Hybrid Document Retrieval With Multi-Model Reranking for Natural Summaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document retrieval methods, such as text matching and deep learning-based approaches, face limitations in performance, context dependency, and require continuous model advancement, while conventional summarization technologies are biased towards extractive summarization, lacking naturalness.
Innovation Solution
A method and apparatus that combines text matching and deep learning-based retrieval to identify relevant documents and generate abstractive or extractive summaries by using bi-encoders, cross-encoders, and BM25F scoring, enhancing document retrieval and summarization efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If text matching-based retrieval is used, then retrieval speed is improved, but retrieval accuracy deteriorates due to dependency on exact word matches
Solution Approach 1:
The patent combines text matching-based retrieval with deep learning-based retrieval into a hybrid system. The text matching component (using BM25, phrase matching, shingle matching) provides fast initial retrieval, while the deep learning component (using bi-encoder and cross-encoder models) refines the results to improve accuracy. This merging allows the system to leverage the speed of traditional methods while achieving the precision of modern AI approaches.
2Measurement precision
If deep learning-based retrieval is used, then retrieval accuracy is improved, but system complexity and resource requirements worsen
Solution Approach 1:
The patent segments the retrieval system into distinct modules: text matching module, bi-encoder deep learning module, cross-encoder deep learning module, and hybrid integration module. Each module performs a specific function at increasing levels of complexity. This segmentation allows the system to progressively refine results without requiring all components to operate simultaneously at full capacity, managing complexity through modular architecture.
Solution Approach 2:
The system applies deep learning methods partially - using bi-encoder for initial scoring and cross-encoder only for top-k re-ranking. This partial application of complex deep learning techniques reduces computational overhead while still achieving significant accuracy improvements compared to full deep learning deployment.
3Productivity
If conventional summarization methods are used, then processing speed is improved, but summary quality deteriorates due to extractive bias
Solution Approach 1:
The patent implements a dynamic summarization system that can adapt between extractive and abstractive modes based on input characteristics and user needs. The system uses extractive summarization for speed when appropriate, and abstractive summarization for quality when needed, making the summarization approach flexible rather than fixed. This dynamic switching resolves the contradiction by allowing both modes to serve different scenarios.
Data Source
AI summary
A method of retrieving a document according to an embodiment of the present application includes: acquiring a user retrieval query; calculating a user inquiry vector in a unit of sentence from the user retrieval query and acquiring a first document candidate group based on similarity between the calculated user inquiry vector and an embedding vector of a document stored in a retrieval database; acquiring a second document candidate group based on similarity between a text included in the user retrieval query and a text of the document stored in the retrieval database; and determining a summarization target document based on the first document candidate group and the second document candidate group.