Document Question Answering with Page Retrieval and Sub-Question Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document question answering methods lack robustness and consistency in handling multimodal information, leading to insufficient planning for sub-questions and susceptibility to errors during inference.
Innovation Solution
A large model-based question answering method that employs a tree inference-based approach, using pre-trained visual language models to segment questions into sub-questions and recursively process document pages, enhancing robustness and accuracy by identifying candidate pages and generating answers or sub-questions as needed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing document question answering methods are used, then the system can process questions, but the robustness and consistency in handling multimodal information deteriorates
Solution Approach 1:
The patent segments the question answering process into distinct modules: question segmentation divides complex questions into sub-questions, candidate page identification isolates relevant document sections, and answer generation handles response creation. This modular segmentation improves robustness by making each component independent and testable, while maintaining consistency across multimodal information processing.
Solution Approach 2:
The patent performs preliminary actions before the main answer generation: questions are pre-segmented into sub-questions, candidate pages are pre-identified and ranked, and relevant information is pre-extracted. This preliminary processing enhances robustness by preparing structured data in advance, reducing errors during the actual answer generation phase.
2Reliability
If existing question answering methods are used, then processing can be completed, but susceptibility to errors during inference worsens
Solution Approach 1:
The patent implements feedback mechanisms where the system evaluates the quality of generated answers and adjusts its processing accordingly. Candidate pages are ranked based on relevance feedback, and the system can iterate through sub-questions if initial answers are insufficient. This feedback loop reduces error susceptibility while maintaining reasonable inference efficiency through targeted reprocessing.
Solution Approach 2:
The patent prepares cushioning measures in advance by generating multiple candidate pages with varying relevance scores and preparing alternative answer pathways through question segmentation. If one inference path encounters errors, the system has pre-prepared alternative routes through other candidate pages or sub-questions, reducing overall error susceptibility.
3Measurement precision
If comprehensive document analysis is performed, then answer accuracy improves, but resource consumption increases
Solution Approach 1:
The patent applies local quality by focusing computational resources on specific relevant portions of the document rather than uniformly processing the entire document. Candidate page identification ranks and selects only the most relevant pages, and answer generation focuses on extracting information from these selected portions. This localized processing maintains high answer accuracy while significantly reducing resource consumption compared to comprehensive full-document analysis.
Solution Approach 2:
The patent performs partial action by generating a limited number of top-ranked candidate pages (e.g., top 3-5) rather than analyzing all possible document pages. This partial processing provides sufficient accuracy for most questions while avoiding the excessive resource consumption that would result from exhaustive analysis of every page in the document.
4Adaptability or versatility
If question segmentation into sub-questions is implemented, then planning for sub-questions improves, but operational steps increase
Solution Approach 1:
The patent applies the nested doll principle by embedding sub-questions within the main question framework. Each sub-question is a smaller version of the original question structure, processed through the same question answering pipeline. This nesting approach improves adaptability by handling complex questions through recursive application of the same simple rules, while the modular nature prevents excessive overall process complexity.
Data Source
AI summary
A method includes: obtaining a document comprising at least one page for question answering; determining a first vector corresponding to each of the at least one page; determining a second vector corresponding to a target question text to be answered; performing the following first operations: determining, based on the second vector and the first vector corresponding to each of the at least one page, a first similarity between the target question text and each of the at least one page; determining, based on the first similarities, at least one candidate page with the highest similarity to the target question text among the at least one page; and generating, based on the at least one candidate page and the target question text, a first identifier and first content, or second identifier and second content, using a large language model.