Extractive vs Abstractive Summarization Compared

How Extractive Summarization Works

Extractive systems treat summarization as a sentence selection problem. Given a document with 200 sentences, the task is to select the 10-20 sentences that, together, best represent the full document. The output is always a subset of the input, using the exact original wording.

Classical algorithms like TextRank build a graph where each sentence is a node and edges represent similarity between sentences (measured by shared words or embedding similarity). Sentences that are similar to many other sentences score highest, because they represent common themes. The top-scoring sentences form the summary. This runs in milliseconds, requires no API calls, and works offline.

Modern extractive approaches use transformer encoders (BERT, RoBERTa) to classify each sentence as "include" or "exclude" based on contextual understanding of the full document. These models learn which sentences carry key information by training on document-summary pairs where the training signal indicates which source sentences appear in human-written summaries. This achieves better selection quality than graph-based methods but requires a trained model.

LLMs can also perform extractive summarization when instructed to "select the most important sentences verbatim without modification." This gives you the intelligence of an LLM (understanding what is important) combined with the safety of extractive methods (no hallucination risk). The tradeoff is that you pay LLM API costs for what is fundamentally a selection task that cheaper models handle adequately.

How Abstractive Summarization Works

Abstractive systems generate new text that has never appeared in the source document. An LLM reads the source, constructs an internal representation of its meaning, then generates a compressed version using its own words, sentence structures, and organization. The output may combine ideas from different paragraphs into a single sentence, substitute technical terms with simpler equivalents, generalize specific examples into broader claims, or restructure information into a more logical sequence than the original.

The power of abstraction enables compression ratios that extraction can never achieve. An extractive summary of a 5,000-word article is limited by the shortest meaningful sentences in that article, typically producing 500-1,000 word summaries (10-20% compression). An abstractive summary can compress the same article to 100 words (2% of original) by combining multiple facts into dense sentences and expressing ideas more concisely than the original author did.

LLMs dominate abstractive summarization in 2026. GPT-4o, Claude Sonnet, and Gemini 1.5 Pro all produce fluent, well-structured abstractive summaries with minimal prompting. The quality difference between these frontier models is small for typical summarization tasks; the choice between them depends more on pricing, context window size, and API ecosystem fit than raw summarization ability.

Accuracy and Hallucination Comparison

The most important practical difference is factual reliability.

Extractive summaries have a 0% hallucination rate by definition. Every word in the output exists in the source. If the source says "revenue grew 12%," the extractive summary says "revenue grew 12%" verbatim or does not mention revenue at all. There is no mechanism for introducing false information because the system never generates new text.

Abstractive summaries hallucinate at measurable rates. Research consistently shows that LLM-generated summaries contain unsupported claims 2-15% of the time, depending on the model, source complexity, and compression ratio. Common hallucination patterns in summarization include: slightly altered statistics ("revenue grew 15%" when the source says "12%"), fabricated attributions ("according to the CEO" when no such quote exists), merged claims from different entities, and invented causal relationships between correlated facts.

For applications where factual errors carry serious consequences (legal document summarization, medical record summaries, financial reporting, regulatory compliance), extractive methods provide a hard guarantee that abstractive methods cannot match. For applications where occasional minor inaccuracies are tolerable (meeting notes, content curation, educational summaries), abstractive methods produce far more useful output despite the hallucination risk.

Compression and Fluency Comparison

Extractive summaries suffer from two coherence problems. First, selected sentences were written as part of paragraphs with surrounding context. Pronouns may reference entities introduced in unselected sentences ("This approach works well..." but "this approach" was described in a sentence the system did not select). Second, the logical flow between selected sentences is often poor because they were chosen independently for importance, not for narrative coherence.

Abstractive summaries read like they were written as summaries from the start. The model generates transitions between ideas, resolves pronoun references, maintains consistent terminology, and organizes information in a logical sequence. For human readers, abstractive summaries are almost always more pleasant and easier to understand than extractive ones of equivalent length.

Compression ratio differences are substantial. Extractive methods typically achieve 10-30% of original length as a practical minimum before coherence degrades severely. Abstractive methods routinely achieve 2-10% of original length while remaining coherent and informative. For memory systems that need to compress 10,000 tokens of conversation history into 200 tokens, only abstractive methods can achieve the necessary compression.

Cost and Latency Comparison

Extractive summarization using classical algorithms (TextRank, LSA) runs in milliseconds on commodity hardware with zero API cost. A Python implementation of TextRank processes a 10,000-word document in under 100 milliseconds. This makes extraction ideal for high-throughput pipelines processing millions of documents where per-document cost must be negligible.

Extractive summarization using transformer classifiers costs more (GPU inference or API calls to smaller models) but still runs significantly faster and cheaper than LLM-based abstraction. A BERT-based extractive model processes documents in 50-200 milliseconds with inference costs under $0.0001 per document.

Abstractive summarization with frontier LLMs costs $0.005-$0.05 per document depending on input length and model choice. Processing takes 1-5 seconds per document. At 100,000 documents per month, abstractive summarization costs $500-$5,000, while extractive methods cost $0-$10 for the same volume. This cost difference is the primary reason production systems do not use LLM abstraction for every summarization task.

When to Use Each Approach

Use extractive when: Factual accuracy is non-negotiable and any hallucination is unacceptable. Legal documents, medical records, financial statements, and regulatory filings. When you need citation traceability (pointing to exact source locations). When processing volume is massive and per-document cost must be near zero. When latency requirements are under 200 milliseconds. When the source text is already well-written and concise enough that its own sentences serve as adequate summary content.

Use abstractive when: Output quality and readability matter more than perfect factual fidelity. Meeting summaries, conversation compression, content curation, educational materials. When compression ratios need to exceed what extraction achieves (target under 10% of original). When the summary needs restructuring (grouping related topics that appear scattered in the source). When the audience differs from the source's intended audience (summarizing a technical paper for executives).

Use hybrid when: You need the quality of abstractive output with reduced hallucination risk. The hybrid pattern: use extractive methods to select the K most important passages from the source, then pass those selected passages to an LLM for abstractive summarization. The extractive step ensures the model works only with the most relevant source material, reducing distractions that cause hallucination. The abstractive step produces fluent, compressed output. This approach reduces hallucination rates by 40-60% compared to full-document abstractive summarization while maintaining comparable output quality.

Implementation Patterns for Each Approach

Extractive pipeline: Segment text into sentences using a sentence boundary detector (spaCy, NLTK, or regex for simple cases). Generate sentence embeddings using a fast model (sentence-transformers/all-MiniLM-L6-v2 runs in 5ms per sentence). Score sentences by centrality (average similarity to all other sentences), position (earlier sentences get a boost), and length (filter out very short or very long sentences). Select top-K sentences maintaining original order. Optionally post-process to resolve dangling references.

Abstractive pipeline: Check if source fits within context window. If yes, send directly with summarization prompt. If no, apply map-reduce or refine pattern (see the long document summarization article). Include explicit instructions for length, format, audience, and what to prioritize. Optionally run hallucination detection on output. Cache result for repeated requests.

Hybrid pipeline: Run extractive selection to identify top 30-50% of source by importance. Pass selected passages (maintaining original order) to LLM with abstractive prompt. Include instruction: "Summarize the following passages. All information in your summary must come from these passages." This instruction reinforces faithfulness. Run entailment verification on output against selected passages.