Note: This model & model card are based on the finetuned XLM-T for Sentiment Analysis

twitter-XLM-roBERTa-base for Emotion Analysis

This is a XLM-roBERTa-base model trained on ~198M tweets and finetuned for emotion analysis on Spanish language. This model was presented to EmoEvalEs competition, part of IberLEF 2021 Conference, where the proposed task was the classification of Spanish tweets between seven different classes: anger, disgust, fear, joy, sadness, surprise, and other. We achieved the first position in the competition with a macro-averaged F1 score of 71.70%.

Example Pipeline with a Tweet from @JaSantaolalla

from transformers import pipeline
model_path = "daveni/twitter-xlm-roberta-emotion-es"
emotion_analysis = pipeline("text-classification", framework="pt", model=model_path, tokenizer=model_path)
emotion_analysis("Einstein dijo: Solo hay dos cosas infinitas, el universo y los pinches anuncios de bitcoin en Twitter. Paren ya carajo aaaaaaghhgggghhh me quiero murir")

[{'label': 'anger', 'score': 0.48307016491889954}]

Full classification example

from transformers import AutoModelForSequenceClassification
from transformers import AutoTokenizer, AutoConfig
import numpy as np
from scipy.special import softmax
# Preprocess text (username and link placeholders)
def preprocess(text):
    new_text = []
    for t in text.split(" "):
        t = '@user' if t.startswith('@') and len(t) > 1 else t
        t = 'http' if t.startswith('http') else t
        new_text.append(t)
    return " ".join(new_text)
model_path = "daveni/twitter-xlm-roberta-emotion-es"
tokenizer = AutoTokenizer.from_pretrained(model_path )
config = AutoConfig.from_pretrained(model_path )
# PT
model = AutoModelForSequenceClassification.from_pretrained(model_path )
text = "Se ha quedao bonito día para publicar vídeo, ¿no? Hoy del tema más diferente que hemos tocado en el canal."
text = preprocess(text)
print(text)
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)
scores = output[0][0].detach().numpy()
scores = softmax(scores)
# Print labels and scores
ranking = np.argsort(scores)
ranking = ranking[::-1]
for i in range(scores.shape[0]):
    l = config.id2label[ranking[i]]
    s = scores[ranking[i]]
    print(f"{i+1}) {l} {np.round(float(s), 4)}")

Output:

Se ha quedao bonito día para publicar vídeo, ¿no? Hoy del tema más diferente que hemos tocado en el canal.
1) joy 0.7887
2) others 0.1679
3) surprise 0.0152
4) sadness 0.0145
5) anger 0.0077
6) disgust 0.0033
7) fear 0.0027

Limitations and bias

The dataset we used for finetuning was unbalanced, where almost half of the records belonged to the other class so there might be bias towards this class.

Training data

Pretrained weights were left identical to the original model released by cardiffnlp. We used the EmoEvalEs Dataset for finetuning.

BibTeX entry and citation info

@inproceedings{vera2021gsi,
  title={GSI-UPM at IberLEF2021: Emotion Analysis of Spanish Tweets by Fine-tuning the XLM-RoBERTa Language Model},
  author={Vera, D and Araque, O and Iglesias, CA},
  booktitle={Proceedings of the Iberian Languages Evaluation Forum (IberLEF 2021). CEUR Workshop Proceedings, CEUR-WS, M{\'a}laga, Spain},
  year={2021}
}