Subjective QA scoring algorithm based on two-BERTs model

• 1、Institute of Computer Science，Beijing University of Posts and Telecommunications , Beijing

Abstract：With the development of natural language processing, computers have been able to construct good question answering algorithms, but they cannot evaluate questions and answers well from a subjective perspective. Evaluating questions and answers from different subjective perspectives can enhance automatic understanding of the content of complex questions and answers, and further promote the development of question answering systems. The data set in this paper comes from the "Google QUEST Q&A Labeling" competition held by Google in 2019, and a method based on two-BERTs model is proposed to score the subjective labels related to the question and answer separately. On this basis, four different text truncation methods are compared to process long question and answer texts; in view of each layer of BERT captures the different features of the input text, a multi-layer feature fusion method is proposed in this paper to effectively combine different layer features. Based on the pre-training model of BERT, this paper adopts a fine-tuning training strategy with different layer learning rate, so that the model can be better fitted. Experimental results show that the algorithm proposed in this paper performs well on subjective QA scoring tasks.

Keywords： Computer application technology BERT Feature fusion Subjective QA scoring