Multi-domain dialogue state tracking method and device based on reading understanding

Abstract

The invention provides a multi-field dialogue state tracking method and device based on reading understanding. The method comprises the following steps: receiving a dialogue text of a current user; based on dialogue history, the slots of the dialogue text of the current user and the corresponding natural language description, generating input sequences respectively aiming at the replicated slots and the classified slots, sending the input sequences into an XLNET encoder for encoding, and respectively obtaining a replicated slot code and a classified slot code; utilizing a neural network to obtain starting and ending positions of slot values of the replicated slots and the classified slots in the input sequence; based on the duplicated slot code and the classified slot code, utilizing a classifier to predict a duplicated slot position state and a classified slot position state respectively, determining whether to predict a corresponding slot value or not based on a prediction result, and determining the corresponding slot value based on the starting position and the ending position under the condition of needing to predict the slot value; generating a system dialog based on the determined slot value.

Description

technical field [0001] The invention relates to the field of natural language processing, in particular to a multi-field dialogue state tracking method and device based on machine reading comprehension, which can be applied to fields such as automatic question answering. Background technique [0002] At present, automated task-oriented dialogue systems are increasingly applied to practical scenarios, such as ticket booking, hotel reservations, helping users find videos, etc. Task-oriented dialogue systems can reduce customer support costs, which is almost An essential service in every business. Dialogue State Tracking (DST) is one of the core components of task-oriented dialogue systems. With the development of the Dialogue State Tracking Challenge, it has gradually become a research hotspot. A dialog state can be thought of as structured data that records user goals given a dialog history. For each dialogue turn, the dialogue state typically includes a set of all slot-val...

AI Technical Summary

Problems solved by technology

However, not only is the initial annotated dialogue system expensive to purchase, but human errors, delays in annotation, inconsistent conventions, and normalization issues make correct and timely dialogue annotation very challenging.

Images