics by identifying lexical chains. They composed summaries of one sentence from each of the strongest topics, as to maximize coverage. The sum- marization system of [2] prefers to include sentences in the summary which have a coherence relation to another summary …

Content truncated to 500 words