Date of Award
Master of Science (MS)
School of Information Technology: Information Systems
Extractive summarization has been widely studied, but the summaries generated by most current extractive summarization works usually disregard the article structure of the source document. Furthermore, the produced summaries are sometimes not representative sentences in the article. In this thesis, we propose an extractive summarization algorithm with knowledge graph enhancement that leverages both the source document and a knowledge graph to predict the most representative sentences for the summary. The aid of knowledge graphs enables deep learning models with pre-trained language models to focus on article structure information in the process of generating extractive summaries. Our proposed method has an encoder and a classifier: the encoder encodes the source document and the knowledge graph separately. The classifier inter-encodes the encoded source document and knowledge graph information by the cross-attention mechanism. Then the classifier determines whether the sentences belong to summary sentences or not. The results show that our model produces higher ROUGE scores on the CNN/Daily Mail dataset than the model without the knowledge graph for assistance, compared to the extractive summarization work based on the pre-trained language model.
Gao, Yichen, "A Deep Learning Approach to Extractive Text Summarization Using Knowledge Graph and Language Model" (2023). Theses and Dissertations. 1667.