Multi-Task Element Recognition Model for Domain Entity Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing domain element recognition systems face limitations in generalization, requiring extensive rule establishment and high manual labeling costs, especially in professional fields like medicine and finance.
Innovation Solution
A method and apparatus utilizing a pre-trained multi-task element recognition model that combines sequence labeling and entity word predicting networks, with a fusion module, to identify entity words, categories, and positions, trained using a multi-task machine learning approach, incorporating embedding, context coding, memory networks, and dictionary representation layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a pure rule system is used for domain element recognition, then the system can identify elements covered by existing rules, but the system is weak in generalization and requires extensive rule establishment effort
Solution Approach 1:
The patent replaces the mechanical rule-based system with a neural network model (BiLSTM-CRF) that automatically learns patterns from data. The model substitutes manual rule creation and maintenance with automated machine learning, where the network learns domain-specific patterns through training on labeled data, eliminating the need for extensive rule establishment while improving generalization to unseen texts
Solution Approach 2:
The patent transforms the rigid rule-based approach into a flexible parameter-based neural network model. Instead of fixed rules, the system uses learnable parameters (weights and biases) that are optimized during training. This allows the system to adapt to different domains and contexts by adjusting parameters rather than rewriting rules, resolving the contradiction between reliability and complexity
2Adaptability or versatility
If a model-based domain element recognition system is used, then generalization capability is improved, but a large amount of manual labeling data is required and labeling costs are very high
Solution Approach 1:
The patent applies pre-training on large-scale general domain data before fine-tuning on specific domain data. This preliminary action allows the model to learn general language patterns and entity recognition capabilities from abundant unlabeled or weakly labeled data, reducing the amount of expensive professional domain labeling needed for the final application
Solution Approach 2:
The patent merges general domain pre-training with domain-specific fine-tuning in a unified multi-task learning framework. The model simultaneously learns from both general language patterns and domain-specific entity patterns, allowing it to achieve good generalization with less domain-specific labeled data by leveraging transfer learning from the pre-trained general language model
3Adaptability or versatility
If extensive rule establishment is performed to improve element recognition coverage, then more elements can be identified, but the establishment requires a lot of engineer energy and time
Solution Approach 1:
The patent replaces the time-consuming manual rule establishment process with automated neural network training. The BiLSTM-CRF model automatically learns entity recognition patterns from labeled data without requiring engineers to manually create and maintain rules, dramatically reducing the time and effort needed to achieve comprehensive element recognition coverage
Solution Approach 2:
The system enables self-service element recognition by automatically learning and adapting to domain-specific patterns without continuous manual intervention. Once trained, the model can independently identify elements across different texts and domains, eliminating the need for ongoing rule maintenance and allowing the system to serve itself in recognizing elements without engineer energy input
Data Source
AI summary
Embodiments of the present disclosure provide methods and apparatus for outputting information. The method may include: obtaining a sentence to be identified; Performing word segmentation on the to be identified sentence to obtain a word sequence; Inputting a word sequence into a pre-trained multi-task element recognition model based on sequence labeling and entity word prediction, and outputting the identified entity words, entity categories and entity word positions, where the multi-task element recognition model includes a sequence labeling network for performing sequence labeling tasks and an entity word predicting network for performing entity word predicting task, and the sequence labeling network is fused with the entity word predicting network through a fusion module.