sentikit

A sentiment analysis library that goes beyond lexicon tools like VADER and TextBlob by actually fixing their known failure modes instead of just adding more words to a dictionary.

Why not just use VADER?

Honest scope note: sarcasm and fully implicit sentiment (no sentiment words at all) are open research problems. sentikit ships best-effort modules for these with documented confidence handling — it does not claim to solve them outright.

Install

pip install sentikit                 # lite mode only, zero heavy deps
pip install sentikit[parse]          # + spaCy for accurate negation scope
pip install sentikit[transformers]   # + transformer sentiment/sarcasm/ABSA
pip install sentikit[all]            # everything

python -m spacy download en_core_web_sm   # needed if you installed [parse]

Lite mode works completely offline with zero downloads. Full mode downloads models from HuggingFace on first use of that feature (not at import time).

Usage

Lite mode (default) — fast, offline, no downloads

from sentikit import Analyzer

sa = Analyzer()  # mode="lite" by default

r = sa.analyze("I do not think this movie was particularly good")
print(r.compound, r.label)          # -0.356 negative
print(r.negation_applied)           # True — correctly scoped "not ... good"
                                     # despite them being far apart

r = sa.analyze("The food was great, but the service was awful")
print(r.label)                      # negative — post-"but" clause dominates

r = sa.analyze("This laptop is a nightmare to work with")
print(r.idioms_matched)             # ['a nightmare']

Full mode — adds transformer sentiment, sarcasm gate, ABSA

sa = Analyzer(mode="full")

result = sa.analyze(
    "The battery life is terrible but the camera is excellent",
    aspects=True,
)
print(result.label)                 # blended rule + transformer score
for a in result.aspects:
    print(a["aspect"], a["sentiment"], a["confidence"])
    # battery life  negative  0.91
    # the camera    positive  0.94

result = sa.analyze("Oh great, another Monday", sarcasm=True)
print(result.sarcasm_probability, result.sarcasm_adjusted, result.label)

Architecture

text
 ├─ idiom phrase matching (masks matched spans before word-level scoring)
 ├─ tokenize
 ├─ negation scope (spaCy dependency parse, or clause-bounded regex fallback)
 ├─ contrastive discourse split ("but"/"however"/...) + asymmetric reweighting
 ├─ lexicon lookup + intensifier scaling
 └─ [full mode only]
     ├─ transformer sentiment backbone (blended with rule-based score)
     ├─ sarcasm gate (confidence-weighted dampen/invert, not a hard switch)
     └─ aspect extraction (noun chunks) + per-aspect transformer classification

Each module (negation.py, contrast.py, idioms.py, sarcasm.py, aspect.py) is independently testable and swappable — this is intentionally a pipeline of small fit-for-purpose components, not one model trying to do everything.

What's still unsolved / roadmap

• Conditionals ("it would be amazing if it worked") — not yet detected; currently scores the embedded positive word at face value.
• Rhetorical questions — not yet detected as a distinct category.
• Quoted/reported speech attribution (whose opinion is it?) — not handled.
• Word sense disambiguation ("sick" as slang vs. illness) — lexicon uses a single fixed value per word; a WSD stage would help here.
• Idiom coverage — the bundled idiom lexicon (data/idioms.json) is a small starter set (~30 entries); swap in a larger corpus (e.g. MAGPIE) for production use.

Development

pip install -e .[dev,all]
pytest tests/