Recuperação de informação com auxílio de extratos automáticos

Abstract

This dissertation investigates the use of Automatic Summarization (AS) techniques on Information Retrieval (IR), two areas that have attracted growing attention due to the continuous growth of information repositories in digital format. The main goal of this work is to verify the contribution of extracts generated automatically in two stages of the IR: indexing and Pseudo Relevance Feedback (PRF). In general, the main goal of both phases is to find the most descriptive terms of a given document. This goal in turn is strongly related to that of the AS techniques to summarize the main parts of a document which justifies the study. In order to verify the effectiveness of the AS techniques, we have developed five systems that use extracts generated by a summarizing system that was considered useful in the task of indicating the content of documents to human readers. These systems were assessed through a set of documents written in Portuguese to test IR. In general, the results show that the generated extracts were not useful for indexing, presenting worse performance compared to when using a full document for IR. In the PRF case, however, the results obtained using specific extracts were better than those obtained by a system that does not embed PRF. The best results were obtained when using query-biased multi-documents extracts, indicating that this type of extract may be useful for PRF.