Zero-shot Entity Linking via Neuro-symbolic Multi-task Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing zero-shot entity linking models are computationally expensive and suffer from popularity bias, struggling to accurately map entity mentions to corresponding entities in evolving knowledge bases with limited training data.
Innovation Solution
A neuro-symbolic, multi-task learning approach that combines primary entity linking with an auxiliary hierarchical entity type prediction task, leveraging symbolic information from knowledge graphs to improve model performance with less training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing zero-shot entity linking models are used, then entity mentions can be mapped to corresponding entities, but the models are computationally expensive and suffer from popularity bias
Solution Approach 1:
The patent segments the entity linking task into two distinct components: (1) entity type classification that leverages symbolic knowledge graph information to predict entity types, and (2) entity mention mapping that uses the predicted types as constraints. This segmentation allows each component to be optimized independently, reducing overall computational complexity while maintaining accuracy.
Solution Approach 2:
The patent introduces entity type prediction as an intermediary step between entity mention input and final entity mapping. This intermediary component uses symbolic information from knowledge graphs to guide the mapping process, reducing reliance on computationally intensive neural network operations and mitigating popularity bias by incorporating structured knowledge.
2Measurement precision
If existing zero-shot entity linking models are used, then entity mentions can be mapped to corresponding entities, but extensive training data is required
Solution Approach 1:
The patent performs preliminary entity type classification using symbolic information from knowledge graphs before conducting entity mention mapping. This preliminary action leverages pre-structured knowledge to constrain the search space, reducing the amount of training data needed for the subsequent mapping task since the model only needs to learn mappings within type-constrained categories.
Solution Approach 2:
The patent changes the parameter space by incorporating entity type predictions as additional constraints in the mapping process. Instead of directly mapping entity mentions to entities using only neural network parameters, the system transforms the problem by adding type-level parameters derived from symbolic knowledge, thereby reducing dependency on extensive training data.
3Adaptability or versatility
If knowledge bases evolve with new entities, then the knowledge base remains up-to-date, but existing models struggle to adapt without retraining
Solution Approach 1:
The patent creates a universal entity linking framework where the entity type classification component uses symbolic information from knowledge graphs that can accommodate new entities without retraining. The type prediction module serves multiple functions: it classifies entity types, constrains the mapping search space, and adapts to new entities through the knowledge graph structure, maintaining accuracy while enabling adaptability.
Solution Approach 2:
The entity type prediction acts as an adaptable intermediary that can incorporate new entities through the knowledge graph without requiring model retraining. When new entities are added to the knowledge base, their type information is automatically available to the type prediction component, which then guides the mapping process for these new entities, maintaining accuracy while enabling seamless adaptation.
Data Source
AI summary
Methods, systems, and computer program products for zero-shot entity linking based on symbolic information are provided herein. A computer-implemented method includes obtaining a knowledge graph comprising a set of entities and a training dataset comprising text samples for at least a subset of the entities in the knowledge graph; training a machine learning model to map an entity mention substring of a given sample of text to one corresponding entity in the set of entities, wherein the machine learning model is trained using a multi-task machine learning framework using symbolic information extracted from the knowledge graph; and mapping an entity mention substring of a new sample of text to one of the entities in the set using the trained machine learning model.