Extractive Summarization

The core mechanism assigns a numerical importance score to each sentence. Algorithms like TextRank (a graph-based PageRank adaptation) or TF-IDF vectors compute centrality based on inter-sentence similarity. - Graph Methods: Sentences are nodes; edges represent similarity. Highly connected nodes are extracted. - Feature-Based Methods: Score sentences based on position, length, presence of cue phrases, and term frequency. The top-k highest-scoring sentences are selected verbatim to form the summary.

A critical failure mode is selecting sentences that repeat the same information. Maximum Marginal Relevance (MMR) is a greedy algorithm that balances relevance against novelty. - It iteratively selects sentences that are highly relevant to the query but minimally similar to already-selected sentences. - This ensures the final summary maximizes information coverage and avoids the 'echo effect' of duplicated facts.

Extractive systems depend on accurate text segmentation. The pipeline must correctly identify sentence boundaries, even in noisy text with abbreviations, decimals, or bullet points. - A segmentation error (e.g., splitting 'Dr. Smith' into two sentences) propagates downstream, causing fragmented or nonsensical extracted units. - Robust NLP preprocessing using rule-based or statistical models is a prerequisite for coherent extraction.

In well-structured documents like news articles or technical reports, sentence position is a strong heuristic for importance. - Lead-3 Baseline: A surprisingly strong summarization method that simply extracts the first three sentences of a document. - Extractive models often incorporate positional features, assigning higher prior weights to sentences in introductions, abstracts, and conclusions, reflecting the document's discourse structure.

Because sentences are lifted verbatim from different contexts, the resulting summary often suffers from coherence breakdowns. - Dangling pronouns ('It', 'He') may lack their antecedent. - Discourse markers ('However', 'Furthermore') may connect to absent prior context. - Repairing these artifacts requires post-processing or falls outside the scope of pure extraction, representing a key limitation compared to abstractive methods.

The primary advantage of extraction is zero hallucination at the sentence level. Since output strings are exact copies of source text, they cannot introduce factual errors or fabricated entities. - This makes extractive methods highly suitable for legal, medical, and compliance domains where verbatim accuracy is non-negotiable. - The risk shifts from fabrication to contextual omission, where removing a sentence from its surrounding context changes its implied meaning.

Generates novel phrasing rather than copying sentences verbatim. Unlike extractive methods, this approach can rephrase, paraphrase, and compress information—often producing more fluent summaries. Modern implementations use sequence-to-sequence transformers fine-tuned on summarization datasets. The trade-off is higher hallucination risk, as generated text may introduce facts not present in source documents.

A greedy algorithm that selects sentences by balancing two competing objectives:

• Relevance: How similar a candidate sentence is to the query or document centroid
• Novelty: How dissimilar it is to sentences already selected This redundancy penalty prevents extractive summaries from repeating the same information, producing more information-dense output. Widely used in multi-document summarization pipelines.

The computational task of assigning importance scores to sentences or text spans. Common signals include:

• TF-IDF and centroid similarity
• Graph-based centrality (TextRank, LexRank)
• Positional heuristics (lead bias in news)
• Supervised classifiers trained on human summaries Salience ranking is the core scoring mechanism that determines which sentences an extractive summarizer selects.

Recall-Oriented Understudy for Gisting Evaluation using the Longest Common Subsequence. Measures how well a candidate summary captures the structural word sequence of a reference summary. Unlike ROUGE-N (n-gram overlap), ROUGE-L accounts for sentence-level word order and naturally handles variable-length sequences. The F1 variant balances precision and recall for extractive output assessment.

Quantifies whether a summary is factually grounded in source documents. For extractive methods, faithfulness is inherently high since sentences are copied verbatim. However, context truncation can create misleading implications. Modern faithfulness evaluation uses:

• Natural Language Inference models to detect contradictions
• Factual Consistency Scoring against source spans
• Entailment ratio calculations per sentence

Generates summaries tailored to a specific information need rather than providing a generic overview. The query acts as a relevance filter—only sentences addressing the user's question are candidates for extraction. This transforms extractive summarization from a document-centric to a user-centric task, commonly used in search engines and question-answering systems.