Unified Vector Space Entity Linking Across Knowledge Bases
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current entity linking methods require significant time and computing resources to determine the appropriate knowledge graph for linking mentions across multiple domains, as they treat the problem as zero-shot learning, making it challenging to link mentions across different knowledge graphs efficiently.
Innovation Solution
A method for training a model to map entities from multiple knowledge bases into a vector space, allowing for direct processing and linking of mentions across multiple knowledge graphs by evaluating a loss function that measures similarity and dissimilarity between representations, enabling efficient selection of the best candidate entity from multiple knowledge bases in a single step.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If zero-shot learning is used to link mentions across multiple knowledge graphs, then the model can handle multiple domains, but it requires significant time and computing resources to determine the appropriate knowledge graph
Solution Approach 1:
The patent applies preliminary action by pre-training the model on multiple knowledge graphs during the training phase. The model learns to map entities from different knowledge graphs into a unified vector space beforehand, so that during inference, no additional time is needed to determine which knowledge graph to use. The loss function incorporates dissimilarity measures between entities from different knowledge graphs to enforce distinct representations during this preliminary learning stage.
Solution Approach 2:
The patent implements universality by creating a single model that can handle multiple knowledge graphs simultaneously. The model uses a unified vector space that accommodates entities from different knowledge graphs, allowing it to perform entity linking across multiple domains without requiring separate models or manual selection of knowledge graphs, thus eliminating the time cost of domain determination.
2Measurement precision
If multiple knowledge bases are considered for entity linking, then the model can find the best match, but the number of possible entities increases to millions
Solution Approach 1:
The patent applies merging by combining entities from multiple knowledge graphs into a single unified vector space. Instead of treating each knowledge graph separately with millions of candidate entities, the model integrates them into one cohesive representation space, reducing the effective search space while maintaining the ability to distinguish entities from different sources through their learned representations.
Solution Approach 2:
The patent uses parameter changes by transforming entities from different knowledge graphs into a unified vector space representation. This change in representation parameters allows the model to handle millions of candidate entities efficiently by operating in a continuous vector space where similarity can be computed rapidly using dot products or cosine similarity, rather than comparing discrete entity attributes.
3Measurement precision
If manual selection of knowledge graph is required, then the model can focus on specific domain, but it requires user expertise and additional processing steps
Solution Approach 1:
The patent implements self-service by enabling the model to automatically determine the appropriate knowledge graph for entity linking without user intervention. During training, the model learns to associate different knowledge graph entities with their respective domains through the dissimilarity loss function. At inference time, the model autonomously selects the most appropriate knowledge graph based on the input mention, eliminating the need for user expertise in knowledge graph selection.
Data Source
AI summary
A device and method for training a model for linking a mention in textual context to an entity across knowledge bases. I the method, depending on training data, training the model for mapping an entity of a first knowledge base to its first representation in a vector space, for mapping an entity of a second knowledge base to its second representation in the vector space, for mapping the mention to a third representation in the vector space. The training data includes a set of pairs in which each pair includes a mention in a textual context and its corresponding reference entity in either the first knowledge base or the second knowledge base. Training the model includes evaluating a loss function.