Conversational recommendation method, method of training model, device and medium

Patent metadata

Field	Value
Publication number	US-2023088445-A1
Application number	US-202218059386-A
Country	US
Kind code	A1
Filing date	Nov 28, 2022
Priority date	Feb 18, 2022
Publication date	Mar 23, 2023
Grant date	—

Abstract

A conversational recommendation method, a method of training a conversational recommendation model, an electronic device, and a storage medium are provided, which are related to a technical field of data processing, in particular to technical fields of voice interaction, deep learning, artificial intelligence and the like. The conversational recommendation method includes: acquiring a historical conversation information; determining a target conversation object to be generated, from a conversation target graph based on the historical conversation information, the conversation target graph includes an object node, the object node is configured to represent a conversation object, and the target conversation object is determined based on the object node; and generating a target conversation information for recommendation based on the target conversation object.

First claim

What is claimed is:

1 . A conversational recommendation method, comprising:
acquiring a historical conversation information; determining a target conversation object to be generated, from a conversation target graph based on the historical conversation information, wherein the conversation target graph comprises an object node, the object node is configured to represent a conversation object, and the target conversation object is determined based on the object node; and generating a target conversation information for recommendation based on the target conversation object.

2 . The method according to claim 1 , wherein determining the target conversation object to be generated, from the conversation target graph based on the historical conversation information comprises:
determining, based on the historical conversation information and a target conversation guiding information, the target conversation object from the conversation target graph, wherein the historical conversation information is an information generated during a conversation, and the target conversation guiding information is configured to guide the generation of the target conversation object to be generated during the conversation.

3 . The method according to claim 2 , wherein determining, based on the historical conversation information and the target conversation guiding information, the target conversation object from the conversation target graph comprises:
determining a sequence of historical target conversation objects in the historical conversation information based on the historical conversation information; determining a cost parameter of a candidate object node for the conversation target graph based on the sequence of historical target conversation objects, the target conversation guiding information and the conversation target graph, wherein a type of the candidate object node matches a type of the sequence of historical target conversation objects; and determining the target conversation object from the candidate object node based on the cost parameter of the candidate object node.

4 . The method according to claim 3 , wherein determining the cost parameter of the candidate object node for the conversation target graph based on the sequence of historical target conversation objects, the target conversation guiding information and the conversation target graph comprises:
determining a transition matrix for the candidate object node based on the conversation target graph; determining a first initial cost parameter of the candidate object node based on the sequence of historical target conversation objects and the transition matrix for the candidate object node; determining a second initial cost parameter of the candidate object node based on the sequence of historical target conversation objects, the target conversation guiding information and the transition matrix for the candidate object node; and determining the cost parameter of the candidate object node based on the first initial cost parameter and the second initial cost parameter.

5 . The method according to claim 4 , wherein determining the target conversation object from the candidate object node based on the cost parameter of the candidate object node comprises:
determining a probability of switching the target conversation object node based on the cost parameter of the candidate object node; determining the target conversation object from the candidate object node based on the cost parameter of the candidate object node, in response to determining that the probability of switching is greater than or equal to a predetermined switching threshold; and determining the target conversation object from the sequence of historical target conversation objects, in response to determining that the probability of switching is smaller than the predetermined switching threshold.

6 . The method according to claim 4 , wherein determining the target conversation object from the candidate object node based on the cost parameter of the candidate object node comprises:
determining a probability of generating the target conversation object node based on the cost parameter of the candidate object node; and determining the target conversation object from the candidate object node based on the cost parameter of the candidate object node, in response to determining that the probability of generating is greater than or equal to a predetermined generation threshold.

7 . The method according to claim 4 , wherein the conversation target graph comprises a heterogeneous hierarchical conversation target graph, the heterogeneous hierarchical conversation target graph comprises a plurality of conversation target sub-graphs having a hierarchical relationship with each other, each conversation target sub-graph among the plurality of conversation target sub-graphs comprises a plurality of object nodes of the same type, a connection edge between the plurality of object nodes of the same type is configured to represent a homogeneous association relationship, the type of object nodes of one of two adjacent conversation target sub-graphs is different from the type of object nodes of the other of the two adjacent conversation target sub-graphs, and a connection edge between a plurality of objects nodes in the conversation target sub-graph at a current level and a plurality of objects nodes in the conversation target sub-graph at a higher level is configured to represent a heterogeneous association relationship.

8 . The method according to claim 7 , wherein the target conversation object comprises a plurality of target conversation objects having levels; and
wherein determining the transition matrix for the candidate object node based on the conversation target graph comprises: determining a candidate object node in the conversation target sub-graph at the current level based on the heterogeneous association relationship in the heterogeneous hierarchical conversation target graph and a determined target object node in the conversation target sub-graph at the higher level, wherein the target object node in the conversation target sub-graph at the higher level corresponds to a target conversation object at the higher level, and the candidate object node in the conversation target sub-graph at the current level corresponds to a candidate conversation object at the current level; and determining a transition matrix for the candidate object node at the current level based on the candidate object node for the conversation target sub-graph at the current level, and taking the transition matrix for the candidate object node at the current level as the transition matrix for the candidate object node.

9 . The method according to claim 1 , wherein generating the target conversation information based on the target conversation object comprises:
generating the target conversation information based on the target conversation object, the historical conversation information and the sequence of historical target conversation objects.

10 . The method according to claim 1 , wherein a type of the target conversation object comprises at least one of a conversation type of recommendation, a conversation topic and a topic attribute.

11 . A method of training a conversational recommendation model, comprising:
training the conversational recommendation model by using a training sample, to obtain a trained conversational recommendation model; wherein the trained conversational recommendation model is configured to: acquire a historical conversation information; determine a target conversation object to be generated, from a conversation target graph based on the historical conversation

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Liu Zeming
• Liu Hao
• Niu Zhengyu
• Wu Hua
• Wang Haifeng
• Xiong Hui

Classifications

• G06F40/30 Primary

  Semantic analysis · CPC title

• H04L51/02

  using automatic reactions or user delegation, e.g. automatic replies or chatbot-generated messages · CPC title

• G10L15/063 Primary

  Training · CPC title

• G10L2015/228

  of application context · CPC title

• G06F16/35

  Clustering; Classification · CPC title