Incorporating Self-training And Active Learning For Intention Detection

In recent years,with the development of artificial intelligence technology,its practical application has achieved remarkable results,and natural language understanding is an important part of the task-based dialogue system has also received widespread attention.When a dialogue system interacts with a user,the system first needs to translate human language into machine-readable structured information and ”understand” the intent contained in human language,and then the system responds accordingly to the intent.However,in a real-world scenario,labeling large-scale,high-quality data requires significant labor and time costs.Moreover,unlike text data in other domains,customer service conversation histories in e-commerce have much longer text,and current natural language understanding technologies are limited by hardware performance to handle excessively long text input or require significant computational overhead.Therefore,the intention detection system in e-commerce domain faces two major challenges: on the one hand,the long text problem,i.e.,how to simplify the text information without losing too much of the user intention information it contains to reduce the computational overhead;on the other hand,the lack of annotated data problem,i.e.,how to utilize a small amount of annotated data and large-scale unannotated data to reduce the labor cost.To address the above issues,the following research work is conducted in this thesis:(1)Self-supervised learning-based key dialogue discrimination model.This model identifies and extracts key sentences from long texts by simulating the thinking process of human reading comprehension through the features of self-supervised tasks in pre-trained language models.In addition,we used the model to construct a Customer Service Intent dataset(CSI)for e-commerce post-sales conversations based on the JDDC dataset.(2)Intent detection model based on self-training and active learning.The model is based on pre-trained language model for self-training learning on a small amount of data,and we propose an active learning strategy,which combines the self-training process to find the highest uncertainty group for each batch of candidate samples in each round of self-training and send them to the database for manual annotation,so as to achieve the performance of the model with the least annotation cost.This maximizes the performance of the model with minimum labeling cost.In addition,to reduce the pseudo label noise problem during the self-training process,we propose an adaptive active threshold module,which balances the number of generated pseudo labels and the number of samples to be labeled according to the performance of the model and the size and number of categories of the training target dataset.Our experiments on AG'News,IMDB and CSI datasets show that the model effectively improves the fitting speed and accuracy.(3)USS intent detection system based on commodity event augmentation.In order to solve the problems of large ambiguity of intention representation and rapid growth of new intentions in real scenarios,we developed a USS intention detection system based on commodity event augmentation.To improve the accuracy of intent detection in real applications,we preprocessing historical dialogues,extract key dialogues,and generate commodity events,which are added to the USS intent detection model training as a kind of knowledge information to help the downstream intent detection tasks achieve better results.The system is able to maintain the accuracy of the intent detection system at all times while achieving the goal with minimal human cost,and can cope with realistic intent detection scenarios that are constantly being iterated and updated.