Text Summarizer

Extractive summarization: selecting, not rewriting

There are two ways to shorten a document. Abstractive summarization, the approach used by large language models, reads the text and writes new sentences that paraphrase it. Extractive summarization, which this tool performs, does not write anything new — it scores the sentences already in your text and returns the highest-ranking ones verbatim. The trade-off is clear: extractive summaries can never hallucinate a fact or invent a claim, because every sentence is your own, and the whole process is fast, deterministic, and runs without any AI model or server call. The cost is that the result is only as coherent as the original sentences allow. For articles, reports, and essays where you want a faithful, traceable digest, that is exactly the right trade.

How TF-IDF scores a sentence

The engine ranks sentences using TF-IDF — Term Frequency–Inverse Document Frequency — a classic information-retrieval measure of how characteristic a word is to a document. It works in stages:

1. Tokenize and clean. The text is split into lowercase words. Common stop words (the, is, and, of, to…) and very short words are discarded, because they appear everywhere and signal nothing about topic.
2. Stem. A light suffix-stripper folds word variants together so that running, runs, and run reinforce one another instead of being counted as three unrelated terms.
3. Term frequency. Each remaining word's count is divided by the total word count, giving its frequency in the document.
4. Inverse document frequency. Treating each sentence as a "document," a word that shows up in only a few sentences earns a higher weight than one sprinkled across many. A word's TF-IDF is its frequency multiplied by the log of this rarity factor.
5. Sentence score. Each sentence's score is the average TF-IDF of its meaningful words. Averaging rather than summing is deliberate — it stops long sentences from winning simply by containing more words.

The top-scoring sentences, however many you request, are then re-sorted into their original order before display, so the summary reads in the same sequence the author wrote it rather than in score order, which would jump around incoherently.

A worked example

Imagine a 600-word article about urban beekeeping. Words like hive, forage, pollinator, and rooftop appear several times but only in a handful of sentences, so they score high. A sentence such as "Rooftop hives give urban pollinators reliable forage through summer" packs three of those high-value terms into few words, earning a strong average score. A filler sentence like "It is worth noting that there are many things to consider" contains almost nothing but stop words and short connectors, so its score collapses toward zero. Set the slider to 5 and the tool returns the five densest sentences in document order — a tight précis built entirely from the author's own words.

Three views of the result

• Summary — the selected sentences as clean paragraphs, with a copy button.
• Highlighted Text — the full original with the chosen sentences marked in amber, so you can see exactly which parts were lifted and judge the selection in context.
• Key Phrases — the top ten multi-word phrases. The tool scans adjacent word pairs (bigrams) and triples (trigrams) that don't cross sentence boundaries, scores each by the TF-IDF of its component words, and lists the strongest. These give a topic fingerprint of the document at a glance.

A statistics row reports word count, sentence count, an estimated reading time (calculated at 200 words per minute), and a compression ratio — the percentage of sentences removed. Asking for 5 sentences from a 40-sentence article yields about an 88% compression.

Where it shines, and where it doesn't

Extractive TF-IDF summarization is strongest on well-structured informational prose: news articles, research abstracts, documentation, meeting notes, and long blog posts, where important ideas are stated in self-contained sentences. It is weaker on:

• Narrative and dialogue, where meaning is spread across sentences and pronouns, so a lifted sentence may lose its referent ("She agreed" tells you little without the prior line).
• Very short texts, where there aren't enough sentences for the statistics to separate signal from noise.
• Highly repetitive marketing copy, where the same buzzwords inflate scores without adding information.

Practical tips and common mistakes

• Tune the slider to the text length. Five sentences suits a single article; for a long report, raise it so the summary doesn't drop essential context. For a tweet-length gist, lower it to one or two.
• Clean the input first. Navigation menus, footers, and boilerplate pasted along with the article add noise sentences. Paste just the body for the sharpest result.
• Use the highlight view to sanity-check. If the amber sentences skip the actual conclusion, the article likely buries its point in a long sentence that scored lower — read the highlighted version before trusting the summary alone.
• Sentence detection relies on punctuation. Text with missing periods or unusual formatting may be split oddly, which skews scoring. Properly punctuated prose summarizes best.
• Don't expect paraphrase. The summary is always your exact sentences. If you need a reworded, condensed rewrite, that is a different (AI-based) task.

The entire analysis happens in your browser. Your text is never uploaded, which makes the tool safe for confidential drafts, internal documents, and anything you would not paste into an online AI service.