Neural Retrieval System for Multi-Hop Question Answering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing open domain question answering systems fail to effectively handle multi-hop question answering tasks, especially entity-centric questions, due to limitations in capturing lexical information and inter-document relationships, leading to incomplete evidence retrieval and poor performance in scenarios requiring sequential reasoning paths of arbitrary length.
Innovation Solution
A neural retrieval system that learns to select sequences of evidence paragraphs as reasoning paths, sequentially retrieving each paragraph conditioned on previously retrieved evidence, using a multi-head transformer neural network to determine selection probabilities and estimate answer likelihood, allowing for the identification of a minimal set of required evidence paragraphs from a large knowledge base.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If independent document retrieval is used, then retrieval speed is improved, but retrieval quality for multi-hop questions deteriorates
Solution Approach 1:
The patent implements dynamic retrieval where the retrieval process adapts based on previously retrieved documents. The system sequentially retrieves documents, using each new document to update the query and retrieve the next document, creating a dynamic multi-hop retrieval path rather than static independent retrieval
Solution Approach 2:
The patent employs feedback mechanisms where the output of each retrieval step (retrieved documents and extracted entities) feeds into the next retrieval step. The system uses the retrieved documents to refine subsequent queries, creating a feedback loop that improves retrieval quality for multi-hop questions while maintaining reasonable speed
2Measurement precision
If entity linking is used to guide retrieval, then retrieval precision for entity-centric questions is improved, but ability to capture inter-document relationships deteriorates
Solution Approach 1:
The patent segments the retrieval process into multiple independent hops, where each hop retrieves documents based on specific entities or queries. This segmentation allows precise entity-focused retrieval at each step while the sequence of hops collectively captures inter-document relationships that would be lost in a single retrieval pass
3Device complexity
If fixed-length reasoning paths are used, then computational complexity is reduced, but ability to handle arbitrary length reasoning deteriorates
Solution Approach 1:
The patent implements dynamic reasoning path length where the number of retrieval hops is not fixed but adapts based on the complexity of the question and the information needed. The system can retrieve one document for simple questions or multiple documents for complex multi-hop questions, providing versatility while managing computational complexity through on-demand retrieval
Data Source
AI summary
A question answering system includes: a neural network tokenizer module configured to determine a token representation of a question to be answered and token representations of candidate paragraphs of a present reasoning path for the question, respectively; a neural network module configured to: transform the token representation of the question and the token representations of the candidate paragraphs of the present reasoning path into vector representations; and append a first variable to the vector representations to produce a second variable; a search module configured to: select the candidate paragraphs from a knowledge database to extend a present reasoning path based on lexical relevance of the candidate paragraphs to the question; and selectively add ones of the candidate paragraphs to the present reasoning path; and an answer inference network module configured to selectively determine an answer to the question based on multiple different portions of the present reasoning path.