README.md
| 1 | --- |
| 2 | language: multilingual |
| 3 | widget: |
| 4 | - text: "🤗" |
| 5 | - text: "T'estimo! ❤️" |
| 6 | - text: "I love you!" |
| 7 | - text: "I hate you 🤮" |
| 8 | - text: "Mahal kita!" |
| 9 | - text: "사랑해!" |
| 10 | - text: "난 너가 싫어" |
| 11 | - text: "😍😍😍" |
| 12 | --- |
| 13 | |
| 14 | |
| 15 | # twitter-XLM-roBERTa-base for Sentiment Analysis |
| 16 | |
| 17 | This is a multilingual XLM-roBERTa-base model trained on ~198M tweets and finetuned for sentiment analysis. The sentiment fine-tuning was done on 8 languages (Ar, En, Fr, De, Hi, It, Sp, Pt) but it can be used for more languages (see paper for details). |
| 18 | |
| 19 | - Paper: [XLM-T: A Multilingual Language Model Toolkit for Twitter](https://arxiv.org/abs/2104.12250). |
| 20 | - Git Repo: [XLM-T official repository](https://github.com/cardiffnlp/xlm-t). |
| 21 | |
| 22 | This model has been integrated into the [TweetNLP library](https://github.com/cardiffnlp/tweetnlp). |
| 23 | |
| 24 | ## Example Pipeline |
| 25 | ```python |
| 26 | from transformers import pipeline |
| 27 | model_path = "cardiffnlp/twitter-xlm-roberta-base-sentiment" |
| 28 | sentiment_task = pipeline("sentiment-analysis", model=model_path, tokenizer=model_path) |
| 29 | sentiment_task("T'estimo!") |
| 30 | ``` |
| 31 | ``` |
| 32 | [{'label': 'Positive', 'score': 0.6600581407546997}] |
| 33 | ``` |
| 34 | |
| 35 | ## Full classification example |
| 36 | |
| 37 | ```python |
| 38 | from transformers import AutoModelForSequenceClassification |
| 39 | from transformers import TFAutoModelForSequenceClassification |
| 40 | from transformers import AutoTokenizer, AutoConfig |
| 41 | import numpy as np |
| 42 | from scipy.special import softmax |
| 43 | |
| 44 | # Preprocess text (username and link placeholders) |
| 45 | def preprocess(text): |
| 46 | new_text = [] |
| 47 | for t in text.split(" "): |
| 48 | t = '@user' if t.startswith('@') and len(t) > 1 else t |
| 49 | t = 'http' if t.startswith('http') else t |
| 50 | new_text.append(t) |
| 51 | return " ".join(new_text) |
| 52 | |
| 53 | MODEL = f"cardiffnlp/twitter-xlm-roberta-base-sentiment" |
| 54 | |
| 55 | tokenizer = AutoTokenizer.from_pretrained(MODEL) |
| 56 | config = AutoConfig.from_pretrained(MODEL) |
| 57 | |
| 58 | # PT |
| 59 | model = AutoModelForSequenceClassification.from_pretrained(MODEL) |
| 60 | model.save_pretrained(MODEL) |
| 61 | |
| 62 | text = "Good night 😊" |
| 63 | text = preprocess(text) |
| 64 | encoded_input = tokenizer(text, return_tensors='pt') |
| 65 | output = model(**encoded_input) |
| 66 | scores = output[0][0].detach().numpy() |
| 67 | scores = softmax(scores) |
| 68 | |
| 69 | # # TF |
| 70 | # model = TFAutoModelForSequenceClassification.from_pretrained(MODEL) |
| 71 | # model.save_pretrained(MODEL) |
| 72 | |
| 73 | # text = "Good night 😊" |
| 74 | # encoded_input = tokenizer(text, return_tensors='tf') |
| 75 | # output = model(encoded_input) |
| 76 | # scores = output[0][0].numpy() |
| 77 | # scores = softmax(scores) |
| 78 | |
| 79 | # Print labels and scores |
| 80 | ranking = np.argsort(scores) |
| 81 | ranking = ranking[::-1] |
| 82 | for i in range(scores.shape[0]): |
| 83 | l = config.id2label[ranking[i]] |
| 84 | s = scores[ranking[i]] |
| 85 | print(f"{i+1}) {l} {np.round(float(s), 4)}") |
| 86 | |
| 87 | ``` |
| 88 | |
| 89 | Output: |
| 90 | |
| 91 | ``` |
| 92 | 1) Positive 0.7673 |
| 93 | 2) Neutral 0.2015 |
| 94 | 3) Negative 0.0313 |
| 95 | ``` |
| 96 | |
| 97 | ### Reference |
| 98 | ``` |
| 99 | @inproceedings{barbieri-etal-2022-xlm, |
| 100 | title = "{XLM}-{T}: Multilingual Language Models in {T}witter for Sentiment Analysis and Beyond", |
| 101 | author = "Barbieri, Francesco and |
| 102 | Espinosa Anke, Luis and |
| 103 | Camacho-Collados, Jose", |
| 104 | booktitle = "Proceedings of the Thirteenth Language Resources and Evaluation Conference", |
| 105 | month = jun, |
| 106 | year = "2022", |
| 107 | address = "Marseille, France", |
| 108 | publisher = "European Language Resources Association", |
| 109 | url = "https://aclanthology.org/2022.lrec-1.27", |
| 110 | pages = "258--266" |
| 111 | } |
| 112 | |
| 113 | ``` |
| 114 | |
| 115 | |