Er-EIR: A Chinese Question Matching Model Based on Word-Level and Sentence-Level Interaction Features

The semantic matching of questions is a fundamental aspect of retrieval-based question answering (QA) systems. Text representations containing rich semantic information are required to achieve a deeper understanding of question intent. While existing large pre-trained models can obtain character-based text representations with contextual information, the specificity of Chinese sentences makes word-based text representation superior to character-based text representation. In this paper, we propose a question semantic matching method based on word-level and sentence-level interaction features. We utilize a Bidirectional Long Short-Term Memory (BiLSTM) approach to enhance the contextual information of the word representations. Additionally, we incorporate a co-attention mechanism to capture the interaction information between sentence pairs. By comparing our model with several baseline models on a self-built dataset of university financial question pairs, we have achieved remarkable performance.