Extractive Summarization

The defining mechanism of extractive summarization is the direct copying of source sentences. Unlike abstractive methods, no new words or paraphrases are generated. This guarantees factual consistency by design, as the output is a subset of the input. The process relies on a selection function that scores each sentence for salience, often using Maximum Marginal Relevance (MMR) to balance importance against redundancy. This makes the technique ideal for legal and medical domains where precise wording is non-negotiable.

The core computational task is ranking sentences by importance. Algorithms calculate a salience score for each sentence based on features such as:

• Term Frequency-Inverse Document Frequency (TF-IDF): Weighs the importance of words relative to the document.
• Positional Bias: Sentences near the beginning (primacy) or end (recency) often receive higher weights.
• Entity Density: The concentration of named entities (people, organizations, locations) in a sentence.
• Centroid Similarity: Proximity to the document's central theme vector.

A naive salience ranking often selects sentences that repeat the same information. To produce a comprehensive summary, extractive systems apply a redundancy penalty. The Maximum Marginal Relevance (MMR) algorithm explicitly subtracts the similarity of a candidate sentence to already-selected sentences from its relevance score. This diversity constraint ensures the final summary covers distinct aspects of the source document, maximizing information coverage within a fixed token budget.

Modern extractive systems often use graph algorithms like TextRank, inspired by Google's PageRank. The document is modeled as a graph where nodes are sentences and edges represent semantic similarity. A recursive algorithm scores each node based on the number and quality of its connections. Sentences that are highly similar to many other important sentences are considered central to the document's meaning and are extracted. This approach is fully unsupervised and language-agnostic.

A primary limitation of extractive summarization is discourse incoherence. Because sentences are lifted verbatim from different parts of the source, the resulting summary often suffers from:

• Dangling Anaphora: Pronouns like 'it' or 'they' that refer to entities not included in the extracted text.
• Abrupt Transitions: A lack of logical connectors between juxtaposed sentences.
• Redundant Information: Despite MMR, some overlap may persist. This makes the output feel choppy compared to fluent abstractive summaries.

Extractive summarization can be constrained by a user query to perform query-focused summarization. Instead of a generic overview, the salience scoring function is modified to prioritize sentences with high cosine similarity to the query embedding. This ensures the extracted text directly answers a specific question, making it a foundational technique for early Retrieval-Augmented Generation (RAG) systems and search engine snippets that highlight relevant passages.

The generative counterpart to extraction. Instead of copying sentences, this technique paraphrases and rephrases the source content using novel words. It relies on sequence-to-sequence models to generate fluent, human-like summaries that may contain recombined concepts not found verbatim in the original text.

• Key difference: Generates new text vs. selecting existing text
• Risk: Higher propensity for factual hallucination
• Model example: PEGASUS, BART, T5

A greedy algorithm foundational to extractive summarization that selects sentences by balancing relevance against redundancy. It iteratively adds the passage most similar to the query while penalizing passages that overlap with already-selected content.

• Formula: λ * Relevance - (1-λ) * MaxSimilarity
• Purpose: Produces a diverse, non-repetitive summary
• Origin: Carbonell & Goldstein, 1998

The computational process of predicting which words, phrases, or entities in a text are most important to its central topic. Modern models use cross-attention scores or graph centrality algorithms to rank sentences by their semantic importance.

• Graph-based: TextRank applies PageRank to sentence similarity graphs
• Neural: BERT-based models fine-tuned on summary datasets
• Feature-based: Position, term frequency, and entity density

A critical evaluation metric that verifies whether an extracted summary contains only statements directly supported by the source document. Since extractive methods copy verbatim, they are inherently more factually grounded than abstractive methods.

• Measurement: FactCC, SummaC, or natural language inference models
• Advantage of extraction: Zero hallucination risk from paraphrasing
• Limitation: Can still misrepresent context through decontextualization

A scoring mechanism that explicitly reduces the weight of a candidate sentence if it is highly similar to content already selected for the summary. This prevents the extractive algorithm from repeatedly selecting the same information expressed in slightly different ways.

• Implementation: Cosine similarity thresholding on sentence embeddings
• Common approach: Subtract max pairwise similarity from relevance score
• Result: Maximizes information coverage in limited token budgets

An advanced NLP task that synthesizes a single, coherent summary from multiple source documents. Extractive systems must resolve redundancy and contradiction across sources, often using clustering to group related sentences before selection.

• Challenge: Cross-document coreference resolution
• Technique: Hierarchical clustering of sentence embeddings
• Application: News aggregation, systematic literature reviews