Multi-Document Summarisation Using Generic Relation Extraction

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 420–429, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Multi-Document Summarisation Using Generic Relation Extraction Centre for Languate TecnologyBen Hachey Macquarie University NSW 2109 Australia Capital Markets CRC Limited GPO Box 970 Sydney NSW 2001 bhachey@cmcrc.com Abstract Experiments are reported that investi- gate the effect of various source docu- ment representations on the accuracy of the sentence extraction phase of a multi- document summarisation task. A novel representation is introduced based on generic relation extraction (GRE), which aims to build systems for relation iden- tification and characterisation that can be transferred across domains and tasks with- out modification of model parameters. Re- sults demonstrate performance that is sig- nificantly higher than a non-trivial base- line that uses tf*idf-weighted words and at least as good as a comparable but less gen- eral approach from the literature. Anal- ysis shows that the representations com- pared are complementary, suggesting that extraction performance could be further improved through system combination. 1 Introduction The goal of summarisation is to take an informa- tion source, extract content from it, and present the most important content in a condensed form (Mani, 2001). The field of automatic summarisa- tion (Mani, 2001; Sp¨arck Jones, 2007) aims to cre- ate tools that address various summarisation tasks with minimal human intervention. Extractive ap- proaches to automatic summarisation create rep- resentations of the source document that are gen- erally based on an easily identified text sub-unit such as sentences or paragraphs. These represen- tations are then used to identify representative or otherwise important snippets of text to place in the summary. Following Sp¨arck Jones (2007), summarisation systems can be characterised with respect to their approach to three main sub-tasks: 1) interpreta- tion, 2) transformation and 3) generation. The input consists of the source document (or a col- lection of source documents in the case of multi- document summarisation). The first step (interpre- tation) creates a representation of the source doc- ument by performing some level of interpretation. A simple approach here represents sentences by their tokens (i.e., as an unordered bag-of-words). The next step (transformation) is the compaction step where the source representation is converted into the summary representation, e.g. by identify- ing sentences whose words are most representative of the full text. Finally, in the last step (genera- tion), the output summary is created. In the case of sentence extraction, this includes various opera- tions to maximise coherence such as ensuring that entity references are comprehensible and arrang- ing the sentences in a sensible order. The current work investigates several represen- tations of source documents. In particular, an ap- proach from the literature based on atomic events (Filatova and Hatzivassiloglou, 2004) is compared to a novel approach based on generic relation ex- traction (GRE), which aims to build systems for relation identification and characterisation that can be transferred across domains and tasks without modification of model parameters (Hachey, 2009). The various representations are substituted in the interpretation phase …

Content truncated to 500 words