Tibetan Question Intention Classification Based on Feature Perception of Question Words

Question intent classification is an important part of the Tibetan Q&A system. It is a challenging task to use deep learning methods to classify the intent of Tibetan questions. This paper aims to: (1) The traditional Tibetan question intent classification scheme is too coarse, which is not conducive to the problem intent recognition task in the question answering system. (2) The Tibetan question intent type is highly dependent on the local spatial characteristics with interrogative virtual words as the core, which leads to the problem of poor classification effect, and a novel Tibetan question intent classification scheme is proposed, and each question intent type is defined. On this basis, through a detailed analysis of the semantic features and writing rules of Tibetan interrogators, a semantic information labeling algorithm for Tibetan interrogators was designed, and an BERT_RCNN intentional classification model for Tibetan questions was trained by combining the advantages of Tibetan Pre-Training (TiBERT) and recurrent convolutional neural network. The F1 value of this method can reach 94.87% on the test set, which can accurately identify user needs, compress the search space and optimize model decision-making, and significantly improve the performance of the Tibetan Q&A system.