Question Group Extraction via State Transition Labeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies fail to efficiently extract question groups related to an arbitrary topic from conversations, particularly when conversations include multiple topics and lack annotations, and they cannot provide user-determined actions based on user responses.
Innovation Solution
A question group extraction method that labels problem, question, and answer sentences in conversation data using a state transition model, allowing for the association and extraction of question groups based on these labels, enabling the identification of question groups and user actions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If a topic dialogue method is used to automatically generate questionnaires, then the time required to generate templates is reduced, but the system cannot extract question groups related to arbitrary topics from conversation histories without annotations
Solution Approach 1:
The state transition model is designed to be topic-agnostic and annotation-independent, enabling the system to universally extract question groups for any topic from unannotated conversation histories. The model learns general conversational patterns (problem→question→answer states) that can be applied across different domains and topics without requiring topic-specific annotations or manual template generation for each new topic.
Solution Approach 2:
The system automatically learns the state transition model directly from raw conversation history data without requiring manual annotation or preprocessing. The model self-adapts to the conversational patterns in the data and can independently extract question groups for arbitrary topics, eliminating the need for human annotators to create topic-specific annotated datasets.
2Ease of manufacture
If conventional extraction methods are used, then the extraction process is simple, but the extracted question groups lack accuracy and relevance to arbitrary topics
Solution Approach 1:
The system performs preliminary learning of the state transition model from conversation history data before the actual question group extraction. This pre-training phase enables the model to understand the structural patterns of conversations (problem statements, questions, and answers) in advance, so that when extraction is performed, it can accurately identify and extract relevant question groups for arbitrary topics with high precision.
3Reliability
If manual template generation is used in slot fill-type dialogue systems, then the system can achieve accurate slot filling, but it takes significant time for administrators to generate templates and conversation flows
Solution Approach 1:
The system automatically generates question groups that can serve as templates for slot fill-type dialogue systems by processing conversation histories through the learned state transition model. This self-service approach eliminates the need for administrators to manually create templates while maintaining the quality and accuracy needed for reliable slot filling, as the extracted question groups preserve the problem-question-answer structure essential for accurate information extraction.
Data Source
AI summary
An addition unit 11, with regard to data indicating a conversation history including one or more sets of sentences formed from a problem sentence being a sentence indicating one problem, a question sentence being a sentence indicating a question for the one problem, and an answer sentence being a sentence indicating an answer to the question, adds a label indicating a problem state to the problem sentence within the data, a label indicating a question state to the question sentence within the data, and a label indicating an answer state to the answer sentence within the data. An extraction unit 12 extracts, from the data, a set of sentences with which the states indicated by the labels have been associated according to a state transition model that is a model configured from the one problem state, question state, and answer state, and that represents a transition of the states.