Question Answering Reasoning Layers for Specialized Domain Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Pre-trained language models exhibit weaker question-answering abilities on specialized domains such as physics research and medical knowledge compared to general-purpose domains, necessitating improved methods for handling complex questions.
Innovation Solution
A method and apparatus that splice questions with candidate answers into question-answer pairs, perform reasoning operations using recurrent neural networks with horizontal and vertical reasoning layers to determine feature combination weights, and select target candidate answers based on feature similarity, enhancing the question-answering capability on specialized domains.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pre-trained language models are used for question-answering on specialized domains, then the system can provide question-answering capability, but the accuracy and effectiveness are significantly weaker compared to general-purpose domains
Solution Approach 1:
The patent segments the question-answering process into multiple granularity levels (word level, phrase level, sentence level, passage level) and processes each level separately through dedicated reasoning layers. This segmentation allows the model to handle specialized domain content more effectively by analyzing different linguistic structures at appropriate levels of detail, thereby improving accuracy on specialized domains while maintaining general applicability.
Solution Approach 2:
The patent introduces a vertical dimension to the reasoning process by adding multiple reasoning layers that operate at different granularity levels. Instead of a single horizontal processing layer, the model now has stacked reasoning layers (first reasoning layer, second reasoning layer, etc.) that progressively refine the answer selection. This dimensional expansion enables deeper analysis of specialized domain content without sacrificing broad domain adaptability.
2Measurement precision
If multiple reasoning layers at different granularity levels are introduced to improve specialized domain question-answering, then the accuracy improves, but the model complexity increases
Solution Approach 1:
The complex reasoning process is segmented into modular reasoning layers, each handling a specific granularity level. This modular segmentation makes the overall complex system more manageable and interpretable. Each layer can be independently trained and optimized, reducing the practical complexity despite the increased number of components.
Solution Approach 2:
The reasoning layers are designed with universal functionality that can handle multiple types of reasoning tasks across different domain granularities. The same basic reasoning mechanism is applied recursively at different levels (word, phrase, sentence, passage), reducing the need for domain-specific customizations and making the complexity more systematic rather than arbitrary.
Data Source
AI summary
A method and apparatus for determining an answer to a question are provided. The method includes: splicing an acquired to-be-queried question with each candidate answer into each question-answer pair; performing reasoning operations of feature combination parameters on different granularity features of each question-answer pair at a preset number of steps in a horizontal direction based on recurrent characteristics of a recurrent neural network; determining feature combination weights of the different granularity features using multiple preset vertical reasoning layers at different reasoning focuses respectively, at each step of the reasoning operations in the horizontal direction; obtaining a candidate answer feature corresponding to each question-answer pair, respectively, through a final step of the reasoning operations; and determining a target candidate answer matching the to-be-queried question based on a feature similarity between a question feature of the to-be-queried question and each candidate answer feature.