Sentence Extractor

Pulling specific sentences from a long block of text — the first one of each paragraph, sentences containing a keyword, the longest or shortest — is exactly the kind of micro-task that takes longer to do manually than to write a tool for.

Paste your text and pick the extraction mode: by position (first, last, nth), by length (longest, shortest, above/below word count), by keyword (sentences containing a phrase), or by paragraph (one per paragraph). Sentence boundaries are detected with a tokenizer that handles common abbreviations (Dr., Mr., U.S.A.) so it doesn't split mid-abbreviation.

Non-Latin scripts and unusual punctuation work, but performance varies. The tokenizer is tuned for English; Chinese and Japanese text uses different sentence-ending punctuation, which is supported but not as polished. For specialized domains (legal text with embedded citations, scientific papers with formulas), expect occasional misdetections that need manual cleanup.

The tokenizer splits on terminal punctuation (period, question mark, exclamation point) followed by whitespace and a capital letter. The capital-letter requirement avoids splitting on abbreviations: 'Dr. Smith said yes.' is one sentence because 'Smith' is treated as continuing 'Dr.' (no whitespace between). Common abbreviation patterns (U.S.A., e.g., i.e., Mr., Ms., Mrs., Dr., Prof., vs., etc.) are exception-listed so they don't trigger false splits. Quoted dialogue stays in its containing sentence: 'He said, "Yes."' is one sentence with embedded quoted material.

The pain this addresses: pulling specific content from long documents. You have a 5,000-word transcript and need every mention of a specific term in context. Or every first sentence of every paragraph for a topical summary. Or the longest sentences (probably the most information-dense). Or the shortest (often the conclusions). Doing it manually means scrolling and copying. Doing it programmatically means writing a one-off script. The tool handles common cases without scripting.

Worked example: paste a 2,000-word article and pick 'extract sentences containing keyword: climate.' Output is the eight sentences from the article that mention 'climate' (case-insensitive) in document order. Each one comes with surrounding context if you toggle that option. Useful for fact-checking, building topical summaries, or finding all mentions of a competitor in a market report. The same approach works for finding every quoted statement, every numerical claim, or every imperative sentence (commands).

Where this can produce bad output: text with non-standard punctuation. Twitter threads with em-dashes and ellipses, transcribed speech with run-on sentences, technical writing with parenthetical citations like '(Smith et al., 2019).' The tokenizer does its best, but expect occasional errors that need manual cleanup. For specialized text (legal, medical, scientific), more sophisticated NLP tokenizers exist — spaCy, NLTK, Stanza — that handle domain-specific punctuation better. For typical English prose, the heuristic approach works well enough that the failures are rare and obvious.