sloria/TextBlob

> A thin, Pythonic convenience layer over NLTK and pattern for common NLP tasks — POS tagging, noun-phrase extraction, and lexicon sentiment in a few lines.

GitHub repo · Official website · License: MIT

Overview

TextBlob is a text-processing library created by Steven Loria, first released in 2013¹. Its stated goal is ergonomics: wrap the two established but verbose NLP toolkits — NLTK and pattern — behind a single TextBlob object so that tagging, tokenization, sentiment, and noun-phrase extraction read like attribute access rather than pipeline assembly. The README puts it plainly: it "stands on the giant shoulders of NLTK and pattern, and plays nicely with both."²

The library is best understood as a teaching and prototyping tool. It is a fixture of introductory NLP courses, tutorials, and quick scripts precisely because TextBlob(text).sentiment.polarity requires no model download decisions, no pipeline configuration, and no ML background. The defining tradeoff is that this simplicity is bought with dated, lexicon-based methods: the default sentiment analyzer is pattern's hand-built adjective lexicon, and the tagger and parser are classical (rule/statistical) rather than neural. Accuracy on modern text — social media, domain jargon, sarcasm — is mediocre by 2020s standards.

TextBlob went through a long quiet period and has since had a maintenance revival: packaging was modernized, Python 2 and old Python 3 versions were dropped, and the Google-Translate-backed features were removed after the underlying endpoint stopped being usable³. It remains genuinely maintained for compatibility, not actively developed toward new NLP capabilities.

Getting Started

pip install -U textblob
python -m textblob.download_corpora   # pulls the NLTK data TextBlob needs

from textblob import TextBlob

blob = TextBlob("TextBlob is amazingly simple to use. What great fun!")

blob.tags            # [('TextBlob', 'NN'), ('is', 'VBZ'), ('amazingly', 'RB'), ...]
blob.noun_phrases    # WordList(['textblob'])
blob.sentiment       # Sentiment(polarity=0.39, subjectivity=0.62)

for sentence in blob.sentences:
    print(sentence.sentiment.polarity)   # per-sentence scores

blob.words[0].pluralize()          # 'TextBlobs'
TextBlob("I havv goood speling.").correct()   # 'I have good spelling.'

The download_corpora step is not optional for most features — POS tagging, tokenization, and WordNet access all read NLTK data files at runtime, and a fresh environment will raise LookupError from deep inside NLTK if they are missing.

Architecture / How It Works

TextBlob is a facade, not an engine. Almost every method delegates to NLTK or to a vendored copy of pattern's code, and the design is deliberately pluggable through small strategy classes:

• Sentiment — two analyzers. The default PatternAnalyzer scores text against pattern's English adjective lexicon and returns a (polarity, subjectivity) pair; polarity is a lexicon average, not a model prediction, so it is fast, deterministic, and blind to context and negation beyond simple heuristics. The alternative NaiveBayesAnalyzer is trained on an NLTK movie-review corpus and returns (classification, p_pos, p_neg); it is slower and domain-biased toward reviews.
• POS tagging — PatternTagger (from pattern) or NLTKTagger. You select by passing pos_tagger= to the constructor.
• Noun-phrase extraction — FastNPExtractor by default, or ConllExtractor (chunker trained on the CoNLL-2000 corpus) for better phrase boundaries at higher cost.
• Tokenization — pluggable BaseTokenizer subclasses wrapping NLTK's word and sentence tokenizers.
• Spelling correction — a direct implementation of Peter Norvig's edit-distance-plus-word-frequency approach, driven by a bundled frequency file. It is per-word and context-free.
• Classification — NaiveBayesClassifier and DecisionTreeClassifier thinly wrap NLTK's classifiers, adding TextBlob's feature-extraction convenience over (text, label) tuples.

The TextBlob, Sentence, and Word types share a BaseBlob mixin so that a blob, a sentence, and a word expose the same interface where it makes sense. WordNet access on Word is a straight pass-through to NLTK's wordnet corpus reader. Because the analyzers, taggers, and extractors are constructor-injected, swapping in a custom implementation is the intended extension path — but the abstractions are shallow, and anything beyond the built-ins means reaching back down into NLTK directly.

Production Notes

• This is a prototyping library, not a production NLP stack. Its sentiment and tagging quality lags spaCy and transformer models by a wide margin. Teams routinely start with TextBlob for a demo and then discover the polarity scores are too coarse for real decisions.
• Corpus data is a deployment footgun. textblob.download_corpora writes NLTK data to a user/home directory that may not exist or be writable in containers, serverless functions, or read-only images. The fix is to run the download during image build and/or set NLTK_DATA to a baked-in path. A missing corpus surfaces as a runtime LookupError on first use, not at import.
• Translation and language detection are gone. Earlier versions offered TextBlob.translate() and .detect_language() backed by an undocumented Google Translate endpoint. Google's changes broke it, and the feature was removed³. Code and tutorials that still call these methods will fail — do not adopt TextBlob for translation.
• Sentiment is a lexicon average, and it shows. PatternAnalyzer has no real handling of negation scope, intensifiers beyond a small table, sarcasm, or domain vocabulary. Scores cluster near zero on neutral factual text and are easily fooled. For social-media tone specifically, cjhutto/vaderSentiment is a better rule-based choice; for accuracy, a fine-tuned transformer.
• Performance is adequate but single-threaded and CPU-bound. There is no batching API; processing large corpora means your own loop plus multiprocessing. First use of the Naive Bayes analyzer or the CoNLL extractor pays a corpus-load cost.
• Determinism cuts both ways. The lexicon path gives stable, reproducible scores across runs and versions — useful for tests — but that stability is also why the numbers never improve.

When to Use / When Not

Use when:

• You want the shortest path from a string to POS tags, tokens, noun phrases, or a rough sentiment score.
• You're teaching, prototyping, or scripting, and readability matters more than accuracy.
• You need a light dependency for classical NLP without pulling in large model files.
• You already rely on NLTK and want a friendlier surface over it.

Avoid when:

• You need production-grade accuracy on sentiment, NER, or tagging — use spaCy or transformers.
• Your text is social media, multilingual, or domain-specific jargon.
• You need translation or language detection (removed).
• You need GPU acceleration, batching, or modern embeddings — TextBlob offers none.

Alternatives

• nltk/nltk — the toolkit TextBlob wraps; use it directly when you need control over tokenizers, corpora, or classifiers rather than convenience defaults.
• explosion/spaCy — use when you need fast, accurate production tagging, parsing, and NER with a real pipeline API.
• huggingface/transformers — use when sentiment/classification accuracy is the priority and you can run (or fine-tune) a model.
• cjhutto/vaderSentiment — use instead of TextBlob's default when you specifically want rule-based sentiment tuned for social-media text.
• clips/pattern — TextBlob's other parent library (web mining plus NLP); largely unmaintained, but the source of the default sentiment lexicon.
• flairNLP/flair — use when you want strong sequence labeling (NER/POS) via contextual embeddings with a still-approachable API.

History

References

1. ^ TextBlob project history and authorship — Steven Loria. https://github.com/sloria/TextBlob
2. ^ TextBlob README — "stands on the giant shoulders of NLTK and pattern." https://github.com/sloria/TextBlob/blob/dev/README.rst
3. ^ TextBlob changelog (translation/detection removal, packaging and Python-version changes). https://textblob.readthedocs.io/en/latest/changelog.html

Tags

python, nlp, natural-language-processing, sentiment-analysis, pos-tagging, text-processing, nltk, pattern, prototyping, library