Dual-decoder neural architecture with uncertainty-based task weighting for named entity recognition in injection molding defect diagnosis.

Abstract

As a core technique for extracting and categorizing information from unstructured text, Named Entity Recognition (NER) plays a pivotal role in converting injection molding process data into structured knowledge, thereby facilitating intelligent defect diagnosis and optimization. Addressing the challenges of capturing long entities, mitigating fragmented span recognition, and improving the integration of character-level with entity-level information in the injection molding domain, this paper proposes a dual-decoder neural architecture incorporating uncertainty-based dynamic task weighting. The framework first uses a shared pre-trained model to generate contextual embeddings, followed by parallel decoding through a Conditional Random Field with domain constraints and a Global Pointer under a multi-task framework. This complementary design simultaneously models local label dependencies and global boundary semantics, enhancing recognition accuracy for both short and long entities in injection molding texts. Besides, an uncertainty-aware weighting mechanism dynamically adjusts the loss contributions between the two decoders during training, further enhancing inter-task synergy and improving model robustness in complex entity recognition. Extensive experiments on a self-constructed injection molding corpus demonstrate that the proposed method achieves an 86.09 % overall F₁ score in recognizing critical entities, including causes, defects, and solutions, outperforming a series of existing advanced NER models. Additional validation on the general Chinese datasets with similar long-entity distribution characteristics further confirms the method's generalizability and effectiveness.