Text Summarization

NLP interview extractive abstractive seq2seq

Extractive vs abstractive — what's the difference?

Extractive summarization selects the most important sentences verbatim; abstractive generates new sentences that paraphrase the meaning, like a human would.

Extractive
  • Picks/ranks existing sentences — always factual to the source.
  • Can read choppy and redundant; can't condense or rephrase.
Abstractive
  • Fluent, concise, can merge ideas across sentences.
  • Can hallucinate facts not in the source.

How is each one modeled?

Extractive is essentially sentence ranking/classification (which sentences to keep) — classically TextRank, now transformer encoders. Abstractive is a sequence-to-sequence generation task, today handled by encoder-decoder transformers (BART, T5, Pegasus) or decoder-only LLMs.

Follow-ups
  • “Why did abstractive take off recently?” — Transformers and large pretraining made fluent, coherent generation reliable enough for real use.

What are the factuality / hallucination risks?

Abstractive models can produce fluent summaries that contradict or invent details — a serious problem in news, medical, or legal text. See hallucinations.

Mitigations

Ground generation in the source (constrained decoding, copy mechanisms), add a factual-consistency check, or use extractive/hybrid methods when faithfulness matters more than fluency.

How do you summarize documents longer than the context?

When a document exceeds the model's context window, you can't feed it all at once.

Chunk & merge map-reduce

Summarize chunks, then summarize the summaries.

Refine running summary

Iteratively update a summary as you scan chunks.

Retrieve first salient chunks

Pull the most relevant sections, then summarize those.

How do you evaluate a summary?

ROUGE (n-gram/longest-subsequence overlap with reference summaries) is the classic metric, but it only measures word overlap — not factuality or coherence.

Follow-ups
  • “What does ROUGE miss?” — A faithful paraphrase with different words scores low; a fluent but wrong summary can score high. Pair it with factual-consistency metrics, LLM-as-a-judge, or human review.