Abstract:
"This thesis tackles the issue of optimizing Named Entity Recognition (NER) by bridging the gap between sequential and hierarchical information. Many previous studies have focused solely on either sequential or hierarchical relationships, missing the potential benefits of combining both. To address this, the study introduces an innovative ensemble strategy that amalgamates Hierarchical Attention Networks (HAN) and Recurrent Neural Networks (RNN) to enhance NER performance.
The HAN model is employed for its ability to gather information at varying levels of precision through its hierarchical structure, while the RNN is utilized to capture sequential dependencies within the text. The objective is to improve NER systems' accuracy, resilience, and adaptability by synergizing these two models. The ensemble approach ensures the effective utilization of both sequential and hierarchical data structures, resulting in a more comprehensive representation of named entities within the text.
A comprehensive set of experiments is conducted using an unexplored dataset specifically chosen for assessing the efficacy of HAN and the ensemble model. The results are compared to several state-of-the-art NER techniques. The RNN model achieves a test accuracy of 97.5%, the HAN model scores 97.48%, and the ensemble model surpasses both with an impressive test accuracy of 97.56%. These findings underscore the effectiveness of the ensemble method, indicating its potential to substantially enhance named entity recognition tasks.
In conclusion, the thesis introduces an ensemble model that effectively combines HAN and RNN, leveraging both sequential and hierarchical information to enhance NER performance. The experimental results showcase the model's superiority and suggest its significant potential for advancing NER systems, offering greater precision, robustness, and generalizability in identifying named entities in text
"
