CN-121979993-A - Answer recommendation method and device

CN121979993ACN 121979993 ACN121979993 ACN 121979993ACN-121979993-A

Abstract

The invention provides an answer recommendation method and device, which are used for automatically classifying questions of a user, summarizing similar questions into the same question classification, generating corresponding answer recommendation sets, and enabling a target user to select answer recommendations in the answer recommendation sets to directly answer, so that fragmented consultation which is required to be processed piece by piece is converted into efficient batch decision and response processes, the interaction efficiency is remarkably improved, and the use experience of the user and the target user is effectively improved.

Inventors

PENG ZHANKUI
HE YU
ZHAO YIHONG
ZHAO WANLIN

Assignees

湖南快乐阳光互动娱乐传媒有限公司

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. An answer recommendation method, comprising: Receiving problem information of a user; Performing cluster analysis based on the problem information and the label problems in the knowledge base to obtain a sentence semantic vector list; Generating a semantic association information list based on the sentence semantic vector list and an association model, wherein the semantic association information list comprises at least one semantic association information, and each semantic association information comprises two sentence semantic vectors and the association degree between the two sentence semantic vectors; generating a correlation undirected graph based on the semantic correlation information list, wherein nodes in the correlation undirected graph are generated by sentence semantic vectors in the semantic correlation information, and edges in the correlation undirected graph are established based on the correlation degree between the two sentence semantic vectors; classifying and analyzing the nodes in the association undirected graph to generate a classifying and analyzing result, wherein the classifying and analyzing result comprises a problem classifying label of at least one node partition; And aiming at each question classifying label, generating an answer recommendation set corresponding to the question classifying label based on the questions associated with the question classifying label and the historical dialogue records corresponding to the question classifying label in the knowledge base.
2. The answer recommendation method according to claim 1, wherein said generating a semantic association information list based on said sentence semantic vector list and an association model comprises: Inputting any two sentence semantic vectors in the sentence semantic vector list into a relevancy analysis model to generate relevancy between the two sentence semantic vectors, wherein the relevancy analysis model comprises a linear transformation layer and a mixed attention layer; and generating semantic association information based on the two sentence semantic vectors and the association degree between the two sentence semantic vectors.
3. The answer recommendation method of claim 1, wherein the nodes in the associated undirected graph comprise template nodes and input nodes, the classifying analysis is performed on the nodes in the associated undirected graph to generate a classification analysis result, and the method comprises: And aiming at each template node, determining a first target input node belonging to a node partition corresponding to the template node layer by using a dynamic relevance threshold and the relevance among the nodes by taking the template node as a starting point to obtain a first classification analysis result, wherein the dynamic relevance threshold of each layer is determined based on an initial relevance threshold and a control factor, and the control factor is determined based on the number of nodes in different node partitions and the number of nodes which should be included.
4. The answer recommendation method of claim 3, further comprising: If the association degree of the current input node and all the template nodes is smaller than the minimum association degree threshold value, the input node is used as a target node; For each target node, determining a second target input node belonging to a node partition corresponding to the target node layer by using the dynamic association threshold and the association degree between nodes by taking the target node as a starting point until the association degree between no input node and all the template nodes is smaller than the minimum association degree threshold at present, and obtaining a second classification analysis result; And generating a final classification analysis result according to the first classification analysis result and the second classification analysis result.
5. The answer recommendation method of claim 1, further comprising: When an online event of a target user is monitored, determining the size of a first dynamic virtual congestion window based on real-time load information of a current pipeline; and based on the size of the first dynamic virtual congestion window, after acquiring the problem information corresponding to the target user in a problem pool, starting to execute the cluster analysis based on the problem information and the label problems in a knowledge base to obtain a statement semantic vector list.
6. The answer recommendation method according to claim 5, wherein the generating, for each of the question classification tags, an answer recommendation set corresponding to the question classification tag based on the questions associated with the question classification tag and the history dialogue record corresponding to the question classification tag in the knowledge base, further comprises: Receiving target answer information of the target user, wherein the target answer information is a recommended answer selected by the target user in an answer recommendation set corresponding to a question classification label or an answer modified by the target user for the recommended answer in the answer recommendation set corresponding to the question classification label; determining a second dynamic virtual congestion window size based on real-time load information of the current pipeline; and feeding back target answer information of the target user to a corresponding user of the corresponding question in the question classification label based on the second dynamic virtual congestion window size.
7. The answer recommendation method of claim 2, further comprising: Generating rewarding parameters based on the evaluation information of the problem classification labels; constructing a state transition process based on the semantic association information and the reward parameter; And further optimizing the association degree analysis model by using the state transfer process to obtain an optimized association degree analysis model.
8. The answer recommendation method of claim 6, further comprising: After receiving the question and answer information, auditing the content in the question and answer information to obtain an auditing result, wherein the question and answer information comprises the question information of the user and the target answer information; And intercepting question and answer information with risk as the auditing result.
9. The answer recommendation method of claim 1, further comprising, after receiving the question information of the user: If the language type of the problem information of the user is inconsistent with the common language type of the target user, the problem information of the user is translated into the common language type of the target user.
10. An answer recommendation device, comprising: a first receiving unit for receiving problem information of a user; the cluster analysis unit is used for carrying out cluster analysis based on the problem information and the label problems existing in the knowledge base to obtain a sentence semantic vector list; The related information generation unit is used for generating a semantic related information list based on the statement semantic vector list and the related model, wherein the semantic related information list comprises at least one semantic related information, and each semantic related information comprises two statement semantic vectors and a degree of association between the two statement semantic vectors; The generation unit of the association undirected graph is used for generating the association undirected graph based on the semantic association information list, wherein nodes in the association undirected graph are generated by sentence semantic vectors in the semantic association information, and edges in the association undirected graph are established based on the association degree between the two sentence semantic vectors; the classification analysis unit is used for carrying out classification analysis on the nodes in the association undirected graph to generate a classification analysis result, wherein the classification analysis result comprises a problem classification label of at least one node partition; And the answer recommendation generation unit is used for generating an answer recommendation set corresponding to the question classification label according to the question associated with the question classification label and the historical dialogue record corresponding to the question classification label in the knowledge base aiming at each question classification label.

