Dialogue state tracking method based on multi-field and multi-round questions and answers

Abstract

The invention provides a dialogue state tracking method based on multi-field and multi-round questions and answers. The method comprises the following steps that: S1, a dialogue state model is established, and the dialogue state model is encoded; S2, on the basis of a Domain Interest Prediction module, an interest field of dialogue round is judged; S3, a mode of generating a value for each slot type is judged through a Slot gate Predication module; and S4, the value in the slot is judged on the basis of a Span Prediction module, and a word is generated finally.

Description

technical field [0001] The invention relates to the technical field of man-machine question answering, in particular to a dialogue state tracking method based on multi-field multi-round question answering. Background technique [0002] Existing multi-domain DST models, such as TRADE and SOM-DST, have their own problems. First of all, there is inherent heterogeneity in various types of slot values ​​in multiple domains. For example, restaurant-name, attraction-name and other slot values ​​can be classified as named entities. The values ​​of slots such as restaurant-book people and train-book time can all be classified as number-type entities; there are also some slots such as hotel-parking and hotel-internet that belong to the Boolean type, which require users to judge yes or no. Some models, such as HyST, DS-DST, etc., capture these heterogeneous information by manually defining rules. Although effective, it increases the complexity of the model. In addition, models using ...

AI Technical Summary

Problems solved by technology

However, SOM-DST did not show the accuracy or recall of the classifier in the paper, and its final performance actually has a huge room for improvement.

Images