DEIM: An effective deep encoding and interaction model for sentence matching
Kexin Jiang, Yahui Zhao, De Li, Zhenguo Zhang · 2022
Sentence matching is to compare the relevance of two paragraphs of sentences. The current mainstream approach is a deep learning-based approach, usually through the attention mechanism to interact with sentence pairs, which can be a good match for the relationship between two sentences. However, these methods cannot handle sentences with high semantic complexity. Therefore, we propose a method that is based on a deep encoder to extract deep semantic information. To avoid the problem of an insufficient number of interactions to lack information, we encode sentences at the coding layer using the method of multiple encoded interactions. After we get the representation information of the sentence, in this paper, the encoded sentences are processed for secondary interaction of information, while using self-attention to obtain more complex feature interaction information. We conducted experimental validation on some text-matching datasets, and the experimental results show that our proposed DEIM model plays an irreplaceable and important role in the task of text matching.