Retrieved Support-Document Fusion for Traceable Answer Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current question answering systems face challenges in processing multi-source multi-answer scenarios, lack summarizing capabilities, and generate uncontrollable answers with hallucinated content, poor traceability, and inaccurate suggestions.
Innovation Solution
An answer generation method that integrates information from multiple support documents, using a fusion distribution technology to generate controlled and reliable answers by encoding, aggregating, and decoding the information, and presenting it in a multi-source format with source information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a pre-trained model is used to generate answers based on massive data, then the model can provide comprehensive knowledge coverage, but the training consumes excessive time and the model may generate hallucinated content with poor traceability
Solution Approach 1:
The patent segments the answer generation process into distinct components: an encoder processes individual support documents to extract information, a decoder generates answers based on encoded representations, and a fusion mechanism integrates multiple encoded documents. This segmentation allows the system to avoid training a single ultra-large model while achieving comprehensive knowledge coverage through multiple specialized components working together.
Solution Approach 2:
The patent introduces an intermediate encoding layer that acts as a mediator between support documents and final answers. The encoder transforms documents into structured representations, and the decoder uses these representations to generate answers. This intermediary structure enables traceability by maintaining connections between source documents and generated content without requiring an ultra-large pre-trained model.
2Ease of operation
If a pre-trained model generates answers by copying from training data, then the model can provide fluent and comprehensive answers, but the model generates hallucinated content and provides inaccurate suggestions
Solution Approach 1:
The patent implements feedback mechanisms where the encoder-decoder architecture processes multiple support documents and integrates their information systematically. The fusion mechanism provides feedback by comparing and reconciling information from different sources, ensuring that generated answers are grounded in actual support documents rather than hallucinated content, while maintaining fluency through the decoder's language generation capabilities.
Solution Approach 2:
The patent creates a composite answer generation system that combines multiple encoded support documents rather than relying on a single pre-trained model's memorized knowledge. The fusion mechanism integrates information from multiple sources, creating a composite representation that ensures accuracy by grounding answers in actual source materials while maintaining fluency through the decoder component.
3Productivity
If an extractive question answering manner is used, then the system can process questions efficiently, but the system cannot handle multi-source multi-answer scenarios and lacks summarizing capability
Solution Approach 1:
The patent transitions from a static extractive approach to a dynamic generative approach. The encoder-decoder architecture dynamically processes multiple support documents, adapting to different question types and source combinations. The system can selectively extract, summarize, or generate answers based on the input documents, providing versatility for multi-source scenarios while maintaining efficiency through the encoder's optimized processing of individual documents.
Solution Approach 2:
The patent creates a universal answer generation system that can handle multiple question types and source configurations through a single encoder-decoder framework. The system universally processes any combination of support documents, whether for extraction, summarization, or generation, making it adaptable to multi-source multi-answer scenarios while maintaining processing efficiency through the standardized encoding approach.
4Quantity of substance
If knowledge is implicitly stored in model parameters, then the model can provide comprehensive knowledge, but the amount of stored knowledge depends on model parameter quantity requiring ultra-large models
Solution Approach 1:
The patent extracts knowledge from multiple external support documents rather than storing it implicitly in model parameters. The encoder processes and extracts relevant information from each document, and the fusion mechanism integrates these extracted knowledge elements. This approach achieves comprehensive knowledge coverage without requiring an ultra-large model, as knowledge is pulled from external sources during inference rather than embedded in parameters.
Data Source
AI summary
This application relates to the field of artificial intelligence technologies, and in particular, to an answer generation method and apparatus, and a storage medium. The method includes: retrieving, based on an input target question, k support documents related to the target question, where k is a positive integer greater than 1; concatenating, for each of the k support documents, the target question and the support document, to obtain a combination pair; performing information aggregation on the k combination pairs, to obtain target aggregation information, where the information aggregation includes information encoding and exchanging; and generating, based on the target aggregation information, a target answer corresponding to the target question.