Hierarchical Entity Classification via End-to-End Recurrent Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional fine-grained entity recognition methods face challenges due to the complexity of learning with a large set of fine entity labels and the presence of noise and redundant labels in datasets, often relying on distant supervision and knowledge bases, which complicates hierarchical classification.
Innovation Solution
A method and apparatus for hierarchical entity classification using an end-to-end trained decoding recurrent neural network that computes ontological classes levelwise by aligning contextual and mention representations, employing bidirectional recurrent neural networks and attention mechanisms to refine context and mention vectors for accurate classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large set of fine entity labels is used for fine-grained entity recognition, then the classification granularity is improved, but the learning complexity and state space complexity increase drastically
Solution Approach 1:
The patent segments the entity typing task into hierarchical levels (e.g., coarse-grained types first, then fine-grained subtypes). This segmentation reduces the state space at each decoding step by only considering relevant subtypes of the parent type, rather than all possible types simultaneously. The hierarchical structure allows the model to progressively narrow down the classification space.
Solution Approach 2:
The patent introduces a hierarchical dimension to the classification problem, transforming a flat multi-class classification into a multi-level hierarchical classification. This adds a structural dimension that organizes the label space, allowing the model to navigate through types systematically from general to specific, reducing the effective search space at each step.
2Adaptability or versatility
If knowledge bases are used to obtain candidate types for entity mentions, then the coverage of entity types is improved, but noise and redundant labels are introduced
Solution Approach 1:
The patent performs preliminary filtering of candidate types from knowledge bases by leveraging the hierarchical structure. Instead of treating all candidate types equally, the model first identifies the appropriate parent type and then selects from relevant subtypes. This preliminary action of hierarchical filtering removes unrelated and redundant labels before the final classification decision.
Solution Approach 2:
The hierarchical type structure acts as an intermediary between the knowledge base candidate types and the final classification. The parent-child relationships in the hierarchy mediate the selection process, ensuring that only contextually relevant subtypes are considered for a given mention, thereby filtering out noise while maintaining comprehensive coverage.
3Ease of manufacture
If a pipeline approach with separately trained segmentation and entity typing models is used, then the training complexity is reduced, but the overall performance and accuracy deteriorate
Solution Approach 1:
The patent merges the entity segmentation and entity typing tasks into a unified end-to-end trainable model. The model jointly learns both segmentation boundaries and type classification through shared representations and a unified loss function, allowing information to flow bidirectionally between the tasks and improving overall accuracy compared to separate pipeline models.
Solution Approach 2:
The patent creates a multi-functional model that simultaneously performs entity segmentation, mention extraction, and hierarchical type classification. This universal model shares underlying representations across all tasks, improving efficiency and accuracy by leveraging common patterns in the data for multiple purposes rather than requiring separate specialized models.
Data Source
AI summary
A method for performing hierarchical entity classification of an entity mention within a context, wherein ontological classes are computed for the entity mention levelwise using a contextual representation of the context and a state representation obtained by running an end-to-end trained decoding recurrent neural network on a mention representation of the entity mention.