A hierarchical self-attentive neural extractive summarizer via reinforcement learning (HSASRL)

摘要

In recent years, deep neural extractive-based summarization approaches have achieved vast popularity over conventional approaches. However, previously proposed neural extractive-based models have issues that limit their performance. One of these issues is related to the architecture of the used neural network that skips some details about the document hierarchical structure. Moreover, these models are optimized to maximize the probabilities of the training data ground truth labels rather than the evaluation metric that actually measures the quality of the summarization; this way of optimization might neglect important information related to sentence ranking. To address these issues, we combined reinforcement and supervised learning to train a hierarchical self-attentive reinforced neural network-based summarization model to rank sentences according to their significance by directly optimizing the ROUGE evaluation metric. The proposed model employs a hierarchical self-attention mechanism to generate document and sentence embeddings that reflect the hierarchical structure of the document and give better feature representation. While reinforcement learning enables direct optimization with respect to evaluation metrics, the attention mechanism adds an extra source of information to direct the summary extraction. The model was evaluated on the basis of three well-known datasets, namely, CNN, Daily Mail, and their combined version CNN/Daily Mail. Experimental results showed that the model achieved higher ROUGE scores than state-of-the-art models for extractive summarization on the three datasets.