Dialogue-enriched Knowledge Point Recommendation for Consultation Task

Abstract

The customer service agent of the government service hotline has been providing great help to people’s daily lives. To facilitate further work analysis, after the customer service agent completes the dialogue consultation each time, it is necessary to select the related knowledge point for summarizing and record the dialogue from the large-scale knowledge point base according to the current consultation dialogue content. However, this work adds a great burden to the customer service staff. In order to reduce the workload of customer service agents in selecting knowledge points, we explore several approaches based on the consultation dialogue in the customer service scenario, e.g., the text classification model based on BERT, the word vector convolutional neural network classification model, and the pseudo-siamese neural network model. Extensive experiments demonstrated that our methods perform well in terms of accuracy and scalability, and further we designed an efficient and well-structured API module which is easily integrated into service application.