QUERY-BASED AMHARIC LEGAL DOCUMENT SUMMARIZATION

Abstract

Automatic text summarization is a natural language processing application that proposes the extraction of salient information from a source document to produce a condensed version for a particular user. Due to the increasing number of interested users in searching for Amharic documents, Automatic text summarization becomes a tool to help users through pointing which document is relevant to their need (search) by putting a summary for each particular document in a particular domain. However, most researches have been conducted for Amharic text summarization and almost all are done for generic summarization, but in this study query-oriented summarization have been done to extract query related sentences from the document based on their rank. In this study, two distinct approaches have been used to extract Amharic legal judgment summaries from a single legal document based on the incoming user query. Both TF*IDF weighting and TextRank approaches generate a summary based on the query with different performances results. The performance evaluation has been done objectively and subjectively by comparing manual summary with system generated summary. The objective evaluation result shows that Amharic document TextRank registers the best result of 71.36% F-measure at extraction rate 30%. The subjective evaluation also conducted by the experts to compare the coherence and informativeness of the summarizer. In terms of informativeness and coherency of the summary, 85.87% accuracy is registered at extraction rate of 30%. The experiment shows the study achieves promising result. However, TF*IDF weighting and TextRank approaches employed in this study are incapable of extracting a summary from documents which are too long this needs further investigation. One of the challenges of this study is when the document length increases the noise happen and important part of a summary may be missed.