Summarization of news articles

Abstract

In this work, two neural summarization models, Seq2seq and the Transformer, with several variations, were implemented and evaluated on the task of abstractively summarizing news articles. Seq2seq yielded poor results, likely due to not being flexible enough to fit the data set. The Transformer yielded promising results, and it was discovered that the quality of the output was heavily dependent on the quality of the input data, indicating that the implementation might be good but the performance bottle necked by the data set. For future work, specifically in developing a summarizer of clusters of documents, a recommended approach would be to combine an abstractive summarizer such as the Transformer, with extractive methods. In such a case, the Transformer could be further improved upon by pre-training it on word embeddings such as Google BERT, or training it on additional data sets such as CNN/Daily Mail. Finally, it was discovered that the used evaluation metric, ROUGE, could not be considered complete for the given task, and it would thus be advised to explore additional evaluation metrics for summarization models.