Search

CN-121996748-A - Large model dialogue control method, large model dialogue control system, storage medium and computer equipment

CN121996748ACN 121996748 ACN121996748 ACN 121996748ACN-121996748-A

Abstract

The application discloses a large model dialogue control method, a large model dialogue control system, a storage medium and computer equipment. The large model dialogue method comprises the steps of converting user input data and dialogue history data into state perception vectors, selecting a plurality of expert models matched with the state perception vectors from a preset expert model pool, respectively inputting the user input data and the dialogue history data into the selected expert models to obtain a plurality of candidate replies, calculating the value scores of the candidate replies, and outputting the candidate replies with the highest value scores as final dialogue replies. Through the method, the dialogue context semantics and the user core intention can be accurately captured, the knowledge of the multi-expert model is dynamically activated in a sparse mode and integrated, the consumption of computing resources and the reasoning delay are obviously reduced, the accuracy and the reliability of dialogue reply and the adaptability across styles and fields are improved, the process has strong interpretability and controllability, and the large model is convenient to debug and optimize.

Inventors

  • HU BIN

Assignees

  • 深圳市有方科技股份有限公司

Dates

Publication Date
20260508
Application Date
20251211

Claims (10)

  1. 1. A method for large model dialog control, comprising: Converting user input data and dialogue history data into state sensing vectors; selecting a plurality of expert models matched with the state sensing vector from a preset expert model pool; Respectively inputting the user input data and the dialogue history data into each expert model to obtain a plurality of candidate replies; and calculating the value score of each candidate reply, and outputting the candidate reply with the highest value score as a final dialogue reply.
  2. 2. The large model dialog control method of claim 1, wherein the state-aware vector comprises a dialog state vector; the converting the user input data and the dialogue history data into state sensing vectors comprises: and carrying out vector coding on the user input data and the dialogue history data to obtain the dialogue state vector.
  3. 3. The large model dialog control method of claim 2, wherein the state-aware vector further comprises a memory pool aggregate vector; The converting the user input data and the dialogue history data into state sensing vectors further comprises: extracting descriptive objects and associated information of the descriptive objects in the user input data and the dialogue history data; Vectorizing the mapping relation between each description object and the associated information, and updating the mapping relation to a dialogue state memory pool; And carrying out weighted aggregation on a plurality of vectors in the dialogue state memory pool to obtain the memory pool aggregate vector.
  4. 4. The large model dialogue control method according to claim 1, wherein selecting a plurality of expert models matching the state-aware vector from a preset expert model pool includes: calculating the matching probability of the state sensing vector and each expert model in the preset expert model pool; And selecting and activating a preset first number of expert models with highest matching probability with the state sensing vector in the preset expert model pool.
  5. 5. The large model dialog control method of claim 1, further comprising, after the converting the user input data and the dialog history data into state-aware vectors: Inquiring a preset second number of knowledge pieces matched with the state sensing vector from a preset knowledge base; the knowledge segments are introduced into the calculation of a value score.
  6. 6. The large model dialog control method of claim 5, wherein the value score comprises a routing probability factor and at least one quality assessment factor, the quality assessment factor comprising a factual assessment factor; Before calculating the value score of each candidate reply and outputting the candidate reply with the highest value score as a final dialogue reply, the method further comprises: Extracting a claim statement from the candidate reply; calculating the implication probability of each claim statement and each knowledge segment; and taking the representative value in each implication probability as the facts assessment factor.
  7. 7. The large model dialog control method of claim 6, wherein the calculating a value score for each of the candidate replies and outputting the candidate reply with the highest value score as a final dialog reply comprises: Carrying out weighted summation on the routing probability factors and the quality assessment factors to obtain the value scores corresponding to the candidate replies; comparing the value scores of the candidate replies, and outputting the candidate replies with the highest value scores as the final dialogue replies.
  8. 8. A large model dialog control system, comprising: The dialogue state sensing module is used for converting user input data and dialogue history data into state sensing vectors; the multi-expert routing module is used for selecting a plurality of expert models matched with the state sensing vector from a preset expert model pool; the expert model pool module is used for respectively inputting the user input data and the dialogue history data into the selected expert models to obtain a plurality of candidate replies; And the knowledge enhancement fusion and rearrangement module is used for calculating the value score of each candidate reply and outputting the candidate reply with the highest value score as a final dialogue reply.
  9. 9. A storage medium having stored thereon program data, wherein the program data, when executed by a processor, implements the steps of the large model dialog control method of any of claims 1 to 7.
  10. 10. A computer device comprising a processor and a memory connected to each other, the memory storing a computer program, the processor implementing the steps of the large model dialog control method according to any of claims 1 to 7 when executing the computer program.

