DC-Graph: a chunk optimization model based on document classification and graph learning

Abstract

AbstractExisting machine reading comprehension methods use a fixed stride to chunk long texts, which leads to missing contextual information at the boundaries of the chunks and a lack of communication between the information within each chunk. This paper proposes DC-Graph model, addressing existing issues in terms of reconstructing and supplementing information in long texts. Knowledge graphs contain extensive knowledge, and the semantic relationships between entities exhibit strong logical characteristics, which can assist the model in semantic understanding and reasoning. By categorizing the questions, this paper filters the content of long texts based on categories and reconstructs the content that aligns with the question category, compressing and optimizing the long text to minimize the number of document chunks when inputted into BERT. Additionally, unstructured text is transformed into a structured knowledge graph, and features are extracted using graph convolutional networks. These features are then added as global information to each chunk, aiding answer prediction. Experimental results on the CoQA, QuAC, and TriviaQA datasets demonstrate that our method outperforms both BERT and Recurrent Chunking Mechanisms, which share the same improvement approach, in terms of F1 and EM score. The code is available at (https://github.com/guohaozhang/DC-Graph.git).