[Paper Review] Machine Learning for Food Review and Recommendation

Tan Khang Le, Siu Cheung Hui|arXiv (Cornell University)|Jan 15, 2022

Sentiment Analysis and Opinion Mining4 citations

TL;DR

This paper proposes a web-based food review system, NTU Food Hunter, integrating BERT and LSTM for sentiment analysis, a POS-based adjective-noun pair extraction method for automatic review tag generation, and a RankNet model for re-ranking in a Solr-based retrieval system. The system significantly improves retrieval performance, achieving 100% MAP@1 and 100% MRR—46% better than BM25—by combining semantic and textual features for more accurate food review search.

ABSTRACT

Food reviews and recommendations have always been important for online food service websites. However, reviewing and recommending food is not simple as it is likely to be overwhelmed by disparate contexts and meanings. In this paper, we use different deep learning approaches to address the problems of sentiment analysis, automatic review tag generation, and retrieval of food reviews. We propose to develop a web-based food review system at Nanyang Technological University (NTU) named NTU Food Hunter, which incorporates different deep learning approaches that help users with food selection. First, we implement the BERT and LSTM deep learning models into the system for sentiment analysis of food reviews. Then, we develop a Part-of-Speech (POS) algorithm to automatically identify and extract adjective-noun pairs from the review content for review tag generation based on POS tagging and dependency parsing. Finally, we also train a RankNet model for the re-ranking of the retrieval results to improve the accuracy in our Solr-based food reviews search system. The experimental results show that our proposed deep learning approaches are promising for the applications of real-world problems.

Motivation & Objective

• To address information overload in food review platforms by enabling intelligent, automated analysis of unstructured food reviews.
• To improve food recommendation accuracy by applying deep learning to sentiment analysis, tag generation, and retrieval ranking.
• To develop a real-world web-based system (NTU Food Hunter) that integrates multiple NLP techniques for practical food review management.
• To enhance retrieval precision by combining semantic understanding (via BERT) with learning-to-rank models (RankNet) in a Solr-based search pipeline.

Proposed method

• Fine-tuned BERT and LSTM models for food review sentiment classification, leveraging pre-trained embeddings and fine-tuning on NTU canteen review data.
• Developed a POS-based algorithm using dependency parsing and part-of-speech tagging to extract adjective-noun pairs for automatic review tag generation.
• Constructed a feature dataset by defining query-document relevance features and labeling pairs as relevant (1) or irrelevant (0) for training.
• Implemented a RankNet model to re-rank Solr retrieval results based on learned feature values, improving ranking accuracy beyond lexical matching.
• Integrated the retrieval pipeline with Solr, using BERT embeddings and feature matching to compute relevance scores before RankNet re-ranking.
• Used a payload-based query interface to transmit features and embeddings to Solr, enabling end-to-end retrieval with semantic-aware re-ranking.

Experimental results

Research questions

• RQ1 Can BERT and LSTM models effectively classify sentiment in food reviews from a university canteen setting?
• RQ2 Can POS-based adjective-noun pair extraction reliably generate meaningful, informative review tags from unstructured food reviews?
• RQ3 Can a RankNet-based learning-to-rank model improve retrieval accuracy beyond traditional methods like tf-idf and BM25 in a food review search system?
• RQ4 To what extent does combining semantic features (BERT) and structured features (adjective-noun pairs, category) enhance retrieval performance?

Key findings

• The proposed system achieved 100% Mean Average Precision at position 1 (MAP@1), significantly outperforming tf-idf (60.00%) and BM25 (60.00%).
• The system achieved a Mean Reciprocal Rank (MRR) of 1.0000, indicating perfect ranking of relevant results in the top position, surpassing BM25 (0.6795) and tf-idf (0.7021).
• The RankNet-based re-ranking improved retrieval performance by 46% compared to BM25 and 29% compared to tf-idf across all evaluation metrics.
• The integration of BERT embeddings and semantic feature matching in Solr significantly enhanced relevance scoring beyond lexical similarity alone.
• The adjective-noun pair extraction method successfully generated meaningful tags that reflect key food attributes, improving review summarization and search discoverability.
• The end-to-end pipeline, combining sentiment analysis, tag generation, and semantic re-ranking, demonstrated strong real-world applicability for food review systems.