README.md

metadata

license: mit

Three classes sentiment analysis (positive, negative, neutral)

Based on https://huggingface.co/j-hartmann/sentiment-roberta-large-english-3-classes

Fine-tuned using:

• annotated sentences from book reviews in English https://www.gti.uvigo.es/index.php/en/book-reviews-annotated-dataset-for-aspect-based-sentiment-analysis
• annotated paragraphs from amateur writers' stories https://arxiv.org/abs/1910.11769

Performance for books:

Num examples = 1666

Batch size = 16

	precision	recall	f1-score	support
Negative	0.83	0.88	0.85	844
Neutral	0.68	0.51	0.58	351
Positive	0.79	0.85	0.82	471
accuracy			0.79	1666
macro avg	0.76	0.75	0.75	1666
weighted avg	0.78	0.79	0.78	1666

Performance for reviews:

Num examples = 205

Batch size = 16

	precision	recall	f1-score	support
Negative	0.89	0.92	0.91	26
Neutral	0.96	0.91	0.94	90
Positive	0.94	0.98	0.96	89
accuracy			0.94	205
macro avg	0.93	0.94	0.93	205
weighted avg	0.94	0.94	0.94	205