Abstractive summarization interprets a document and produces a shorter version in new wording. Unlike extractive summarization, it can paraphrase ideas, change sentence structure, and convey overall meaning, sentiment, or tone. section 6.3.2
It is especially useful for long narratives, books, and detailed discussions where quoting individual sentences would miss the broader essence. Technical documents may benefit from combining extraction of important terms with abstractive synthesis. section 6.3.4
Customer-review summaries
A frequency-based method may reduce a review to isolated words such as “received,” “well,” and “before.” An abstractive system can instead express the idea as “Early delivery,” or describe a mislabeled but effective device as “Efficient, mislabeled device.” section 6.3.4
Table 6.4 Text summarization: Generative AI vs. traditional NLP methods
|
Factor
|
Generative AI
|
Traditional NLP methods
|
|---|---|---|
|
Summary type
|
Extractive
• When prompted correctly can generate extractive summaries, however it tends to drift toward abstractive summaries
Abstractive
• Processes the entire document, understands the key points and relationships, and generates a summary that conveys the essential meaning in its own words
• Can rephrase ideas, use synonyms, and change the sentence structure
• Can incorporate the overall sentiment or tone of the original text, or change it on demand according to the needs
|
Extractive
• Select the most important words and sentences
• Prioritize factual accuracy
• Present the main ideas in a clear and concise way
Abstractive
• Predefined templates are used to generate summaries
• Typically does not generate entirely new text to create a summary
• May not be able to capture the overall sentiment
|
|
Computational efficiency
|
Low
• Computationally expensive training and inference
|
High
• Faster training and inference
|
|
Scalability
|
Low
• May struggle with long documents
• May require significant resources to process complex summaries
• Using an API is slow and expensive
|
High
• Scales well with larger datasets
• Relatively low cost of usage
|
|
Interpretability
|
Low
• Difficult to understand how the model arrives at a summary
• Like a black box
|
High
• Clear picture of chosen sentences
|
|
Model optimization
|
Complex
• Requires complex techniques like reinforcement learning to optimize for specific objectives (e.g., factual accuracy, sentiment analysis)
• Prompt engineering may help at setting summary objectives, but it is not fully deterministic and requires monitoring
|
Simple
• Can be optimized with simpler techniques like sentence scoring functions that prioritize factual keywords or specific emotions
|
Compare abstractive and extractive summarization across summary type, wording, control, and factual-accuracy tradeoffs. section 6.3.2
The flexibility has costs: large generative-AI deployments can become expensive, slow, and dependent on third-party providers. Validate summaries against the original records when accuracy matters. section 6.3.2 section 6.4
