Multi-document summarization for Turkish news

Abstract

In this paper, we introduce our multi-document summarization system for Turkish news. The aim of the summarization system is to build a single document for multi document news that have been collected previously. The news were collected from several Turkish news sources via Real Simple Syndication (RSS). They were separated into clusters according to their topics. We utilized cosine similarity metric for the clustering process. Latent Semantic Analysis (LSA) has been used in the summarization phase. Multi-Document Summarization (MDS) differs from single document summarization in that the issues of compression, speed, redundancy and passage selection are essential inside the formation of ideal summaries. In this study, we utilized term frequency in document scoring which let us select the sentences with higher importance degree. We use ROUGE technique for evaluation of the system and our results show that the average of recall and precision percentage of this system is 43%. In the manual summarization phase, fifteen volunteers took part. The reason of low percentage is interpreted as getting texts randomly without any edit. It has been observed that the number of sentences and rate of summarization affect the accuracy rate.