Description

Answer recommendation method and device Technical Field The invention relates to the technical field of computers, in particular to an answer recommendation method and device. Background At present, the interactive platform dialogue system has the following pain point problems when providing services for users and target users, the target users usually need to face a huge amount of dialogues, the reply cannot be timely and effectively performed, and the huge amount of dialogues can lead to the reduction of the reply will of the target users, so that the interaction enthusiasm of the users is further reduced and the use frequency is reduced under the condition that the users can not receive the reply frequently. Disclosure of Invention In view of the above, the present invention provides an answer recommendation method and apparatus to solve the problem of low interaction efficiency in the prior art. The first aspect of the present invention provides an answer recommendation method, including: Receiving problem information of a user; Performing cluster analysis based on the problem information and the label problems in the knowledge base to obtain a sentence semantic vector list; Generating a semantic association information list based on the sentence semantic vector list and an association model, wherein the semantic association information list comprises at least one semantic association information, and each semantic association information comprises two sentence semantic vectors and the association degree between the two sentence semantic vectors; generating a correlation undirected graph based on the semantic correlation information list, wherein nodes in the correlation undirected graph are generated by sentence semantic vectors in the semantic correlation information, and edges in the correlation undirected graph are established based on the correlation degree between the two sentence semantic vectors; classifying and analyzing the nodes in the association undirected graph to generate a classifying and analyzing result, wherein the classifying and analyzing result comprises a problem classifying label of at least one node partition; And aiming at each question classifying label, generating an answer recommendation set corresponding to the question classifying label based on the questions associated with the question classifying label and the historical dialogue records corresponding to the question classifying label in the knowledge base. Optionally, the generating a semantic association information list based on the sentence semantic vector list and the association model includes: Inputting any two sentence semantic vectors in the sentence semantic vector list into a relevancy analysis model to generate relevancy between the two sentence semantic vectors, wherein the relevancy analysis model comprises a linear transformation layer and a mixed attention layer, and the mixed attention layer comprises a soft attention mechanism and a self-attention mechanism; and generating semantic association information based on the two sentence semantic vectors and the association degree between the two sentence semantic vectors. Optionally, the nodes in the association undirected graph include template nodes and input nodes, and the classifying analysis is performed on the nodes in the association undirected graph to generate a classification analysis result, which includes: For each template node, determining a first target input node belonging to a node partition corresponding to the template node layer by using a dynamic association threshold and association between nodes by taking the template node as a starting point to obtain a first classification analysis result, wherein the dynamic association threshold of each layer is determined based on an initial association threshold and a control factor, and the control factor is determined based on the number of nodes in different node partitions and the number of nodes which should be included; optionally, the answer recommendation method further includes: If the association degree of the current input node and all the template nodes is smaller than the minimum association degree threshold value, the input node is used as a target node; For each target node, determining a second target input node belonging to a node partition corresponding to the target node layer by using the dynamic association threshold and the association degree between nodes by taking the target node as a starting point until the association degree between no input node and all the template nodes is smaller than the minimum association degree threshold at present, and obtaining a second classification analysis result; And generating a final classification analysis result according to the first classification analysis result and the second classification analysis result. Optionally, after receiving the problem information of the user, the method further includes: and storing the problem informa