Matching Verb Attributes for Cross-Document Event Coreference.pdfVIP

Matching Verb Attributes for Cross-Document Event Coreference Eleftheria Tomadaki and Andrew Salway Department of Computing, University of Surrey Guildford, Surrey, UK GU2 7XH {e.tomadaki, a.salway}@surrey.ac.uk Abstract Collateral texts of different genre can describe the same filmed story, e.g. audio description and plot summaries. We deal with the challenge of cross-document coreference for events by matching verb attributes. Cross document coreference is the task of deciding whether two linguistic descriptions from different sources refer to the same event. This is important for reliable information integration, as well as generating richer machine-executable representations of multimedia material in retrieval and browsing systems. Corpora of audio description and plot summaries were analysed to investigate how they describe the same film events. This analysis shows that events are described by different verbs in the two corpora and has inspired the algorithms for cross-document event coreference, which match verb attributes, rather than verbs themselves. The preliminary evaluation was encouraging, showing a significantly better performance than the baseline algorithm. 1 Introduction The present era can be characterised by a vast amount of information available in different forms of media; text documents, images, audio and video files etc. Many kinds of electronic information artefacts convey the same story; a fire event, for example, can be broadcast on television or radio, or narrated in a newspaper by the people that were affected; or a fictional story, for example Cinderella can be presented in films, theatre, books, pantomime etc. Information can be conveyed in the form of stories in history, science, current affairs, financial news, fiction etc. The process of narrating a story comprises a sequence of causally connected events organised in space and time. Matching events can be one way to acquire major infor