Sentiment Analysis and Aspect-Based Opinion Mining of Customer Reviews for Data-Driven Business Insights using NLP

A college NLP project that goes beyond overall "positive/negative" review sentiment and predicts sentiment per aspect (food, service, price, ambience, ...) of a customer review, then rolls those predictions up into a business-facing insights dashboard — the kind of output a real Data Analyst would hand to a restaurant/product manager.

But the staff was so horrible to us → staff: negative The food is uniformly exceptional → food: positive

Why aspect-based, not just overall sentiment?

A single review often contains mixed sentiment about different things. Aspect-Based Sentiment Analysis (ABSA) recovers that nuance, which is what makes the output business-actionable ("service is the problem, not food") rather than a single vague sentiment score per review.

Dataset

SemEval-2014 Task 4 — Restaurant Reviews, the standard academic benchmark for ABSA. 3,608 train / 1,120 test labeled (sentence, aspect term, polarity) examples.

Reference papers

• Instruct-DeBERTa: A Hybrid Approach for Aspect-Based Sentiment Analysis — arXiv:2408.13202
• Cross-domain aspect-based sentiment analysis (IIT Guwahati) — arXiv:2501.08974

Project structure

├── data/
│   ├── raw/                     # original SemEval .xml.seg files
│   └── processed/                # parsed train.csv / test.csv
├── src/
│   ├── data_utils.py              # parse raw SemEval format -> DataFrame
│   ├── preprocessing.py           # text cleaning + aspect-aware input builder
│   ├── aspect_categories.py       # rolls ~100s of raw aspect terms into 6 business categories
│   ├── classical_model.py         # TF-IDF + Logistic Regression / Linear SVM baseline
│   ├── transformer_model.py       # fine-tunes DistilBERT/DeBERTa (run in Colab, needs GPU)
│   ├── business_insights.py       # aggregates predictions into category-level KPIs + charts
│   └── eda.py                     # exploratory data analysis charts
├── notebooks/
│   └── ABSA_Business_Insights.ipynb   # full narrated pipeline, start to finish
├── figures/                        # generated charts (PNG)
├── reports/                         # generated summary tables
├── requirements.txt
└── README.md

Pipeline

1. Data loading & parsing — raw SemEval format → clean DataFrame
2. EDA — class balance, sentence length, most frequent aspect terms
3. Preprocessing — light cleaning, aspect-aware input ("<aspect> [SEP] <sentence>")
4. Baseline model — TF-IDF (uni+bigrams) + Logistic Regression / Linear SVM
5. Transformer model — fine-tuned DistilBERT (or swap in deberta-v3-base), run separately in Colab
6. Business insights — roll aspect-level predictions up into 6 categories (Food Quality, Service, Price/Value, Ambience, Drinks, Location/Wait), compute mention volume, sentiment mix, and a net sentiment score (%positive − %negative) per category
7. Conclusions & limitations

Results (held-out test set)

Model	Accuracy	Macro F1
Logistic Regression (TF-IDF)	0.704	0.590
Linear SVM (TF-IDF)	0.709	0.577
DistilBERT (fine-tuned, 3 epochs, Colab T4 GPU)	0.814	0.704

The classical baselines do well on the majority positive class (F1 ≈ 0.83) but struggle on neutral (F1 ≈ 0.35–0.38) — a known hard case in ABSA. Fine-tuning DistilBERT closes much of that gap, lifting macro-F1 from ~0.59 to 0.704, confirming that contextual embeddings handle negation and aspect-specific context better than TF-IDF bag-of-words features. There's still headroom before matching published DeBERTa-based SOTA on this benchmark (~82–86% accuracy / ~75–80 macro-F1) — see Limitations.

How to run

Classical pipeline (no GPU needed):

pip install -r requirements.txt
python src/data_utils.py         # parses raw data -> data/processed/*.csv
python src/eda.py                # -> figures/
python src/classical_model.py    # -> reports/classical_model_report.txt, figures/confusion_matrix_classical.png
python src/business_insights.py  # -> reports/business_insights_summary.csv, figures/*.png
jupyter notebook notebooks/ABSA_Business_Insights.ipynb

Transformer pipeline (Google Colab, free T4 GPU):

!pip install -q transformers datasets evaluate accelerate
!python src/transformer_model.py --model_name distilbert-base-uncased --epochs 3

Limitations & possible extensions

• Uses SemEval's given aspect terms — aspect term extraction itself (finding which words are the aspects) is a separate, harder subtask not attempted here.
• The aspect→category taxonomy is a manually built keyword map; it won't generalize perfectly beyond restaurant-domain vocabulary.
• Natural next steps: full end-to-end aspect extraction + sentiment, a small demo UI (Streamlit/Gradio) highlighting per-aspect sentiment inline in a pasted review, or applying the pipeline to a larger, real review corpus (e.g. Yelp Open Dataset).

Author

Akil Raza -