Extractive Text Summarization of Long Documents Using Word and Sentence Encoding

Automatic Text Summarization (ATS) for long documents is a very challenging task. A long document includes more than one topic, so it is required to construct a summary that covers the most important contextual information from the different topics in the input document. There are many ATS techniques, but the produced summaries are much less accurate than the human summaries. This paper proposes a supervised extractive Neural-based summarization model that focuses on long documents summarization by integrating both the local and global contexts of the document. The proposed system employs two bidirectional Gated Recurrent Units (GRUs) as word and sentence encoders, respectively, and uses Word2Vec embeddings as word representation. Two ensembles are proposed by applying the average ensemble method to the proposed system with two other Neural-based systems. The proposed system uses the PubMed dataset to evaluate the two ensembles using the ROUGE evaluation metrics. The results of the evaluation show that there are promising and remarkable improvements of 0.14% and 0.97% for ROUGE-1, 0.33% and 1.12% for ROUGE-2, and 0.25% for ROUGE-L.