The problem with A.I. and Documents

Adobe’s VP Engineering for Document Cloud, Phil Ydens, estimates there may be up to 2.5 trillion PDF documents in the world. (2015-2016)

Can voice assistants such as Siri and Alexa help us declutter our lives by understanding our documents?

The history of humanity is etched in billions of documents. The footprints of our lives manifest in a bread crumb trail of paper, in digital and physical form. As part of our lives, we generate and receive an unprecedented volume of documents. 

Technology has delivered us practical solutions to digitize many of these statements-of-record. Humanity has made incredible strides digitizing our history. A Google search will show just how far. There is barely a subject or topic, current or historical that cannot be found with a few keystrokes. 

A.I. has given us voice-powered assistants. Alexa and Siri are powerful assistants that can answer our questions and control our homes. Voice assistants allow us to speak instead of type to get an answer. The rub is that this only works for question and answer scenarios and not when you want to go really in depth into a topic. 

As far as digital assistants have come, they cannot effectively help us with our documents. Sure, they can read an email or document out aloud. As much you may enjoy the dulcet tones of Alexa, nobody wants to spend time being read a long document, unless this a strategic move to cure insomnia. It is still easier to read the document yourself and let our non-artificial intelligence engine, our brain, do the processing. Your brain can make sense and understand a document in seconds and guide you to the relevant parts. 

As simple as documents are for people to consume, A.I. still does not understand the meaning of documents. Strides have been made in Text Summarization, but they still operate at a superficial level and do not intrinsically understand the content.

The challenge is to use machine learning, not to just read documents verbatim but to understand the document. Voice assistants should not only be able to give a quick summary of a document but also answer any questions you may have.

Consider these questions that can readily be answered by most people after reading the corresponding document: 

• What does my insurance policy cover?
• Does my insurance cover my mobile phone?
• When did we sign the contract?
• When do I get my pension?
• Why is my mobile bill so high?
• How long until I pay off my mortgage?
• Did the proposal include exclusivity for the North American market?
• What is the main idea behind the thesis? 

What we find easy, is no cake walk for A.I. today.

While this might sound like the perfect challenge for machine learning, it becomes a gargantuan task once you understand that many different types of documents contain information on every domain and subject known - past to present.

It is not insurmountable; maybe the way forward is not only improved machine learning but new methods of authoring documents. Automated systems generate many of the documents that we receive today. 

Consider your bills, contracts, policies they all created automatically by systems that have a lot more information ("metadata") than is just contained in that document. Enriching documents with metadata already available in our document production workflows makes it possible for machine learning to get to grips with your documents. 

Companies such as Pryon (https://www.pryon.ai) and InfoSlips (https://www.infoslips.com) are already turning the dream to reality. Asking your document a question and getting a relevant answer is definitely in our future.

Personally, I cannot wait for the day when I do not have to personally read every document that makes it way to me. The day when voice assistants make us more productive; free up time and improve the quality of our lives.