Text Summarization: Extractive vs. Abstractive

Two fundamentally different strategies

Automatic text summarization compresses a longer document into a shorter one that preserves its most important content. There are two fundamentally different ways to do this. Extractive summarization selects a subset of existing sentences (or phrases) directly from the source document and stitches them together, unchanged, as the summary. Abstractive summarization generates new sentences that may never appear verbatim in the source, paraphrasing and synthesizing information the way a human would when asked to summarize something in their own words.

The distinction matters because the two approaches have different failure modes, different evaluation challenges, and were historically solved with entirely different techniques — though modern transformer-based systems have blurred the line somewhat.

Extractive summarization: picking the best sentences

Extractive methods treat summarization as a sentence-ranking and selection problem: score every sentence in the document by how representative or important it is, then select the top-scoring sentences (subject to a length budget) and output them in their original order. Because the output sentences are copied verbatim, extractive summaries are always grammatically correct and factually consistent with the source — they can't "hallucinate" a fact that wasn't in the document, because they never generate new text at all.

The classic algorithm here is TextRank, a graph-based method inspired by PageRank: build a graph where sentences are nodes and edges are weighted by sentence similarity, then run a random-walk ranking algorithm to find the most "central" sentences — the ones that are similar to many other important sentences, and therefore likely represent the document's core content. Modern extractive systems instead train a neural classifier (often built on a transformer encoder like BERT) to directly predict, for each sentence, a binary label: should this sentence be included in the summary or not.

Abstractive summarization: generating new text

Abstractive methods frame summarization as a sequence-to-sequence generation task: feed the source document into an encoder, and have a decoder generate a summary word by word, conditioned on the encoded source. This is exactly the architecture used for machine translation, and the same encoder-decoder transformer models (BART, T5, Pegasus) that excel at translation have become the standard tools for abstractive summarization, simply fine-tuned on document-summary pairs instead of translation pairs.

Abstractive summaries read more fluently and can be more concise, because the model can compress multiple sentences' worth of information into one sentence, paraphrase awkward source phrasing, and connect ideas the original document never stated side by side. The cost is that generation introduces the possibility of hallucination — the model can produce a fact that sounds plausible but wasn't actually in the source document, which is a serious problem for any application where factual accuracy matters, like summarizing medical records or legal filings.

Evaluating summaries: ROUGE

The standard automatic metric for summarization is ROUGE (Recall-Oriented Understudy for Gisting Evaluation), which compares a generated summary against one or more human-written reference summaries by measuring n-gram overlap. ROUGE-1 measures unigram (single word) overlap, ROUGE-2 measures bigram overlap, and ROUGE-L measures the longest common subsequence between generated and reference summaries. Each is typically reported as an F1 score balancing precision (what fraction of the generated summary's n-grams appear in the reference) and recall (what fraction of the reference's n-grams appear in the generated summary).

ROUGE is cheap to compute and correlates reasonably well with human judgment for extractive summaries, where overlapping wording is expected. It correlates much less reliably for abstractive summaries, because a summary can be an excellent paraphrase that uses almost entirely different words from the reference and still score poorly, while a summary that copies reference phrasing verbatim scores well regardless of whether it's actually a good summary of the source document.

Extractive vs. abstractive summarization

Where this leaves production systems

In practice, many production summarization systems hedge between the two strategies. A common pattern is to extract the most salient sentences first, then use an abstractive model to compress and smooth that extracted content into a shorter, more fluent final summary — getting some of the faithfulness guarantees of extraction with some of the fluency of generation. Evaluation in production also increasingly supplements ROUGE with LLM-based judges or targeted factuality checks (does every number and named entity in the summary actually appear in the source?), since ROUGE alone is a weak signal for abstractive quality.

What's next

Abstractive summarization is really a special case of sequence-to-sequence generation — the same underlying machinery used to translate between languages. The next lesson traces how machine translation evolved from rule-based systems through statistical phrase-based MT to the sequence-to-sequence and transformer models used today.