TweetNLP: Cutting-Edge Natural Language Processing for Social Media

• Social media text includes noise, multilinguality, slang, immediacy, limited context, length restrictions, and emojis that standard NLP must handle.Source: §1 · p. 1
• TweetNLP integrates Twitter-trained Transformer models fine-tuned for social-media NLP tasks.Source: §1 · p. 1
• The platform provides a Python API, an interactive online demo, and real-time Twitter analysis.Source: §1 · pp. 1–2
• Over 15M downloads made TweetNLP’s sentiment model the most downloaded model in Hugging Face’s model hub in January 2022.Source: §1 · p. 1

• General-purpose libraries such as CoreNLP, Stanza, spaCy, and Hugging Face Transformers are not specialized for social media or Twitter.Source: §2 · p. 2
• Earlier social-media libraries mainly addressed tokenization, part-of-speech tagging, and dependency parsing.Source: §2 · p. 2
• TweebankNLP provides Twitter-specific models for tokenization, lemmatization, NER, part-of-speech tagging, and dependency parsing.Source: §2 · p. 2

3.1 Language models

• TweetNLP uses RoBERTa and XLM-R checkpoints that are continually pre-trained on Twitter-specific corpora.Source: §3.1 · p. 2
• RoBERTa-base and XLM-R-base architectures support efficient execution on standard hardware and free cloud-computing tiers, including without GPU support.Source: §3.1 · p. 2
• The model family includes TweetEval, continually updated TimeLMs, and multilingual XLM-T trained on over thirty languages.Source: §3.1 · p. 2

3.2 Supported tasks

• TweetNLP fine-tunes its models for classification tasks and uses T-NER for named entity recognition.Source: §3.2 · p. 2
• Supported tasks include sentiment, emotion, emoji, irony, hate-speech, offensive-language, stance, topic classification, and named entity recognition.Source: §3.2 · p. 3
• Sentiment analysis predicts positive, neutral, or negative labels, while topic classification assigns one or more of 19 topics.Source: §3.2 · p. 3

3.3 Embeddings

• TweetNLP also releases word and tweet embeddings for applications including similarity analysis and tweet retrieval.Source: §3.3 · pp. 3–4

3.2 Supported tasks

• Topic-classification and named-entity-recognition datasets were internally annotated and were still being prepared for public release.Source: §3.3 · p. 3

• TweetNLP is distributed through PyPI and can be installed with pip.Source: §4 · p. 4
• The library loads task-specific models in the background and exposes predefined functions for predictions on tweets or sentences.Source: §4 · p. 4
• Users can load sentiment, tweet-embedding, and masked-language models through the same simple API pattern.Source: §4 · p. 4
• Custom loading of existing fine-tuned language models not included in TweetNLP is also supported.Source: §4 · p. 4

Tutorial platform

• The tutorials provide starter code and examples for users with varying experience in NLP and social-media processing.They expose the underlying Python code alongside the TweetNLP library.Source: §5 · p. 4

Tutorials

• The introduction tutorial demonstrates specialized social-media models across tasks including sentiment analysis and named entity recognition.Source: §5 · p. 4

Data collection

• The data tutorial explains Twitter API interaction and retrieving tweets using hashtags or keywords.Source: §5 · p. 4

Model adaptation

• The custom fine-tuning tutorial covers adapting language models to specific tasks and evaluating them on TweetEval.It uses TweetEval data and its unified format.Source: §5 · p. 4

Representations

• The word-embeddings tutorial covers training embeddings on custom data and obtaining similarity scores or nearest-neighbor words.Source: §5 · p. 4

Temporal models

• The language-models-over-time tutorial uses TimeLMs to access models trained during short periods from 2019 onward.Source: §5 · p. 4

Representations

• The tweet-embeddings tutorial shows how vector representations support tweet similarity and retrieval.Source: §5 · p. 4

Web demo

• The comprehensive web demo integrates TweetNLP models so users can test them and obtain predictions without writing code.Source: §6 · p. 4

Tweet classification

• Tweet classification accepts sentences, tweets, or tweet URLs and displays prediction confidence, including English tasks and multilingual sentiment analysis.Source: §6 · p. 5

Hashtag analysis

• Hashtag analysis retrieves tweets for a chosen keyword, date interval, task, and language, then computes aggregated results.Source: §6 · p. 5

Word prediction

• Word prediction ranks likely replacements for a masked word using a masked language model.Source: §6 · p. 5

Tweet similarity

• Tweet similarity displays a 0-100 cosine similarity score for two short text pieces using the default tweet-embedding model.Source: §6 · p. 5

Named Entity Recognition

• Named Entity Recognition locates named entities in a tweet or sentence and infers their types.Source: §6 · p. 5

Experimental setting

• Evaluation uses TweetEval datasets for tweet classification tasks except topic classification, alongside the train/validation/test splits described for the supported tasks.Source: §7.1 · p. 5
• Default models use TimeLMs-21 for English and XLM-T for other languages and multilingual tasks, fine-tuned to each corresponding task.Source: §7.1 · p. 6
• Comparison systems include frequency-based SVM, fastText, BiLSTM, RoBERTa variants, the original TweetEval model, and BERTweet.Source: §7.1 · p. 6

Results

• The default TimeLMs-21 model achieves the overall results on most of the nine Twitter-based tasks, especially versus a comparable general-purpose RoBERTa model.Source: §7.2 · p. 6

Caveat

• BERTweet’s Irony result overlaps with its pre-training corpus and the Irony dataset, which was constructed using distant supervision.Source: §7.2 · p. 6

Results

• XLM-T achieves the best overall multilingual sentiment-analysis results in all evaluated languages against XLM-R and fastText.Source: §7.2 · p. 6
• 0.77 (WS-Sim), 0.72 (WS-Rel), 0.69 (SemEval), and 0.79 (MEN) are the Spearman correlations for the Twitter word embeddings, which outperform standard fastText on two datasets.Source: §7.2 · p. 6
• Tweet embeddings significantly outperform all-mpnet-base-v2 on tweet-reply retrieval, while all-mpnet-base-v2 performs best on the STS-Benchmark.Source: §7.2 · p. 7

• TweetNLP provides relatively lightweight language models trained on Twitter and adapted to popular social-media NLP tasks.Source: §8 · p. 7
• The platform combines an online demo, Python library, and tutorial for users with different levels of expertise.Source: §8 · p. 7
• TweetNLP supports inspection by non-programmers to help identify harmful biases or errors and improve models in the future.Source: §8 · p. 7
• Future work will add models and tasks, develop social-media datasets, extend beyond tweet classification, and support additional platforms and languages.Source: §8 · p. 7

• Models trained on social-media data may amplify existing biases present on the platform.Source: §9 · p. 8
• Dataset-related tweets were anonymized with URLs removed, and creators often tried to remove offensive or harmful content.Source: §9 · p. 8

• The sentiment-analysis demo supports English and numerous additional languages, including Arabic, Chinese, French, German, Hindi, Japanese, Korean, Persian, and Vietnamese.Source: A Languages supported · p. 11
• The supported languages use both an XLM-T multilingual model and the Twitter API.Source: A Languages supported · p. 11