Abstract:
Named Entity Recognition (NER) is a pivotal technique for extracting structured information from unstructured or semi-structured data. This project focuses on applying deep learning (DL) tech niques to NER for ingredient phrases, leveraging a manually annotated dataset of approximately 10,000 rows. Multiple approaches were explored, including statistical methods, fine-tuning deep learning-based language models, and few-shot prompting with large language models (LLMs). Mod els such as spaCy-transformer, DistilBERT, BERT, and DistilRoBERTa were initially implemented as described in reference literature, achieving F1 scores in the 90s. Through custom optimizations, the performance of these models improved to 93–94%, yet still fell short of benchmark results. To address these limitations, a second dataset comprising 10,000 rows was integrated after resolving inconsistencies, creating a unified dataset of 18,000 rows. However, combining the datasets led to a performance drop, revealing challenges related to dataset quality and model generalization. Further evaluations, including cross-validation and bucket analysis, were conducted with BERT, the best performing model, to gain deeper insights into its strengths and limitations. This report presents a comprehensive overview of the methods, challenges, and results, emphasizing the need for continuous refinement and highlighting opportunities for enhancing model performance.
