Subjective Prediction of Questions in Q & A System based on the Open Domain of Daily Life

Abstract

People and computers have different understandings of questions, and people have different needs for answers. For some questions, people may not need objective answers, but developmental opinions. This paper analyzes long and difficult questions in an open domain question answering system and provides effective information to the system with subjective predictions. It uses pseudo-label technology and the blending of multiple pre-trained language models to improve the understanding of long and difficult text question sentences. In addition, by designing a variety of subjective labels, the model's prediction of the subjectivity and objectivity of questions can provide effective information for the question-and-answer system. Since there are currently no standard definitions or standards for subjective labels and long and difficult text question sentences, we have conducted a subjective analysis of long text questions based on 30 question sentence subjective labels and long text question longer than 512 characters, using Spearman's relative coefficient as the evaluation standard for model prediction. This work is the first to implement subjective prediction of long and difficult text in the open domain area by designing 30 subjective labels.