Research On Task Oriented Dialogue System For Travel Domain

Nowadays,task-oriented dialogue system is currently a research hotspot in academia and industry,and has received extensive attention from domestic and foreign scholars and companies in recent years.Task-oriented dialogue system can help users complete task-based instructions through multiple rounds of dialogues,such as querying bus routing,restaurant reservations,and recommendations for tourist attractions.The current mainstream research methods including dividing dialogue system into three modules of natural language understanding,dialogue management and natural language generation,furthermore,more and more studies are focusd on using end-to-end neueral network on task-oriented dialogue system.This paper focuses on multi-round task-oriented dialogue system for the travel domain.The main modules of the dialogue system include questions standardization module,dialogue understanding module,dialogue management module,and dialogue generation module.The main research contents of this topic include:(1)In this paper,we establish a database of intent-detection corpus and slotfilling corpus for the travel domain,so as to facilitate the intent-detection and slotfilling experiments.(2)This paper attempts to introduce question normalization module before the dialogue understanding module,aiming at solving the problem of garbled characters,emoji expressions,demonstrative pronouns and sentence substructures that may appear in the question sentence.We tried to combine with the above slot information for deletion and substitution,filling and other operations.After the standardization of the question is more likely to be correctly parsed.(3)In terms of the intent detection problem,this paper mainly studies classification method based on CNN model and LSTM model.In this paper,traditional statistical learning method SVM and random forest are used as benchmark models to compare the effects of DAN model,CNN model and LSTM model.The experimental results show that the intent detection model based on CNN and LSTM model achieves better results than other algorithms.And the LSTM based model achieved the best results with Marco-F1 value of 92.57%.(4)In terms of slot identification problem,this paper attempts the conditional random field model,BLSTM model,BLSTM-CRF model,BLSTM-CRF model with intent feature,and BLSTM-CRF model with intentional feature vocabulary.In addition,this paper attempts to identify the problem and the slot identification problem jointly,through the BLSTM-CRF model for joint learning.The experimental results show that the BLSTM-CRF model with the intent feature vocabulary has the best performance,with F1 value reaches 88.77.(5)Finally,this paper completed the dialogue management module based on strategies,and completed the dialogue generation module in a template-based manner,and completed the construction of the task-oriented dialogue system for the travel domain.