Review Aspect Analyzer (BERT, ABSA)
BERT fine-tuned for aspect-based sentiment classification. Given a sentence and an aspect term from it, the model predicts whether the sentiment toward that aspect is negative, neutral or positive.
Example: sentence "The food was great but the service was slow", aspect "service" -> negative.
Training
- Base model:
bert-base-uncased - Dataset: SemEval-2014 Task 4 (restaurant reviews), "conflict" label removed
- Split: 80/10/10 by sentence
- 4 epochs, learning rate 2e-5, batch size 16, seed 42
Results (test set)
| Accuracy | Macro-F1 |
|---|---|
| 0.7726 | 0.7089 |
Limitations
Trained on restaurant reviews only and on single sentences, so it may perform worse on other domains. The neutral class is the hardest.
Usage
The model is meant to be used with the full app, which extracts aspects with spaCy and then classifies each one: https://github.com/spiyush09/review_Aspect_Analyzer
Input format: the sentence and the aspect are passed to the model as a text pair.