Text Summarization — Extractive and Abstractive
- Type
- NLP text-generation / text-reduction task
What does it do?
Automatic text summarization is a natural-language-processing task that condenses long documents into shorter summaries while preserving their key information.
It works through one of two families of approaches — extractive summarization, which selects the most important spans from the source, or abstractive summarization, which generates new text. The field was consolidated by Nenkova and McKeown (2011), and sequence-to-sequence models such as BART (Lewis et al., 2020) advanced the abstractive side.
The intuition behind the method
Reading a long document to extract its core message is slow and does not scale across many documents. The goal of summarization is to keep the key information while making the text dramatically shorter. An extractive approach behaves like a highlighter, lifting the most informative sentences out of the source; an abstractive approach behaves more like a person rewriting the gist in fewer words. Either way, the output is a compact summary that a reader can absorb far faster than the original.
History
Automatic summarization was surveyed and consolidated by Nenkova and McKeown (2011), who framed the task of reducing text while preserving its key information across extractive and abstractive lines. The abstractive side advanced markedly with pretrained sequence-to-sequence models such as BART (Lewis et al., 2020), which use denoising pre-training to generate condensed text.
When to use it
Text summarization fits when you have documents long enough to be worth condensing and the source language is supported by the chosen approach. It suits descriptive and exploratory goals over text data, working on cross-sectional or longitudinal collections. With too few documents an abstractive model cannot be trained reliably — fall back to keyword extraction for the key information — and with no text data at all summarization cannot run.
Limitations and considerations
- Abstractive models cannot be trained reliably on very few documents.
- Performance depends on the source language being supported by the chosen approach.
- Documents must be long enough to have meaningful content to compress.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Condenses long documents into shorter summaries while preserving the key information.
- Offers two complementary approaches: extractive selection of source spans and abstractive generation of new text.
- Scales the reading effort across many or lengthy documents that would be slow to read in full.
Applications
- Producing short summaries of long reports or articles so readers absorb the gist quickly.
- Condensing large document collections for exploratory review.
- Generating abstracts or briefs from lengthy source text.
- Summarizing text corpora in humanities, social-science, business, and natural-science research.
How it works
How it works
One or more source documents of sufficient length are loaded as the input. Summarization needs text with enough content to compress, so very short fragments offer little to reduce; the method targets longer documents where condensing genuinely saves the reader time.
The raw text is cleaned and segmented into analysable units such as sentences or tokens. Because the source language must be supported by the chosen approach, this step also confirms that the document language is one the model or extractive technique can handle.
A shorter summary is produced. An extractive approach scores and selects the most important spans from the source text, while an abstractive approach uses a sequence-to-sequence model to generate new, condensed wording. The key information of the original is preserved while the length is reduced.
The summary is checked against the source for coverage of key information and for length, and outputs are inspected before the method is put to use. This guards against summaries that drop essential content or that, in the abstractive case, introduce wording not grounded in the source.
Common pitfalls
- Running an abstractive model on too small a corpus (under ~100 documents) — fall back to keyword extraction to obtain the key information instead.
- Applying summarization to a language the model or extractive technique does not support degrades the output.
- Trying to summarize when there is no real text data; with structured data, descriptive statistics are the appropriate alternative.
- Trusting a summary without checking it against the source for dropped key information or, in the abstractive case, unsupported wording.
Frequently asked questions
Extractive or abstractive — which should I use?
Extractive summarization selects and stitches together the most important spans directly from the source, so it stays faithful to the original wording. Abstractive summarization uses a sequence-to-sequence model to generate new, condensed text, which can read more naturally but requires a trained model and enough data. Extractive is the safer default on small or sensitive corpora; abstractive suits larger collections where natural rewriting matters.
How many documents do I need?
Summarization targets documents of sufficient length, with a practical minimum around 20. Training or relying on an abstractive model on fewer than roughly 100 documents is unreliable; in that case use keyword extraction to capture the key information instead.
Does it work in languages other than English?
Yes, provided the source language is supported by the chosen approach. Applying a model or extractive technique to an unsupported language is a common cause of poor summaries.
What if I have no text data?
Summarization cannot run without text. If your data is structured rather than free text, descriptive statistics are the appropriate alternative for conveying its key information.
Related methods
Sources
Nenkova, A. & McKeown, K. (2011). Automatic Summarization. Foundations and Trends in Information Retrieval.
Open DOI referenceCopied.Lewis, M. et al. (2020). BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. ACL.
Open DOI referenceCopied.