Understanding Extractive vs Abstractive Summarization

With the large volume of digital data to process, professionals who have to deal with data management and processing find automatic text summarization an inevitable tool to navigate their day-to-day business. The sole objective of AI-driven summarization tools is to express all information in the input text in a vivid, concise, and comprehensive manner, enabling users to save effort and time.

There are basically two types of AI document summarization methods – extractive vs abstractive summarization. While the former serves like a highlighter, the latter functions like a pen or pencil. A better explanation would be that the extractive method pull out relevant sentences from a passage and stich it together to give you a comprehensive summary, while the abstractive method will ‘read and understand’ the passage, and then write a summary in its own words.

Let’s look at the two AI solutions and their applications in detail for a better understanding.

How Does Extractive Summarization Work?

Extractive summarization is largely based on statistical, machine learning and unsupervised learning methods:

• Statistical Method: Statistical methods include word frequency analysis, graph-based ranking algorithms like TextRank and LexRank, and TF-IDF weighting. While word frequency analysis identifies the most common and representative words in the text, TF-IDF weighs the importance of words based on their frequency in the text document.Graph-based methods represent sentences as nodes and their similarity as edges, and then rank the sentences based on their centrality or connectivity.
• Machine Learning Methods: These methods involve training classifiers (Naïve Bayes, SVM) to categorize sentences as summary-worthy or not, or using sequence labelling models (HMM, CRF) to assign labels to each sentence indicating its role in the summary (introductory, conclusive, etc.)
• Unsupervised Learning Methods: These include techniques like using clustering models (K-mean, hierarchical), to partition sentences into coherent clusters and select the most central sentence from each cluster, or employing topics models (NMF, LDA) to discover the latent themes in the text and choose sentences that best represent each topic.

Abstractive Summarization Algorithms Explained

Abstractive summarization is based on encoder-decoder models, and transformer-based models.

• Encoder-decoder Architecture: This technique uses encoders (GRU, LSTM, Transformer) to process the input text, while the decoder (GRU, LSTM, Transformer) generates the summary. Attention mechanisms allow the decoder to focus on different parts of the input during text generation, enabling better capturing of long-range dependencies and important information.
• Pointer-generator Networks: Pointer-generator networks combine copying words from the source text with generating new words, allowing for more faithful summaries while maintaining abstractive capabilities.
• Transformer-based Models: Transformer-based models like BART (Bidirectional and Auto-Regressive Transformers), and T5 (Bidirectional and Auto-Regressive Transformers) have achieved advanced performance on abstractive summarization tasks by leveraging large-scale pre-training and fine-tuning.

Human vs Automated Summary Evaluation Methods

Summaries, either extractive or abstractive, may not be 100% precise and hence you need to adapt evaluation techniques to evaluate their consistency and accuracy. Some of the common automated evaluation techniques used are ROUGE, BLEU, and BERTScore.

• Recall-Oriented Understudy for Gisting Evaluation (ROUGE): This method measures overlap between the generated summary and reference summaries.
• BERTScore: This technique computes the cosine similarity between the contextualized word embedding of the generated and reference summaries, capturing semantic equivalence beyond exact string matching.
• Bilingual Evaluation Understudy (BLEU): This method, originally designed for machine translation but adapted for summarization, calculates the precision of n-grams in the generated summary compared to the references.

Apart from these automated methods of evaluation, there is of course, the human evaluation too, which is critical.

• To evaluate the quality of extractive summaries in terms of relevance, coverage, and readability.
• Having annotators rate the quality of summaries based on criteria like fluency, relevance, factual consistency, and coherence.
• When it comes to abstractive summaries, there could be issues with faithfulness and factuality, as models may generate content that is not supported by the original text. Hence a manual evaluation is a must here.

Such qualitative evaluation of summaries can provide insights into the strengths and weaknesses of different models and help identify areas for improvement.

Real-world Applications

As discussed above, both extractive and abstractive are great AI-driven summarization tools but their real-world applications can widely differ due to the advantages (or disadvantages) they offer.

Since extractive summarization identifies and extracts important sentences or phrases directly from the original text, they are more factual, accurate and preserve the original meaning of the text, though it can result in less coherent or natural-sounding summaries.

They are best suited to summarize:

• News articles
• Scientific papers
• Medical records
• Legal documents

Abstractive summarization uses natural language processing (NLP) to understand the input text and generate a new summary in its own words, paraphrasing the original content. It thus provides more human-like fluent, and concise summaries. Also, it is better suited for capturing the overall meaning of the original text than just capturing key sentences. The drawback, however, is that abstractive summarization tends to ‘hallucinate’ or provide inaccurate summaries.

They are best suited to summarize:

• General content like books, articles or product reviews
• Customer service reports
• Simplifying complex information
• Conversation data like meetings or dialogues

Conclusion

By adopting these automated document summary and processing techniques, one can develop summarizing tools that are driven by artificial intelligence and are capable of catering to a wide range of audiences and use cases.

Nevertheless, using the right model that caters to your needs would give your business the right edge. To know more about these AI-driven summarization tools or to take advantage of the best-in-the-industry experts in implementing these smart tools for your business, feel free to approach DeepKnit AI.