Virtual Knowledge Graphs for Zero-Shot Document Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional document retrieval systems struggle to represent complex knowledge in domain-specific documents, leading to inefficiencies in modeling relationships between queries and candidate documents.
Innovation Solution
The system employs a machine learning model to construct virtual knowledge graphs (VKGs) for queries and documents, using a relation encoder network and contrastive learning to generate relation vectors that encode document aspects, enabling effective indexing and matching of relevant documents.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional document retrieval systems are used, then system simplicity is maintained, but the ability to represent complex domain-specific knowledge is insufficient
Solution Approach 1:
The system segments the knowledge representation task by constructing separate virtual knowledge graphs for queries and documents. Each VKG is built by extracting entities and relations independently, then matching them through relation vectors. This segmentation allows the system to handle complex domain-specific knowledge without requiring a complete system redesign.
Solution Approach 2:
The patent introduces virtual knowledge graphs as an intermediary layer between the query and document retrieval process. The VKGs serve as mediators that transform unstructured text into structured knowledge representations, enabling more effective matching while maintaining system modularity and managing complexity.
2Measurement precision
If human-annotated datasets are used for training, then model accuracy is improved, but the cost and time for dataset preparation increase
Solution Approach 1:
The system employs self-service mechanisms where the model trains on automatically generated data from the corpus itself. The contrastive learning process uses positive samples (mentions from the same document) and negative samples (mentions from different documents) without requiring manual annotation. This self-service approach maintains retrieval accuracy while eliminating time-consuming dataset preparation.
Solution Approach 2:
The patent changes the training parameters from manually annotated labels to automatically generated contrastive samples. By modifying the training data generation process to create positive and negative samples based on document co-occurrence, the system achieves accurate retrieval performance without the time investment required for human annotation.
3Productivity
If conventional retrieval methods are used, then processing speed is maintained, but retrieval performance on domain-specific documents deteriorates
Solution Approach 1:
The system performs preliminary action by pre-construction of virtual knowledge graphs for all documents in the corpus. These VKGs are built in advance with extracted entities, relations, and relation vectors, enabling efficient query matching without complex processing during actual retrieval operations. This preliminary preparation improves retrieval performance while managing processing complexity.
Data Source
AI summary
Systems and methods for text processing are described. Embodiments of the present disclosure receive a query comprising a natural language expression; extract a plurality of mentions from the query; generate a relation vector between a pair of the plurality of mentions using a relation encoder network, wherein the relation encoder network is trained using a contrastive learning process where mention pairs from a same document are labeled as positive samples and mention pairs from different documents are labeled as negative samples; combine the plurality of mentions with the relation vector to obtain a virtual knowledge graph of the query; identify a document corresponding to the query by comparing the virtual knowledge graph of the query to a virtual knowledge graph of the document; and transmit a response to the query, wherein the response includes a reference to the document.