A method for multi-round session framework based on cold start

Abstract

The invention discloses a method for a multi-round conversation framework based on cold start, which is characterized in that it includes setting an ontology, a topic tracker, dialogue flow management, natural language understanding, dialogue state tracking, strategy action, and natural language generation, Step 1, ontology setting; the setting content of the ontology includes topics, intentions, slots and triggering rules of the above three and robot answering skills; Type, the topic tracker as a gate only allows task-based questions to enter the subsequent multiple rounds of conversation flow; Step 3, dialogue flow management; the task-based questions are sequentially processed for natural language understanding, dialogue state tracking, strategic actions and natural language generation. Generate system responses and send them to the user. The present invention can meet the requirement of cold-starting multi-round intelligent conversational robots directly without training corpus, and realizes pluggable, ready-to-use multi-round task-based dialogue function through body setting.

Description

technical field [0001] The invention relates to the field of intelligent conversation systems, in particular to a method for a multi-round conversation framework based on cold start. Background technique [0002] With the development of AI technology, intelligent conversation technology has been vigorously developed. Task-based multi-round conversation is a technology that can conduct multi-round communication sessions with users based on specific domain tasks. Usually multi-round conversation systems are implemented by a pipeline structure. User questions utterance are processed sequentially through the pipeline structure of natural language understanding NLU, dialogue state tracking DST, policy action POLICY and natural language generation NLG, and the system replies. NLU is a natural language understanding module, which is used to understand the user's intention and slot; DST dialogue state tracking and POLICY policy learning module learn the user's state and the strate...

AI Technical Summary

Problems solved by technology

These modules can be learned and implemented by deep learning models, and can also be implemented by rules. However, in practical applications, because task-based conversational systems in different fields need to be configured with slots and pipeline structure models or methods related to different fields and intentions , only a single pipeline structure is not enough to meet the needs of commercial multi-round conversation business for out-of-the-box use, and it is difficult to cold start without corpus

[0003] Patent application document CN202010019770-A multi-round dialogue intelligent customer service system based on proper word correction and cold start, which involves technical features such as intentional recognition, slot recognition, and multi-round conversation management, but the existing problem is that the slot recognition is cold Startup is based on dictionary matching. If it is a number or date type, it is difficult to exhaustively enumerate all number combinations. Intent recognition needs to be trained with a model. The intention is not a cold start, and a complete cold start cannot be achieved.

[0004] Patent application document CN201811202675 - a method for realizing dialogue management in intelligent dialogue, which involves creating the intention flow and intention structure of dialogue management, optimizing the word slot extraction process, adding an intention jump mechanism, and integrating dialogue information in the context to realize Intention jumping during the dialogue process, word slot extraction and historical information fusion, multi-round dialogue tasks, especially the relationship between intention and word slots and how to deal with them, but the problem is that if multiple word slots are found during word slot extraction. Empty, there is no clear order of asking questions about these word slots, and it does not take into account the situation that word slots need to be filled in order in some tasks

[0005] Patent application document CN201910706857 - a training method, device, and equipment for an intelligent interactive platform, which involves providing multiple users with a solution to the content of the same conflicting slot, but this conflict is essentially due to the lack of isolation and concurrent design between users Caused by lack of consideration of intent slot state isolation for multi-user interaction

Images