CN-121980078-A - Conversational content recommendation method, device, equipment and medium

CN121980078ACN 121980078 ACN121980078 ACN 121980078ACN-121980078-A

Abstract

The application belongs to the field of artificial intelligence, and relates to a conversational content recommendation method, device, equipment and medium. And acquiring a time sequence attenuation coefficient, and calculating the weight of each round of historical dialogue interaction data of the user according to the time sequence attenuation coefficient. And then obtaining a target weight coefficient of the target demand vector, combining the weight and the target weight coefficient, and fusing the target demand vector and the historical interaction data to obtain the enhanced demand vector. And adopting a second large language model to carry out demand disassembly and strategy matching on the enhanced demand vector, and obtaining a recommended strategy instruction. Searching the candidate content set in the content database according to the instruction, obtaining a target candidate content set after knowledge graph optimization, and finally generating personalized recommended content based on the target candidate content set and the enhanced demand vector. The method and the device can be applied to the business fields of finance, science and technology and the like, and can improve the accuracy and individuation of content recommendation.

Inventors

DENG YUWEI
KONG LINGWEI

Assignees

平安科技（深圳）有限公司

Dates

Publication Date: 20260505
Application Date: 20260107

Claims (10)

1. A conversational content recommendation method, comprising the steps of: Receiving dialogue text input by a target user in a current dialogue, wherein the target user has multiple historical dialogues in history; Carrying out semantic analysis on the dialogue text through a first large language model to obtain a target demand vector of the target user; acquiring a time sequence attenuation coefficient, and calculating the weight of historical interaction data of each round of historical conversations based on the target demand vector and the time sequence attenuation coefficient to obtain a weight set; acquiring a target weight coefficient of the target demand vector, and fusing the target demand vector and the historical interaction data based on the weight set and the target weight coefficient to obtain an enhanced demand vector of a fused context; adopting a second large language model to carry out demand disassembly and strategy matching on the enhanced demand vector to obtain a recommended strategy instruction; Searching a candidate content set in a content database according to the recommendation strategy instruction, and optimizing the candidate content set through a knowledge graph to obtain a target candidate content set; And generating personalized recommended content of the target user based on the target candidate content set and the enhanced demand vector.
2. The method according to claim 1, wherein the step of performing semantic parsing on the dialog text through a first large language model to obtain the target demand vector of the target user specifically includes: Word segmentation processing is carried out on the dialogue text through a first large language model, so that a word vector sequence is obtained; And inputting the word vector sequence into an encoder of the first large language model for semantic coding, and outputting the requirement representation vector of the target user.
3. The method according to claim 1, wherein the step of calculating weights of the historical interaction data of each historical dialog based on the target demand vector and the time sequence attenuation coefficient to obtain a weight set specifically comprises: Calculating the correlation between the target demand vector and the historical interaction data of each round of historical conversations, wherein the historical interaction data is stored in a historical conversational buffer area; and calculating the weight of the historical interaction data based on the correlation and the time sequence attenuation coefficient to obtain a weight set.
4. The method according to any one of claims 1-3, wherein the historical interaction data includes a historical demand representation vector, a recommendation result vector, and a user feedback vector, and the step of fusing the target demand vector and the historical interaction data based on the weight set and the target weight coefficient to obtain an enhanced demand vector of a fused context specifically includes: based on the weight set, determining weight coefficients corresponding to the historical demand representation vector, the recommendation result vector and the user feedback vector; and carrying out weighted summation on the target weight coefficient, the target demand vector, the weight coefficient, the historical demand expression vector, the recommended result vector and the user feedback vector to obtain an enhanced demand vector integrating the context.
5. The method of claim 1, wherein the step of performing demand disassembly and policy matching on the enhanced demand vector by using a second large language model to obtain a recommended policy instruction specifically includes: Based on the enhanced demand vector, adopting a second large language model to disassemble demands, and determining key sub-demands; Based on the enhanced demand vector, constructing a recommendation strategy instruction matched with the key sub-demand by adopting an autoregressive generation mode through a second large language model.
6. The method according to claim 1, wherein the step of retrieving the candidate content set from the content database according to the recommendation policy instruction, and optimizing the candidate content set through a knowledge graph to obtain the target candidate content set specifically includes: Converting the recommended strategy instruction into a search query vector; Searching a plurality of candidate vectors with the similarity with the search query vector in the front in a vector database through an approximate nearest neighbor search algorithm; obtaining product content corresponding to each candidate vector to obtain a candidate content set; For each product content in the candidate content set, inquiring an associated entity corresponding to each product content from a knowledge graph; calculating the similarity between the associated entity and the search query vector; And if the similarity is greater than a preset threshold, adding the associated entity into the candidate content set to optimize the candidate content set to obtain a target candidate content set.
7. The method according to claim 1, wherein the step of generating personalized recommended content for the target user based on the target candidate content set and the enhanced demand vector, specifically comprises: Based on the target candidate content set and the enhanced demand vector, generating a personalized recommendation document corresponding to each candidate content in the target candidate content set by adopting an autoregressive generation mode through a third large language model; performing association integration on the personalized recommendation file and the multi-mode information of each candidate content to obtain multi-dimensional recommendation content corresponding to each candidate content; Calculating the similarity between the enhancement demand vector and each candidate content; Calculating the display sequence score of each candidate content according to the similarity and the historical feedback information of the target user; And adjusting the display sequence of the multi-dimensional recommended content of each candidate content based on the display sequence score, and generating the personalized recommended content of the target user.
8. A conversational content recommendation device, comprising: The receiving module is used for receiving dialogue texts input by a target user in the current dialogue round, and the target user has multiple historical dialogue rounds historically; the analysis module is used for carrying out semantic analysis on the dialogue text through a first large language model to obtain a target demand vector of the target user; The calculation module is used for acquiring a time sequence attenuation coefficient, and calculating the weight of the historical interaction data of each round of historical conversations based on the target demand vector and the time sequence attenuation coefficient to obtain a weight set; The fusion module is used for acquiring a target weight coefficient of the target demand vector, and fusing the target demand vector and the historical interaction data based on the weight set and the target weight coefficient to obtain an enhanced demand vector of a fused context; The strategy matching module is used for carrying out requirement disassembly and strategy matching on the enhanced requirement vector by adopting a second large language model to obtain a recommended strategy instruction; The retrieval module is used for retrieving the candidate content set in the content database according to the recommendation strategy instruction, and optimizing the candidate content set through the knowledge graph to obtain a target candidate content set; And the generation module is used for generating personalized recommended content of the target user based on the target candidate content set and the enhanced demand vector.
9. A computer device comprising a memory and a processor, the memory having stored therein computer readable instructions which when executed by the processor implement the steps of the conversational content recommendation method of any one of claims 1 to 7.
10. A computer readable storage medium having stored thereon computer readable instructions which when executed by a processor implement the steps of the conversational content recommendation method of any one of claims 1 to 7.

