Parallel learning based Bi-LSTM model for Named Entity Recognition

The increasing usage of chatbots, sentiment analysis, and other text-based applications has dramatically increased the need for accurate Named Entity Recognition (NER) systems. The proper setting of entities such as people, organizations, dates, and times is crucial for these applications to function correctly. However, it is time-consuming and error-prone to manually annotate large datasets for NER purposes, which limits their practicality for large-scale deployment. This has led to the development of several research efforts on the development of automated NER solutions. However, current NER systems are not without their limitations. Although rule-based and dictionary-based approaches are poor in coverage of vocabulary and language rules, traditional encoder-based models such as BERT are computationally expensive, slow, and unfit for lightweight applications. This paper introduces a new BiLSTM based NER system with no reliance on pre-trained models such as BERT. Our model incorporates a parallel learning strategy, a feedback mechanism, and residual connections to improve the network performance and achieves an F1-score of 93% in entity recognition. Most importantly, this architecture offers a lightweight alternative to conventional encoder-based models with improved efficiency and processing speed without sacrificing accuracy.