Anaphora resolution method for multi-round dialogue system

Abstract

The invention provides an anaphora resolution method for a multi-round dialogue system, which comprises the following steps of: S1, detecting statements received by the multi-round dialogue system, judging whether the statements need anaphora resolution or not, and if so, entering the step S2; S2, judging the statements determined to be required to be subjected to anaphora resolution, distinguishing anaphora types of the statements, and screening candidate entities from the statements of which the anaphora types are distinguished; wherein the anaphora type of the statement comprises a back-pointing statement and a co-pointing statement; S3, determining the distances between candidate entities and the demonstrative pronouns in statements, and taking the candidate entities with the smallestdistance as demonstrative linking words; S4, updating the demonstrative pronouns into demonstrative linking words; The demonstrative pronouns of the statements input into the multi-round dialogue system can be accurately recognized, accurate anaphora resolution is carried out, smoothness of the multi-round interaction system can be effectively improved, and user experience is improved.

Description

technical field [0001] The invention relates to the field of man-machine dialog language processing, in particular to a method for resolution of anaphora in a multi-round dialog system. Background technique [0002] Reference is a common linguistic phenomenon. In multi-turn dialogue systems, reference and omission affect the fluency and user experience of human-computer dialogue systems. The human brain has a knowledge background and memory ability, and can reproduce the key information of the dialogue history very well, and automatically complete or replace the pronouns of the other party's current turn to understand the true meaning of the dialogue. Similarly, a reference resolution module is needed in a human-computer dialogue system to understand the real semantics of user input. [0003] The anaphora resolution methods are divided into pipeline-based methods and end-to-end based methods. Pipeline-based approaches perform coreference detection first, followed by coref...

AI Technical Summary

Problems solved by technology

End-to-end methods usually use deep learning models, but require a large amount of labeled data, and the cost of model construction is relatively high in the face of lack of domain data, especially in the early stages of multi-round referral resolution. Encountered problems in the process, the model modification iteration does not meet the timeliness of the online system

Images