Keyword Extractor: Extract Keywords & Keyphrases from Text

This keyword extractor pulls the most important keywords and keyphrases out of any text, web page or dataset field. Paste texts or page URLs, pick a method (YAKE-style, RAKE or TF n-grams), and get ranked keywords with a score and a count for every text. It runs on plain statistics, so there is no AI, no model, no API key, and the same text always gives the same result.

It works in English, Spanish, French, German, Portuguese, Italian and Dutch and detects the language of each text on its own.

What can this keyword extractor do

• Extract keywords from text. One result row per text with keywords: [{keyword, score, count}], best first.
• Extract keywords from a URL. The main article text is pulled out of the page (menus, footers and cookie boxes are dropped) and analysed. The page title is returned too.
• Extract keywords from a dataset. Point it at the output of another Actor and name the field that holds the text (reviews, descriptions, transcripts, articles).
• Keyphrase extraction, not only single words. Choose 1 to 6 words per keyword (default 1 to 3).
• Three scoring methods plus a mix. YAKE-style (single-document statistics: casing, position, frequency, context), RAKE (phrases between stopwords), TF n-grams (repeated words and phrases), or combined.
• Control the output. Top N, minimum score, lowercase or original spelling, light plural merging, extra stopwords of your own, and removal of keywords that are just part of a better one.
• Safe at scale. Up to 5,000 items per run. One bad URL or empty text never stops the run: it gets its own row with a clear status and error.

How to extract keywords from text

1. Paste one or more texts into Texts (one entry per text).
2. Leave Method on YAKE-style and Language on auto, or change them.
3. Run it. Each text becomes one row in the dataset with its keywords.

How to extract keywords from a URL

Add page links to Web page URLs. Pages must show their text without JavaScript (blogs, news, docs, product pages usually do). Pages that need a login, are blocked for bots, or build their text in the browser return status: "failed" with the reason.

Input example

{
  "texts": ["Keyword extraction is the automated process of finding the most important words and phrases in a text."],
  "urls": ["https://www.python.org/about/"],
  "method": "yake",
  "topN": 10,
  "language": "auto",
  "minNgram": 1,
  "maxNgram": 3
}

More examples:

• {"method": "rake", "minNgram": 2, "maxNgram": 4} : only multi-word keyphrases.
• {"datasetId": "<dataset id>", "datasetField": "reviewText", "topN": 5} : top 5 keywords for every review.
• {"language": "de", "lemmatize": true, "extraStopwords": ["brand"]} : force German, merge plurals, ban a word.

Output

One dataset row per text or page:

{
  "index": 1,
  "source": "Keyword extraction is the automated process of finding the most import…",
  "sourceType": "text",
  "language": "en",
  "languageDetected": true,
  "wordCount": 84,
  "method": "yake",
  "keywords": [
    {"keyword": "keyword extraction", "score": 1.0, "count": 3},
    {"keyword": "automated process", "score": 0.85, "count": 1},
    {"keyword": "important words", "score": 0.85, "count": 1}
  ],
  "status": "ok",
  "error": null,
  "note": null
}

score runs from 0 to 1 and the best keyword of a text is always 1.0, so scores are comparable inside a text, not between texts. count is how many times the keyword appears. status is ok, no_keywords, skipped or failed. A SUMMARY record in the key-value store lists totals.

How much does it cost

Pay per event: $0.003 per text or page analysed ($3 per 1,000), event text-analyzed. Failed pages, empty texts, texts without keywords and skipped items are free. Use Max items to cap a run. Platform compute is tiny because there is no browser and no model.

Limits (honest)

• This is statistical extraction, not understanding. It ranks words by how they are used, so verbs and adjectives can slip in ("reads", "unlike machine learning"). Raise minNgram to 2, add extra stopwords, or try another method to tune it.
• Seven languages only. Other languages are analysed with English stopwords and a note.
• Very short texts (a few words) give few or poor keywords. Set the language yourself for short texts.
• Light plural merging is a simple suffix rule, not a dictionary lemmatizer.
• Web pages: no JavaScript rendering, no login pages, private network addresses are refused. PDFs and Office files are not read here.

FAQ

Is this the same as an SEO keyword research tool? No. It finds the keywords that are in your text. It does not report search volume or competition.

Which method should I use? Start with YAKE-style. Use RAKE for longer descriptive phrases, TF for long documents where repetition matters, and combined when you want a middle ground.

Does it use AI or send my text anywhere? No. Everything runs inside the Actor. Texts are not sent to any AI service. Only pages you list are downloaded.

Why do I get fewer keywords than Top N? The text may be short, or minScore, the word range or overlap removal filtered some out.

Can I feed it scraper output? Yes: pick the dataset and name the field. If a field value is only a link it is fetched as a page.

Are results the same every run? Yes. There is no randomness.

Related tools

• Google Autocomplete & Keyword Suggestions: 10 Search Engines to find what people search for.
• Extract Text from PDF: OCR, PDF to Text & Markdown to turn PDFs into text, then run them through this extractor.
• YouTube Transcript Downloader & Scraper to get video transcripts you can pass in as a dataset.