A User Intent Recognition Model for Medical Queries Based on Attentional Interaction and Focal Loss Boost

Yuyu Luo, Yi Xie, Enliang Yan, Lap-Kei Lee, Fu Lee Wang, Tianyong Hao

Abstract: Pre-trained language models such as BERT and RoBERTa have obtained new state-of-the-art results in the user intent recognition task. Nevertheless, in the medical field, the models frequently neglect to make full use of label information and seldom take the difficulty of intent recognition for each query sentence into account. In this paper, a new user intent recognition model based on Text-Label Attention Interaction and Focal Loss Boost named TAI-FLB is proposed to identify user intents from medical query sentences. The model focuses on incorporating a text-to-label attention interaction mechanism based on label embedding to exploit the information from labels. Moreover, during training process, the loss contribution of difficult samples with unclear intention is increased to shift model focus towards difficult samples in medical query statements. Experimental evaluation was performed on two publicly available datasets KUAKE and CMID. The results demonstrated that the proposed TAI-FLB model outperformed other baseline methods and demonstrated its effectiveness.