Loading page
About this tool
Strip stopwords and filler words to see the change in word count, characters and estimated tokens — and what meaning it costs.
Stopword Impact Tester removes the 179-word English stopword list distributed with NLTK, plus common prose filler like really, basically and actually, and reports exactly what that saved in words, characters and estimated tokens. Its more useful half is the cost side: negations and contrastives such as not, no, but and than are protected by default, because stripping them turns 'does not scale' into 'scale' and silently reverses what the text says. Reading time is estimated at 238 words per minute, the mean silent reading rate for English non-fiction from Brysbaert's 2019 meta-analysis.
Open Stopword Impact Tester on AltFTool — it loads instantly in your browser.
Paste or type the text you're working with.
Choose how it should be transformed or analyzed.
Copy the finished text into your document or post.
The words that flip meaning are held back and listed separately, so you can see what a naive filter would have destroyed.
Word count, character count, estimated tokens and reading time all move together, which makes the trade-off concrete.
Add domain terms to strip, or force specific words to survive, without leaving the page.
There is no single official list. The list used here is the 179-word English list distributed with NLTK, which is the one most tutorials and search pipelines start from. Other lists differ substantially: scikit-learn's older English list has 318 entries and includes words many people would not consider stopwords.
It can, badly. Roughly 68 words in the standard list carry meaning — every negation and contraction of one, plus but, if, or, than, against, only, and the directional prepositions. 'The service does not scale' becomes 'service scale', which asserts the opposite. That is why this tool protects them unless you switch the protection off.
Roughly in proportion to the characters removed. English text averages about four characters per token, so stripping 30% of the characters saves close to 30% of the tokens. These figures are rules of thumb rather than a real tokeniser, so treat them as an estimate and confirm against your model's own count before relying on the saving.
Usually not. Modern models use the function words to parse structure, and the token saving is small next to the risk of changing what you asked for. Stopword removal still earns its place in classic keyword search indexes, tag extraction and bag-of-words features, where word order is discarded anyway.