Hierarchical and Multiple-Perspective Interaction Network for Long Text Matching

Long text matching is widely used in various sub-tasks of natural language processing. However, conducting research in this field can be challenging due to excessive redundant and distracting information, the complex semantic context, and the limited availability of high quality public datasets. Existing long text matching methods generally do not fully use the rich local features embedded in text information, and focus more on encoding long text as fixed length vectors to calculate the semantic distance, disregarding the importance of feature interaction in the text matching process. Therefore, the performance of the relevant models needs to be improved. To address these problems, a hierarchical and multiple-perspective interaction network (HMIN) is proposed in this paper. First, the long text is encoded at the word and sentence levels to extract global features, while one-dimensional convolutional neural networks and attention mechanisms are used to focus on important local features in long texts. Second, the different types of features are compared separately using the comparison function, and then, the comparison results are aggregated. Finally, whether long texts are matched is determined in the prediction layer. We have conducted comparative experiments on two datasets, the results show that HMIN has an improvement in accuracy and F1 values compared with the same type of existing algorithms, and the related experimental analysis demonstrates the effectiveness of the proposed method.