TripAdvisor Review Analysis

This repository contains our NLP project for analyzing TripAdvisor hotel reviews. We tackle two core tasks: predicting review ratings and generating hotel review text, using both traditional and deep learning methods.

---

🔍 Project Overview

• Tasks:
  • Predict review ratings from text (1–5 stars)
  • Generate synthetic reviews conditioned on ratings and keywords
• Techniques:
  • Supervised learning (classification)
  • Conditional text generation
• Dataset: 200k+ English TripAdvisor hotel reviews
• Goal: Explore both discriminative and generative NLP pipelines

---

Installation

uv venv python=3.10
source .venv/bin/activate
uv pip install -r pyproject.toml

🚀 Getting Started

Run the project using the repository CLI located at src/main.py. The CLI is configuration-driven and exposes three main commands: train, eval and infer.

Examples (run from repository root):

uv run src/main.py --config configurations/FeedForward/config.yaml train --help
# Show options for the train command

# Train a model using a specific configuration and seed (recommended runner)
uv run src/main.py --config configurations/FeedForward/config.yaml --seed 42 train

# Evaluate a trained model
uv run src/main.py --config configurations/FeedForward/config.yaml eval

# Run inference / generate text from a prompt (provide prompt and rating)
uv run src/main.py --config configurations/Transformer/config.yaml --seed 42 infer -p "clean, location, nice, pretty, comfortable" -r 5

Notes:

• The --config option is required and should point to one of the YAML files in the configurations/ folder (for example configurations/Transformer/config.yaml).
• --seed sets the random seed for reproducibility (default: 42).
• Use uv run src/main.py --config <config> <command> --help or python -m src.main --config <config> <command> --help to see command-specific options.

Example configuration files are available in the configurations directory. All training, preprocessing, and model details are specified via YAML.

---

📚 Dataset

• Name: TripAdvisor Review Rating Dataset
• Size: 201,295 hotel reviews
• Labels: Integer ratings (1–5 stars)
• Fields Used: review (text), overall (label)
• Source: Hugging Face
• Language: English (filtered using fastText)

---

🛠️ Data Processing

Preprocessing Steps

• Lowercasing
• Punctuation removal
• Byte-Pair Encoding (BPE) tokenization
• Optional: Lemmatization, stopword removal (in notebooks)

Imbalance Handling

The dataset is heavily skewed towards high ratings. We addressed this via:

• Class weighting
• Macro-averaged metrics
• Data augmentation (see below)

---

🔁 Data Augmentation

Implemented via a modular class using nlpaug, targeting minority class expansion:

• Synonym replacement
• Sentence reordering
• Word deletion

Augmentation improved classifier performance significantly, e.g., FNN accuracy increased from 0.65 → 0.80.

---

🧠 Classification Models

Model	Accuracy	F1 Score
Naive Bayes (TF-IDF)	0.59	0.57
Logistic Regression (BPE+TF-IDF)	0.66	0.65
Feed-Forward NN (TF-IDF)	0.80	0.80
RNN (Word2Vec)	0.74	0.73
BERT Mini (fine-tuned)	0.77	0.76

Insight: A simple feed-forward model with TF-IDF outperformed more complex RNNs and even a small Transformer.

---

✍️ Generative Models

Input Format

• Rating (1–5) + Keywords → Full review text

Models

Model	BLEU	ROUGE-L	BERTScore F1
N-gram	—	—	—
FNN	0.015	0.194	-0.268
RNN	0.057	0.296	-0.024
Transformer	0.028	0.177	0.073

Limitations: Training from scratch limited performance. Pre-trained models were not fine-tuned due to hardware constraints.

---

🧱 Architecture

Modular, configuration-driven system:

• Plug-and-play support for tokenizers, vectorizers, and models
• Easily switch between CPU, CUDA, and Apple MPS
• YAML-configurable training pipeline

---

📊 Vectorization Methods

• TF–IDF: Best performance for classification
• Word2Vec: Used in RNNs; inferior to learned embeddings in generation tasks

---

📦 Project structure

Repository layout (top-level):

pyproject.toml
README.md
configurations/        # Example YAML configs per pipeline/model
docs/                  # Documentation and figures
notebooks/             # Exploratory notebooks
src/                   # Source code (entrypoint: src/main.py)
  ├── main.py
  ├── data/             # preprocessing, augmentation, datasets
  ├── models/           # model implementations and modules
  ├── my_tokenizers/    # BPE / custom tokenizers
  ├── my_vectorizers/   # TF-IDF, word2vec helpers
  └── pipelines/        # training, evaluation, generation pipelines

Use src/main.py as the CLI entrypoint (via uv run or python -m src.main).

---

📈 Results Summary

• Classification: Feed-forward networks + TF-IDF offer strong baselines
• Generation: Transformer-based models show promise but require pre-training
• Augmentation: Addressed imbalance, significantly boosting performance

---

📄 Technical Report

For detailed methodology, architecture diagrams, results, and limitations, please refer to our full Technical Report (PDF).

---

👥 Authors

• Lucas Duport lucas.duport@epita.fr
• Eugenie Beauvillain eugenie.beauvillain@epita.fr
• Yanis Martin yanis.martin@epita.fr
• Arthur Courselle arthur.courselle@epita.fr
• Flavien Geoffray flavien.geoffray@epita.fr