Preference data set construction method and device based on multiple rounds of dialogues
Through a multi-round dialogue method, the preferred data set is constructed using dialogue history information and clustering technology, which solves the complex and time-consuming problems in the process of building preferred data, and improves the data generation efficiency and dialogue effect.
CN119293501APending Publication Date: 2025-01-10SHANGHAI SHANGHU INFORMATION TECH CO LTD +2
View PDF 0 Cites 0 Cited by
Patent Information
- Application Number
- CN202411315755.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-20
- Publication Date
- 2025-01-10
AI Technical Summary
Technical Problem
There are complex and time-consuming problems in building preferred data, which are mainly due to the need for manual annotation and data aggregation, resulting in inefficiency.
Method used
Using a multi-round dialogue method, by obtaining multiple dialogue historical information, the target user speech and its predecessor dialogue rounds, positive feedback and negative feedback seat speech are determined, and the preferred data samples are constructed, and dialogue trees and clustering technology are used to improve. Data generation efficiency.
Benefits of technology
It improves the efficiency of generating preferred data sets, reduces the subjective preferences of manual labeling, and the generated data is closer to actual usage scenarios, and has better dialogue effects.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN119293501A_ABST
Abstract
The embodiment of the invention provides a preference data set construction method and device based on multiple rounds of dialogues, and the method comprises the steps: obtaining a plurality of pieces of dialogue historical information, and enabling any piece of dialogue historical information to comprise at least one round of dialogue information between an agent and a user; determining a plurality of target user verbal skills from the plurality of pieces of dialogue historical information, wherein at least two pieces of dialogue historical information to which the target user verbal skills belong respectively have positive feedback indexes and negative feedback indexes; for at least one target user verbal skill in the multiple target user verbal skill, constructing a preference data sample corresponding to each target user verbal skill so as to obtain a preference data set; any preference data sample comprises a preorder conversation round, a positive feedback seat verbal skill and a negative feedback seat verbal skill of a target user verbal skill, and both the positive feedback seat verbal skill and the negative feedback seat verbal skill belong to the same round of conversation information with the target user verbal skill.
Need to check novelty before this filing date? Find Prior Art