Description

Conversational content recommendation method, device, equipment and medium Technical Field The application relates to the technical field of artificial intelligence, is applied to online processing of business scenes in financial science and technology, and particularly relates to a conversational content recommendation method, device, equipment and medium. Background In the information age, a recommendation system is used as a key technology of information filtering and personalized service and is widely applied to various fields such as electronic commerce, content platforms and the like. The recommendation system in the current industry mainly covers two types of traditional algorithm driving type and preliminary semantic understanding type. The core of the traditional algorithm-driven recommendation system depends on collaborative filtering algorithms, content filtering algorithms and the like. Taking an e-commerce platform as an example, a recommendation mode based on matching of user purchase history and commodity attributes is often adopted in early stage, and commodities with similar attributes are recommended to a user by analyzing past purchase behaviors of the user. And the video platform recommends similar contents according to the watching time length and the clicking record of the user. However, such systems can only mechanically match based on user past behavior data, lack active interaction capability with the user, and have difficulty capturing dynamic changes in user demand in real time. The preliminary semantic understanding type recommendation system introduces simple natural language processing technology in recent years, such as semantic parsing based on keyword matching. Some content platforms attempt to recommend relevant types of content to a user by identifying core keywords in the user input text, such as "joke", "suspense", and the like. But this approach only captures the surface information of the user's needs, and cannot be accurately understood and processed for complex needs that contain multiple conditions, ambiguous expressions, or implicit preferences. In summary, the existing recommendation system cannot fully and deeply understand the user requirements, and particularly, it is difficult to accurately grasp the current requirements by combining the historical dialogue information of the user, so that the accuracy and individuation degree of the recommendation result are limited. Disclosure of Invention The embodiment of the application aims to provide a conversational content recommendation method, a conversational content recommendation device, computer equipment and a storage medium, so as to solve the problems that an existing recommendation system cannot fully and deeply understand user demands, and particularly the accuracy and individuation degree of a recommendation result are limited because the current demands are difficult to accurately grasp by combining historical conversational information of a user. In a first aspect, a conversational content recommendation method is provided, and the following technical scheme is adopted: the method comprises the steps of receiving dialogue texts input by a target user in a current dialogue, carrying out semantic analysis on the dialogue texts through a first large language model to obtain target demand vectors of the target user, obtaining time sequence attenuation coefficients, calculating weights of historical interaction data of each round of the historical dialogue based on the target demand vectors and the time sequence attenuation coefficients to obtain a weight set, obtaining target weight coefficients of the target demand vectors, carrying out fusion on the target demand vectors and the historical interaction data based on the weight set and the target weight coefficients to obtain enhanced demand vectors of fused contexts, carrying out demand disassembly and strategy matching on the enhanced demand vectors through a second large language model to obtain recommendation strategy instructions, searching candidate content sets in a content database according to the recommendation strategy instructions, optimizing the candidate content sets through knowledge maps to obtain target candidate content sets, and generating personalized recommendation contents of the target user based on the target candidate content sets and the enhanced demand vectors. In a second aspect, a conversational content recommendation apparatus is provided, which adopts the following technical scheme: The receiving module is used for receiving dialogue texts input by a target user in the current dialogue round, and the target user has multiple historical dialogue rounds in history; The analysis module is used for carrying out semantic analysis on the dialogue text through the first large language model to obtain a target demand vector of a target user; The calculation module is used for acquiring a time sequence attenuation coefficient, and calcu