Modeling discourse structure and temporal event relations for automated document summarization with markov logic networks
James C. Lester, Eun Young Ha · 2010
Recent years have seen significant progress in natural language processing. A key challenge posed by many natural language applications ranging from text summarization to question answering and machine translation is designing an effective discourse processing framework. Linguistic phenomena that hold across sentence boundaries are still not well understood, even though they provide crucial knowledge for integrating the information communicated by individual sentences. Our work addresses automatic processing of natural language spanning multiple sentences. More specifically, it concerns modeling discourse structure and temporal event relations in support of automated document summarization. The models for discourse structure and temporal event relations are learned from natural language corpora in a supervised manner by utilizing the Markov logic machine learning framework. Markov logic is a probabilistic extension of first-order logic that allows formulae to be violated. In contrast to first-order logic in which a possible world has a binary value (true or false), Markov logic assigns a weight to each formula, reflecting the strength of the constraint the formula represents. A Markov logic network is a set of weighted first-order clauses that represent objects in the domain, together with constants. An implemented prototype of a document summarizer utilizes a discourse structure model and a temporal event relation model in order to identify salient content from a given document and to structure the identified salient content into a coherent summary. The discourse structure model is learned from the RST-DT corpus, which consists of a collection of newspaper articles that are manually annotated with rhetorical structure based on Rhetorical Structure Theory. The temporal event relation model is learned from the TempEval-2 corpus, which contains a set of newpaper articles that are manually annotated with events and time expressions following the TimeML specifications for events and temporal relations.