anynlp 0.2.0

One-liner NLP utilities -- summarize, classify, extract entities, analyze sentiment -- with rule-based fallbacks and HuggingFace backends.

anynlp

One-liner NLP utilities -- summarize, classify, extract entities, analyze sentiment, and more.

Runs completely offline. Works out of the box with zero dependencies using rule-based backends. Optionally upgrade to HuggingFace transformers for ML-powered results -- models download once and are cached locally.

Built by Viet-Anh Nguyen | GitHub

---

Quick Start

pip install anynlp

import anynlp

# Summarize text
summary = anynlp.summarize("Your long article text here...", ratio=0.3)
print(summary.text)

# Sentiment analysis
sentiment = anynlp.sentiment("I absolutely love this library!")
print(sentiment)  # positive (90.00%)

# Named entity extraction
entities = anynlp.entities("Contact [e-mail] or call 555-123-4567 by January 15, 2024.")
for entity in entities:
    print(f"{entity.text} [{entity.label}]")

# Keyword extraction
keywords = anynlp.keywords("Machine learning is transforming artificial intelligence research.", n=5)
for kw in keywords:
    print(f"{kw.word}: {kw.score:.2f}")

# Text classification
label = anynlp.classify("Great product, highly recommend!", labels=["positive", "negative"])
print(label)  # positive (100.00%)

# Text similarity
score = anynlp.similarity("Hello world", "Hi there world")
print(float(score))  # 0.3333

Features

Task	Function	Rule-Based	HuggingFace
Summarization	anynlp.summarize(text)	Sentence scoring (position, frequency, length)	Abstractive summarization models
Classification	anynlp.classify(text, labels)	Keyword overlap scoring	Zero-shot classification
NER	anynlp.entities(text)	Regex patterns (email, phone, date, URL, etc.)	BERT-based NER models
Sentiment	anynlp.sentiment(text)	Lexicon-based with negation handling	Transformer sentiment models
Keywords	anynlp.keywords(text, n)	TF-IDF-like scoring	TF-IDF-like scoring
Similarity	anynlp.similarity(a, b)	Jaccard similarity on word sets	Jaccard similarity on word sets

Installation Options

# Core (rule-based, zero dependencies)
pip install anynlp

# With HuggingFace transformers
pip install anynlp[transformers]

# Everything
pip install anynlp[all]

Backend Selection

anynlp automatically selects the best available backend:

• If transformers and torch are installed, uses HuggingFace models
• Otherwise, falls back to rule-based methods (zero dependencies)

You can force a specific backend:

# Always use rule-based (fast, no downloads)
result = anynlp.sentiment("Great!", backend="rules")

# Always use HuggingFace (requires transformers + torch)
result = anynlp.sentiment("Great!", backend="huggingface")

Plugin System

Add custom NLP tasks using the registry pattern:

from anynlp import NLPTask, register_task

class TranslateTask(NLPTask):
    name = "translate"

    def run(self, text="", target_lang="en", **kwargs):
        # Your translation logic here
        return {"text": text, "lang": target_lang}

# Register and use
register_task("translate", TranslateTask)

from anynlp.core import registry
result = registry.run("translate", text="Bonjour", target_lang="en")

Result Objects

All functions return rich dataclasses with useful representations:

summary = anynlp.summarize(text)
summary.text                  # The summarized text
summary.sentence_count        # Number of sentences in summary
summary.original_sentence_count  # Number in original
str(summary)                  # The text itself
repr(summary)                 # Summary(sentences=3/10, ratio=0.30)

sentiment = anynlp.sentiment("I love this!")
sentiment.label               # "positive"
sentiment.score               # 0.9 (0-1 scale, 0.5 = neutral)

entities = anynlp.entities(text)
len(entities)                 # Number of entities
entities[0].text              # Entity text
entities[0].label             # Entity type (EMAIL, URL, DATE, etc.)

similarity = anynlp.similarity(text1, text2)
float(similarity)             # Score as float (0.0-1.0)
similarity.method             # "jaccard"

Local-First / Edge AI

This package is designed to work completely offline. The rule-based backend requires zero internet access. For ML-powered results, HuggingFace models download once and are cached locally -- no internet needed after that.

# Pre-download HuggingFace models for offline use
python -m anynlp download

import anynlp

# Pre-download all HuggingFace models
anynlp.download_models()

# Rule-based backend always works offline (no downloads needed)
result = anynlp.sentiment("Great product!", backend="rules")

Development

git clone https://github.com/vietanhdev/anynlp.git
cd anynlp
pip install -e ".[all]"
pytest

See CONTRIBUTING.md for guidelines on adding new tasks and backends.

License

MIT License. See LICENSE for details.

---

Built with care by NRL.AI