Description

Large model dialogue control method, large model dialogue control system, storage medium and computer equipment Technical Field The application relates to the technical field of artificial intelligence, in particular to a large model dialogue control method, a large model dialogue control system, a large model dialogue control storage medium and a large model dialogue control computer device. Background In the field of artificial intelligence dialogue systems, a dialogue system based on a large language model has realized smooth and coherent open-domain dialogue, and the core of the dialogue system depends on a single parameter large-scale pre-training model to learn language rules and world knowledge through mass data. In order to improve the accuracy and the specificity of the dialogue, the technology of search enhancement generation (RETRIEVAL AUGMENTED GENERATION, RAG) is widely applied in the industry, and the technology retrieves relevant information fragments from an external knowledge base before generating replies, and uses the relevant information fragments as contexts to be input in combination with large model prompt words so as to guide the model to generate contents with more practical basis. In addition, aiming at different business scenes or individual demands, the prior art generally adopts a mode of designing specific Prompt words (promts) or carrying out full-parameter fine adjustment on a basic large model to obtain a plurality of independent dialogue models which are adaptive to different styles or fields. However, the prior art still has a number of core drawbacks to be solved, and reliability of large model dialogue reply is poor. On the one hand, the traditional single model architecture has inherent contradiction of capability and efficiency, and is easy to generate actual deviation, namely 'knowledge illusion', and huge monomer models need to activate all parameters when processing simple inquiry, so that the consumption of computing resources is high, the response delay is large, the existing retrieval enhancement generation technology only takes external knowledge as a reference, no effective constraint is applied to the model generation process, and the model can ignore or misunderstand the retrieved knowledge to generate a reply which is inconsistent with the fact. On the other hand, the stylized/specialized scheme based on the prompt words or full-parameter fine adjustment is not flexible enough and low in efficiency, the method for controlling the styles by the prompt words is poor in stability, the styles can be out of control due to slight expression change, a large amount of labeling data and computing resources are needed for the model of each scene in full-parameter fine adjustment, the cost is high, and the model can only be switched in a hard mode when the styles or fields are switched, so that the consistency and the user experience of the dialog are destroyed. Therefore, a session control manner is needed that can effectively avoid the problems of knowledge illusion, computing resource waste, insufficient self-adaptive capacity of style and field, poor session consistency and the like caused by strongly coupling tasks such as knowledge storage, style control and language generation in a single model under the existing architecture, and can control a large model to perform efficient, accurate and flexible session reply. Disclosure of Invention The application mainly provides a large model dialogue control method, a large model dialogue control system, a storage medium and computer equipment, and aims to solve the technical problem that the reliability of the existing large model dialogue reply is poor. In order to solve the technical problems, the technical scheme adopted by the application is to provide a large model dialogue control method. The large model dialogue control method comprises the steps of converting user input data and dialogue history data into state perception vectors, selecting a plurality of expert models matched with the state perception vectors from a preset expert model pool, respectively inputting the user input data and the dialogue history data into the selected expert models to obtain a plurality of candidate replies, calculating the value scores of the candidate replies, and outputting the candidate replies with the highest value scores as final dialogue replies. In some embodiments, the state-aware vector comprises a dialog state vector, and the converting the user input data and the dialog history data into state-aware vectors comprises vector encoding the user input data and the dialog history data to obtain the dialog state vector. In some embodiments, the state sensing vector further comprises a memory pool aggregate vector, and the converting the user input data and the dialogue history data into the state sensing vector comprises extracting description objects in the user input data and the dialogue history data and associated inform