Multi-granularity adaptive extractive document summarization with heterogeneous graph neural networks

• Published in: PeerJ. Computer science 2023-12, Vol.9, p.e1737-e1737, Article e1737
• Main Authors: Su, Wu, Jiang, Jin, Huang, Kaihui
• Format: Article
• Language: English
• Subjects: Ablation; Adaptive; Algorithms; Analysis; Artificial Intelligence; Computational linguistics; Data Mining and Machine Learning; Documents; Extractive summarization; Graph attention network; Graph neural networks; Graphs; Language processing; Natural Language and Speech; Natural language interfaces; Neural Networks; Nodes; Semantics; Sentences; Text Mining
• ISSN: 2376-5992; 2376-5992
• DOI: 10.7717/PEERJ-CS.1737

The crucial aspect of extractive document summarization lies in understanding the interrelations between sentences. Documents inherently comprise a multitude of sentences, and sentence-level models frequently fail to consider the relationships between distantly-placed sentences, resulting in the omission of significant information in the summary. Moreover, information within documents tends to be distributed sparsely, challenging the efficacy of sentence-level models. In the realm of heterogeneous graph neural networks, it has been observed that semantic nodes with varying levels of granularity encapsulate distinct semantic connections. Initially, the incorporation of edge features into the computation of dynamic graph attention networks is performed to account for node relationships. Subsequently, given the multiplicity of topics in a document or a set of documents, a topic model is employed to extract topic-specific features and the probability distribution linking these topics with sentence nodes. Last but not least, the model defines nodes with different levels of granularity-ranging from documents and topics to sentences-and these various nodes necessitate different propagation widths and depths for capturing intricate relationships in the information being disseminated. Adaptive measures are taken to learn the importance and correlation between nodes of different granularities in terms of both width and depth. Experimental evidence from two benchmark datasets highlights the superior performance of the proposed model, as assessed by ROUGE metrics, in comparison to existing approaches, even in the absence of pre-trained language models. Additionally, an ablation study confirms the positive impact of each individual module on the model's ROUGE scores.