Evaluation of Rewrite Rules for References to People in Automatic Multi-document Summaries

• Grant number: 17/15344-8
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: October 01, 2017
• End date: September 30, 2019
• Field of knowledge: Linguistics, Literature and Arts - Linguistics
• Principal Investigator: Ariani Di Felippo
• Grantee: Luana Fonseca Cristini
• Host Institution: Centro de Educação e Ciências Humanas (CECH). Universidade Federal de São Carlos (UFSCAR). São Carlos , SP, Brazil

Abstract

Multi-document Summarization (MDS) aims at automatically condensing a group of source texts on the same topic into a shorter version preserving the main information content of the collection. The extractive summarization methods consist of selecting important sentences from the original documents and concatenating them into shorter form. Consequently, the multi-document extracts often contain several problems of cohesion and coherence. Some of the problems are directly related to the named entities references, such as "first mention without explanation", "subsequent mention with explanation", "acronym without explanation", etc. A possible way to treat these problems is to rewrite the references in a post-editing MDS step. In this project, we aim at evaluating the effect of rewriting references to people on the linguistic quality of extractive summaries in Portuguese. We will evaluate the set of rewriting rules proposed by Di-Felippo (2016) - based on a characterization of the co-reference chains in human multi-document summaries - by comparing the original extractive and rewritten summaries. Our study represents an attempting to shift away from the purely extractive MDS to a partially abstractive one, by rewriting the references and consequently improving the linguistic quality of the extracts. Finally, we emphasize that this research proposal is inserted into Sustento, a project already supported by FAPESP (2012/13246-5) (2012-2014), and it is directly related to the postdoctoral research of the supervisor (FAPESP #2015/01450-5).