Participant-Driven Speech Annotation for Conversation-State Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for estimating conversation states from speech data face challenges due to insufficient data annotation and subjective label determination by third parties, leading to inaccurate annotations.
Innovation Solution
An information processing device that determines whether speech data is an insufficient annotation candidate and requests annotation directly from the conversation's participants, using provisional labels to narrow down the data and reduce user burden.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If annotation is performed by third parties who do not participate in actual conversation, then annotation can be performed without involving conversation participants, but annotation accuracy deteriorates due to subjective view limitations
Solution Approach 1:
The patent enables conversation participants to annotate their own speech data. The notification device sends annotation requests to the actual speakers or listeners, who then provide annotations based on their direct experience of the conversation. This self-service approach ensures that annotations reflect the true context and meaning of the speech, resolving the contradiction between ease of operation and annotation accuracy.
2Quantity of substance
If all speech data is annotated to improve model training, then annotation completeness improves, but user burden increases due to excessive annotation requests
Solution Approach 1:
The patent implements selective annotation by determining whether each speech data requires annotation before sending requests. The notification device evaluates speech data characteristics and only notifies users to annotate when necessary, rather than requesting annotation for all speech data. This partial action approach maintains annotation completeness for critical data while reducing overall user burden.
Solution Approach 2:
The system performs preliminary determination of annotation necessity before sending annotation requests. By pre-assessing which speech data requires annotation based on predefined criteria, the system avoids unnecessary notification and annotation requests, thereby reducing user burden while maintaining necessary annotation completeness.
Data Source
AI summary
An information processing device that is configured to: in a case of annotation of a machine learning model that estimates a state of a party to a conversation from speech data of the conversation, determine whether or not the speech data is an insufficient predetermined annotation candidate, and in a case in which it is determined that the speech data is an insufficient predetermined annotation candidate, request annotation from the party to the conversation, who is at least one of a speaker or a listener of the conversation.