shenghaowang/absa-for-restaurant-reviews

POC implementation of aspect based sentiment analysis using multi-task learning

Aspect Based Sentiment Analysis (ABSA) for Restaurant Reviews

This is a proof-of-concept implementation of aspect based sentiment analysis using multi-task learning. The model architecture was taken from Eram Munawwar's blog post.

The prototype neural network model was trained on ~3000 restaurant reviews with polarity labels over 5 different aspects. The model achieved a test accuracy of 0.7434 compared against the majority baseline of 0.6410. The accuracy of different aspects is reported as follows.

Aspect	Baseline accuracy	ABSA accuracy
food	0.7225	0.7823
service	0.5872	0.8430
price	0.6145	0.7590
ambience	0.6441	0.7542
anecdotes/miscellaneous	0.5427	0.5897
Overall	0.6410	0.7434

Installation

Python >3.8 is required.

virtualenv absa_env
source absa_env/bin/activate
pip install -r requirements.txt
python -m spacy download en_core_web_md

Commands

python src/make_reviews.py will process the restaurant review data and export it into parquet format.

python src/train.py will train a multi-task neural network model for sentiment analysis.

python src/evaluate.py will make prediction for the test data and compute the accuracy for different aspects.

python src/debug_dataset.py can be used for debugging the process of datasets creation.

python src/debug_model.py can be used for debugging the network architecture.

Source data

• SemEval-2014 Task 4
• SemEval-2014 ABSA Test Data - Gold Annotations
• SemEval-2016 Task 5(Optional)

References

• Tunstall, L., Werra, L. von, & Wolf, T. (2022). Natural language processing with transformers, Revised edition. O'Reilly Media, Inc.
• Aspect Based Sentiment Analysis
• Splitting a Dataset for Machine Learning
• SemEval-2014 Task 4: Aspect Based Sentiment Analysis
• SemEval-2016 Task 5: Aspect Based Sentiment Analysis
• A detailed example of how to generate your data in parallel with PyTorch
• RANZCR: Multi-Head Model [training]
• Tutorial 6: Transformers and Multi-Head Attention