Graph based method for Arabic text summarization

The amount of Arabic textual data is growing tremendously, hence the need to reduce it with the aim to be easier to use while keeping only the necessary from the original text. In this regard, several natural language processing researchers are working on the generation of extractive and abstractive summary tools to achieve this aim. In this work, we explore an extractive approach to realize a generative model of summaries for Arabic single-documents. We focus on the use of graph-based methods to find the most important sentences and then extract them with a variety of text representation methods such as TF-IDF, fastText, and Word2Vec-, similarity measures, and graph ranking methods. To test our system we used the EASC (Essex Arabic Summaries Corpus) and the ROUGE metric to evaluate it. The results obtained show that the TF-IDF representation, the ranking by PageRank, and the use of cosine similarity achieve good performance, which can generate a high-quality summary.