Edit model card

Model fine-tuned from roberta-large for sentiment classification of financial news (emphasis on Canadian news).


This model was train on financial_news_sentiment_mixte_with_phrasebank_75 dataset. This is a customized version of the phrasebank dataset in which I kept only sentence validated by at least 75% annotators. In addition I added ~2000 articles validated manually on Canadian financial news. Therefore the model is more specifically trained for Canadian news. Final result is f1 score of 93.25% overall and 83.6% on Canadian news.

Training data

Training data was classified as follow:

class Description
0 negative
1 neutral
2 positive

How to use roberta-large-financial-news-sentiment-en with HuggingFace

Load roberta-large-financial-news-sentiment-en and its sub-word tokenizer :
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("Jean-Baptiste/roberta-large-financial-news-sentiment-en")
model = AutoModelForSequenceClassification.from_pretrained("Jean-Baptiste/roberta-large-financial-news-sentiment-en")

##### Process text sample (from wikipedia)

from transformers import pipeline

pipe = pipeline("text-classification", model=model, tokenizer=tokenizer)
pipe("Melcor REIT (TSX: MR.UN) today announced results for the third quarter ended September 30, 2022. Revenue was stable in the quarter and year-to-date. Net operating income was down 3% in the quarter at $11.61 million due to the timing of operating expenses and inflated costs including utilities like gas/heat and power")

[{'label': 'negative', 'score': 0.9399105906486511}]

Model performances

Overall f1 score (average macro)

precision recall f1
0.9355 0.9299 0.9325

By entity

entity precision recall f1
negative 0.9605 0.9240 0.9419
neutral 0.9538 0.9459 0.9498
positive 0.8922 0.9200 0.9059
Downloads last month
Model size
355M params
Tensor type

Dataset used to train Jean-Baptiste/roberta-large-financial-news-sentiment-en