基于TR-BERT和注意力网络的多文档推理方法

In the task of multi-document inference, information related to the answer may reside across multiple relevant texts, and sometimes this information is not directly associated with the question. To address the challenge of balancing accuracy and efficiency in multi-document inference for the electric utility customer service scenario, this paper proposes a multi-level inference method based on pre-trained models and attention networks. The model utilizes pre-trained models to preserve the rich semantics extracted from paragraph texts and questions, and then evaluates the relevance of candidates through an attention mechanism. Furthermore, due to the real-time requirements of the question-answering scenario, we employ the TR-BERT pre-trained model based on dynamic token reduction and simplify the attention network. Experimental results on the WikiHop dataset demonstrate that the model overall exhibits advantages in both computational speed and accuracy, providing effective methodological support for the multi-turn question-answering functionality in intelligent question-answering systems.