Loading...

Named Entity Recognition in Persian Language using Deep Learning

Abed Azad, Parham | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57411 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Beigy, Hamid
  7. Abstract:
  8. Named Entity Recognition (NER) is one of the most important tasks in natural language processing. Improving performance in this task benefits other tasks. However, with the rapid expansion of text volume and diversity, significant challenges have emerged in this area. This thesis proposes a model to address the problem of multi-domain Named Entity Recognition. Traditional approaches used for multi-domain NER often have considerable limitations. Training a single model for diverse domains does not yield optimal results, while using multiple large models can lead to resource constraints. Additionally, approaches based on generative models for non-English languages, such as Persian, which is addressed in this thesis, do not produce desirable results for such tasks without retraining and fine-tuning these models is costly. Therefore, this thesis introduces a new approach consisting of a core model with multiple domain-specific parameter sets trained for each domain. The thesis uses adapters along with the integration of additional layers to add parameters that can be trained for specific domains. Experimental results on multiple NER datasets show that the proposed model achieves outstanding results across all domains while requiring only one model instance for training and storage, very close to the method of training separate models for each domain. In summary, our method can reduce the required volume for storing models by the number of domains. For example, if we have 5 domains, this model performs like storing 5 separate models, effectively reducing space requirements by 80\%. Moreover, compared to the conventional method for NER, an average improvement of 5 to 7 percent is observed. Finally, we have introduced a document-based domain detection pipeline for scenarios with unknown text domains. When we have access to eight input cases, this method can detect domains with high accuracy
  9. Keywords:
  10. Deep Learning ; Machine Learning ; Named Entity Recognition ; Natural Language Processing ; Multi-Domain Environment

 Digital Object List

 Bookmark