Leveraging Deep Learning and Word Embeddings to Detect Dish Names in Consumer Reviews
Loading...
Date
2024-02
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
The British University in Dubai (BUiD)
Abstract
Named Entity Recognition (NER) is crucial for extracting entities from unstructured text, offering significant insights for businesses through customer review analysis. This study fills a gap in recognizing dish names from customer reviews, as existing literature mainly addresses food entity recognition in recipe datasets and lacks annotated datasets for this specific NER task. Domain adaptation and deep learning approaches like BiGRUs and CNNs remain underexplored. The research proposes a deep learning NER framework to accurately identify dish names in customer reviews with efficient computational resource use. In addition to the existing dataset, MenuNER dataset, an annotated dataset, ReviewsDB, was created from Yelp reviews for evaluation. Initial experiments revealed a notable performance drop in domain adaptation from food names in recipe datasets to dish names in reviews, with the F1-score nearly 50% lower. A comparative analysis of 53 deep learning models using various word embeddings, including Glove, Word2Vec, and Bert variants, showed that a simple architecture with a single-layer Bidirectional Gated Recurrent Unit (BiGRU) and Conditional Random Field (CRF) layer achieved the best performance, with an F1-score of 93.07% using glove-twitter-100 embeddings in the MenuNER dataset. Additionally, a two-layer BiGRU with a CNN and CRF achieved an F1-score of 82.40% on the ReviewsDB dataset. The study attributes performance differences to variability in annotation lengths and the broader range of terms in ReviewsDB. In conclusion, the proposed NER framework, leveraging pre-trained embeddings, provides a valuable tool for the food industry to analyze customer feedback and enhance customer satisfaction.
Description
Keywords
deep learning, consumer reviews, Named Entity Recognition (NER)