Résumé de séminaire

Résumé :

In this talk I will present a methodology for the Summarization of Evolving Events. I propose that such a summarization task entails the capturing of the similarities and differences that exist between the news sources, in two axes: synchronically and diachronically. We manage to capture those similarities and differences by the introduction of the notion of Synchronic and Diachronic Relations (SDRs), which are cross-document relations. SDRs exist between some structures which we call messages.

This talk will begin with an introduction to the concepts of events and topics, and the ways in which events evolve, dividing them into linearly and non-linearly evolving events. We will continue by presenting a general methodology as well as a system architecture for the automatic creation of summaries from evolving events. Two concrete implementations for the above will finally be presented, through two case studies: one for a linearly and one for a non-linearly evolving event. In these case-studies I will present the Machine Learning techniques used for the extraction of the entities, messages and SDRs from the input documents. The extracted messages and SDRs form a graph structure which we call grid. This grid is later passed on to a Natural Language Generation system in order to create the final summary. This seminar will conclude with the presentation of the evaluation results and some words on future work.