Co-referential chaining for coherent summaries through rhetorical and linguistic modeling

Co-referential chaining poses a difficult problem for Automatic Summarization (AS) when a text unit that embeds an anaphoric reference is chosen to compose a summary and its antecedent is not. Coherence, in this case, is often severely damaged and so may the degree of informativity be. This article presents an AS proposal that deals with rhetorical and linguistic knowledge to overcome that. A heuristics-based system is presented that addresses referentiality and rhetorical structuring to identify when an antecedent text unit must be included in a summary, to avoid coherence problems. The model has been assessed so far on both informativity and coherence for news articles written in Brazilian Portuguese. Only coherence is addressed in this paper.