Twitter-roBERTa-base for Emotion Recognition

This is a roBERTa-base model trained on ~58M tweets and finetuned for emotion recognition with the TweetEval benchmark.

• Paper: TweetEval benchmark (Findings of EMNLP 2020).
• Git Repo: Tweeteval official repository.

Example of classification

from transformers import AutoModelForSequenceClassification<br /> from transformers import TFAutoModelForSequenceClassification<br /> from transformers import AutoTokenizer<br /> import numpy as np<br /> from scipy.special import softmax<br /> import csv<br /> import urllib.request<br /> # Preprocess text (username and link placeholders)<br /> def preprocess(text):<br /> new_text = []<br /> for t in text.split(" "):<br /> t = '@user' if t.startswith('@') and len(t) > 1 else t<br /> t = 'http' if t.startswith('http') else t<br /> new_text.append(t)<br /> return " ".join(new_text)<br /> # Tasks:<br /> # emoji, emotion, hate, irony, offensive, sentiment<br /> # stance/abortion, stance/atheism, stance/climate, stance/feminist, stance/hillary<br /> task='emotion'<br /> MODEL = f"cardiffnlp/twitter-roberta-base-{task}"<br /> tokenizer = AutoTokenizer.from_pretrained(MODEL)<br /> # download label mapping<br /> mapping_link = f"https://raw.githubusercontent.com/cardiffnlp/tweeteval/main/datasets/{task}/mapping.txt"<br /> with urllib.request.urlopen(mapping_link) as f:<br /> html = f.read().decode('utf-8').split("\n")<br /> csvreader = csv.reader(html, delimiter='\t')<br /> labels = [row[1] for row in csvreader if len(row) > 1]<br /> # PT<br /> model = AutoModelForSequenceClassification.from_pretrained(MODEL)<br /> model.save_pretrained(MODEL)<br /> text = "Celebrating my promotion "<br /> text = preprocess(text)<br /> encoded_input = tokenizer(text, return_tensors='pt')<br /> output = model(**encoded_input)<br /> scores = output[0][0].detach().numpy()<br /> scores = softmax(scores)<br /> # # TF<br /> # model = TFAutoModelForSequenceClassification.from_pretrained(MODEL)<br /> # model.save_pretrained(MODEL)<br /> # text = "Celebrating my promotion "<br /> # encoded_input = tokenizer(text, return_tensors='tf')<br /> # output = model(encoded_input)<br /> # scores = output[0][0].numpy()<br /> # scores = softmax(scores)<br /> ranking = np.argsort(scores)<br /> ranking = ranking[::-1]<br /> for i in range(scores.shape[0]):<br /> l = labels[ranking[i]]<br /> s = scores[ranking[i]]<br /> print(f"{i+1}) {l} {np.round(float(s), 4)}")<br />

Output:
1) joy 0.9382<br /> 2) optimism 0.0362<br /> 3) anger 0.0145<br /> 4) sadness 0.0112<br />