Weakly Supervised NER Training via Iterative Rule-Model Bootstrapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Supervised training of named-entity recognition (NER) models requires significant manual annotation, which is challenging, especially in low-resource languages and domains without available knowledge bases, and existing distantly supervised methods rely on existing knowledge bases that may not be available in all fields.
Innovation Solution
A weakly supervised mechanism using symbolic seeding rules to label data, iteratively refining the NER model through bootstrapping with neural representations, allowing the model to identify new entities and generate training data without extensive human effort, leveraging both logical rules and neural representations for entity recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If supervised training is used to train NER models, then model performance and reliability are improved, but significant manual annotation effort and time are required
Solution Approach 1:
The system performs self-service by automatically generating training data through rule-based entity recognition. The server autonomously creates labeled training examples by applying extraction rules to unstructured text, eliminating the need for manual annotation while maintaining model training capabilities.
Solution Approach 2:
Extraction rules serve as an intermediary between unstructured text and training data labels. These rules act as a mediator that automatically transforms raw text into structured training examples, bridging the gap without requiring direct human annotation effort.
2Productivity
If distantly supervised methods are used to generate training data, then scaling without human effort is improved, but dependency on existing knowledge bases is worsened
Solution Approach 1:
The system segments the training data generation process into rule-based entity extraction and model training components. By dividing the workflow, it can generate domain-specific training data through customizable extraction rules without relying on pre-existing knowledge bases, enabling flexibility across different domains.
Solution Approach 2:
The extraction rules can be dynamically adjusted and customized for different domains by changing rule parameters and patterns. This allows the system to adapt to various domains (e.g., biomedical, technical) without being constrained by existing knowledge bases, maintaining both scalability and domain versatility.
3Measurement precision
If manual annotation is performed for low-resource languages, then model accuracy is improved, but the difficulty and cost of obtaining domain-expert annotation are worsened
Solution Approach 1:
The system enables self-service training data generation for low-resource languages by automatically applying extraction rules to create labeled examples. This eliminates the need to acquire expensive domain-expert annotations while maintaining sufficient training data quality for accurate entity recognition.
Solution Approach 2:
The system uses inexpensive rule-based methods to generate training data that can be quickly created and updated. Rather than investing in costly manual annotation by domain experts, the system employs disposable, easily generated rule-based labels that suffice for training purposes.
Data Source
AI summary
Systems and methods for weakly-supervised training a machine-learning model to perform named-entity recognition. All possible entity candidates and all possible rule candidates are automatically identified in an input data set of unlabeled text. An initial training of the machine-learning model is performed using labels assigned to entity candidates by a set of seeding rules as a first set of training data. The trained machine-learning model is then applied to the unlabeled text and a subset of rules from the rule candidates is identified that produces labels that most accurately match the labels assigned by the trained machine-learning model. The machine-learning model is then retrained using the labels assigned by the identified subset of rules as the second set of training data. This process is iteratively repeated to further refine and improve the performance of the machine-learning model for named-entity recognition.