Sentiment Analysis on Customer Reviews in R

---

Overview

This project performs sentiment analysis on TripAdvisor hotel reviews using Machine Learning and Lexicon-based techniques in R.

It combines lexicon-based sentiment extraction (syuzhet, bing, afinn) with a supervised TF-IDF + GLMNET classifier, built using the text2vec and caret frameworks.

---

Features

• Text preprocessing and TF-IDF vectorization using text2vec
• Lexicon-based sentiment scoring using Syuzhet, Bing, Afinn
• Emotion analysis using NRC lexicon
• Supervised model training using Logistic Regression (GLMNET)
• Model evaluation using caret and pROC
• Reproducible dependency management with renv

---

Project Structure

aspa-sentiment-r/
├── data/                       # Dataset (20,491 TripAdvisor reviews)
├── scripts/                    # R scripts for analysis
│   ├── 00_project_setup.R
│   ├── 01_exploratory_sentiment.R
│   └── 02_supervised_sentiment_model.R
├── figures/                    # Generated figures
│   ├── wordcloud_text2vec.png
│   ├── top_terms_tfidf_top20.png
│   ├── nrc_emotion_counts.png
│   ├── sentiment_hist_lexicons.png
│   └── sentiment_distribution_bing.png
├── models/                     # Vocabulary, TF-IDF transformer, model
├── reports/                    # CSV outputs and samples
├── renv.lock                   # Dependencies snapshot
└── aspa-sentiment-r.Rproj

---

Visual Results

1. Word Cloud — Top Frequent & Informative Terms

Visualization of the most prominent terms by TF-IDF score.

---

2. Top 20 Terms by TF-IDF

High-weight words that contribute most to the sentiment model.

---

3. Emotion Breakdown (NRC Lexicon)

Distribution of emotional categories such as joy, trust, sadness, anger, etc.

---

4. Sentiment Score Distributions (Lexicon Comparison)

Comparative sentiment score ranges across Syuzhet, Bing, and Afinn lexicons.

---

5. Sentiment Distribution (Pie Chart)

Proportion of positive, neutral, and negative sentiments based on Bing lexicon.

---

Model Performance Summary

Metric	Value	Description
Accuracy	94.6%	Overall classification accuracy
AUC (ROC)	0.9836	Excellent discrimination ability
Sensitivity	98.6%	Correctly detects positive sentiments
Specificity	75.3%	Correctly detects negative sentiments
Kappa	0.798	Strong agreement beyond chance

The model shows outstanding predictive performance, demonstrating the power of TF-IDF + GLMNET for textual sentiment classification.

---

Setup Instructions

1. Clone the repository:

   git clone https://github.com/Ash-the-k/aspa-sentiment-r
   cd aspa-sentiment-r

2. Open the project in RStudio.

3. Restore dependencies (like pip install -r requirements.txt in Python):

   install.packages("renv")
   renv::restore()

4. Run the analysis scripts:

   source("scripts/01_exploratory_sentiment.R")
   source("scripts/02_supervised_sentiment_model.R")

All visualizations will be generated in the figures/ folder, and results in reports/.

---

Methodology Summary

1. Data Cleaning: Removal of punctuation, symbols, and stopwords.
2. TF-IDF Construction: Sparse document-term matrix creation via text2vec.
3. Exploratory Lexicon Analysis: Sentiment and emotion extraction using syuzhet and NRC.
4. Supervised Model Training: Logistic Regression (glmnet) trained on TF-IDF features.
5. Evaluation: Confusion matrix, accuracy, and ROC-AUC metrics.

---

Conclusion

This project demonstrates an end-to-end sentiment analysis workflow in R, integrating both descriptive (lexicon) and predictive (machine learning) approaches. It can serve as a reproducible template for future NLP applications in R.

---

License

This project is shared under the MIT License — free to use, modify, and distribute with attribution.