Related Experiment Video
Updated: Jul 21, 2026
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
990
Dual-decoder neural architecture with uncertainty-based task weighting for named entity recognition in injection
Shuxian Li1, Yalin Wang1, Jingyu Guo1
1School of Automation, Central South University, Lushan South Road, Changsha, Hunan, 410083, China; National Engineering Research Centre of Advanced Energy Storage Materials, Changsha, Hunan, China.
Summary
This study introduces a novel dual-decoder neural network for Named Entity Recognition (NER) in injection molding. The method significantly improves the accuracy of identifying critical process information, enhancing intelligent defect diagnosis.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Natural Language Processing
Background:
- Named Entity Recognition (NER) is crucial for structuring unstructured text data.
- Existing NER methods face challenges with long entities and fragmented spans in specialized domains like injection molding.
- Accurate extraction of injection molding process data is vital for intelligent defect diagnosis and optimization.
Purpose of the Study:
- To propose a novel dual-decoder neural architecture for enhanced NER in the injection molding domain.
- To address challenges in recognizing long entities, fragmented spans, and integrating character-level with entity-level information.
- To improve the accuracy and robustness of extracting critical information (causes, defects, solutions) from injection molding texts.
Main Methods:
- A dual-decoder neural architecture utilizing a shared pre-trained model for contextual embeddings.
- Parallel decoding via a Conditional Random Field (CRF) with domain constraints and a Global Pointer (GP) within a multi-task framework.
- An uncertainty-aware dynamic task weighting mechanism to balance contributions between decoders during training.
Main Results:
- The proposed method achieved an 86.09% overall F1 score on a self-constructed injection molding corpus.
- Demonstrated superior performance compared to existing advanced NER models in recognizing critical entities.
- Validation on general Chinese datasets confirmed the method's generalizability for texts with similar long-entity characteristics.
Conclusions:
- The dual-decoder architecture effectively models local dependencies and global boundary semantics for improved NER accuracy.
- Uncertainty-based dynamic task weighting enhances inter-task synergy and model robustness in complex recognition tasks.
- The proposed approach offers a significant advancement for intelligent defect diagnosis and optimization in the injection molding industry.
