Adversarial Domain Adaptation for Duplicate Question Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large companies face inefficiencies in processing and storage due to duplicate questions from users, which require significant resources and are challenging to identify using machine learning models, especially when training data from a specific domain is insufficient.
Innovation Solution
A machine learning technique using adversarial learning to train a model with both specific domain and general domain data, identifying duplicate questions by distinguishing features between domains and reducing the importance of domain-specific terminology.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning models are used to identify duplicate questions, then question processing efficiency can be improved, but the models struggle to accurately identify duplicates when domain-specific training data is insufficient
Solution Approach 1:
The model architecture is segmented into domain-specific feature extraction components and general linguistic feature extraction components. This segmentation allows the system to separately process domain-specific terminology and general question structures, enabling effective duplicate detection even when domain-specific training data is limited by leveraging general linguistic patterns.
Solution Approach 2:
The patent employs a composite approach by combining domain-specific pre-trained language models with general-purpose duplicate detection models. This composite model integrates the strengths of both specialized domain knowledge and general linguistic understanding, achieving accurate duplicate identification without requiring extensive domain-specific training data.
2Reliability
If all user questions are stored in the database, then complete question-answer records are available for reference, but significant storage space is consumed and resources are wasted on answering previously answered queries
Solution Approach 1:
Before storing new questions in the database, the system performs preliminary duplicate detection by comparing the new question against existing questions using the trained model. If a duplicate is identified, the question is not stored and the existing answer is directly provided to the user. This preliminary action prevents redundant storage and reduces database size while maintaining answer completeness.
3Measurement precision
If domain-specific feature extraction is emphasized, then domain accuracy is improved, but the model becomes sensitive to domain-specific terminology variations that prevent duplicate detection
Solution Approach 1:
The model applies local quality by assigning different levels of importance to different features based on their context. Domain-specific features are weighted appropriately for domain accuracy, while general linguistic features are weighted for capturing semantic equivalence across different wordings. This differential weighting allows the model to balance domain specificity with general adaptability.
Solution Approach 2:
The patent implements parameter changes by dynamically adjusting the weights of domain-specific versus general features based on the input question characteristics. When domain-specific terminology is detected, the model increases domain feature weights; when general linguistic patterns are prominent, it increases general feature weights. This adaptive parameter adjustment enables the model to maintain both domain accuracy and generalization capability.
Data Source
AI summary
Detect duplicated questions using reverse gradient adversarial domain adaptation includes applying a general network to multiple general question pairs to obtain a first set of losses. A target domain network is applied to multiple domain specific network pairs to obtain a second set of losses. Further, a domain distinguishing network is applied to a set of domain specific questions and a set of general questions to obtain a third set of losses. A set of accumulated gradients is calculated from the first set of losses, the second set of losses, and the third set of losses. Multiple features are updated according to the set of accumulated gradients to train the target domain network.