Extractive and abstractive Summarization are two well-known methods of summarising a text.  

A summary can have a blend of both or be one or the other, but they have distinct critical features that outline the difference and mark the use cases for which one or the another is best fitting.

Extractive Summarization selects near-exact sentences from the input text to form the summary. The architecture acts similarly to a binary classification problem for each sentence in the text. The goal, then, is to give a yes or no to extracting that sentence into the summary. Deciding what sentences matter is relative and domain-specific, which is why you'll see variations for different use cases.  

Abstractive Summarization works to generate a paraphrasing-type summary that takes all information in the input text into account to form the outline. Abstractive Summarization differs from extractive as it does not look to use exact sentences from the text but a concise summary of everything. ‍ The common architectures follow the transformer-focused approach we see in extractive Summarization with fine-tuning for specific domains such as research papers, financial reports, and legal documents. These models must usually be able to support the thousands of tokens well, so approaches where Longformer techniques are mixed with high-level transformers are required.  

When coming to complex documents, as a rule of thumb, Abstractive Summarization works much better than an Extractive one. However, there are caveats. Suppose you need to summarize a legal document, and certain entities must appear in the Summarization. With the generative approach of Abstractive Summarization, nothing guarantees that all relevant entities are present in the Summarization, the Extractive Summarization will put entire sentences that encompass the entities in summary. However, it's possible to do better as the Abstract Summarization could miss essential entities, and the Extractive one will progressively look like a collection of phrases without a clear, succinct, logical glue.  

To overcome this problem, in KAPTO, we used novel AI models conceptually mixing the capabilities of entity mapping and recognition with abstract Summarization in a frame of reinforcement learning to train domain-specific models that harness the power of the generative summarization model within a framework that guarantees vital information is always present in the final summary.

‍

‍Why summaries are critical to document intelligence-driven processes

Once KAPTO is deeply automating the document-rooted process in an almost touchless way, and this is an outstanding achievement, another question arises. Who knows what is going on? Yes, process analytics can tell you how well the process is going, but the real question is: who knows what is going on at the process level? As a leading example, we can take the insurance legal workflow process. Here, even if the claim adjuster somewhat demands the actual management of the litigation process, general know-how about the judicial processes in which the company is involved is retained in the legal office. However, once you deeply automate the document recognition, the pairing with the claim, almost no human needs to read the actual document, and then the general knowledge is lost. KAPTO solve this by creating expert process agents that can, at any moment, summarize which is the status of each claim that has, in one way or another, a legal process attached, and the first step to perform this task is to have a piece of succinct but focused information about the content of each relevant document. This is why entity-constrained abstract Summarization is so essential.