Lucy’s Technology Accelerates Research and Document Discovery

RahulEquals 3’s Chief Product Officer, Rahul Singhal, shares how Lucy’s deep learning models have helped Lucy evolve and prepare her for new challenges.

It was just two years ago when we launched Lucy—the AI-powered marketing assistant. One of the first use cases that she solved was in helping marketers uncover insights across the vast amounts of data they own and license. In this use case, Lucy works as a research assistant, an extraordinary one who can read through terabytes of data in seconds. A marketer asks Lucy a question such as:

  • Can you give me a SWOT analysis for Tesla?
  • How much time do different age groups spend watching TV?
  • What are the top sports watched by buyers of Nike?
  • How much did BMW spend vs Audi vs Mercedes Benz by month in 2016?
  • What are the top search terms for Gatorade

Lucy then searches for answers across all of their data, structured and unstructured, spread across disparate systems and returns with the most relevant answers. Lucy was not simply identifying search terms, she understood the intent of the questions in relation to marketing. This set her apart. How did we achieve it? Along with the human training and interaction that all AIs need to improve, we also created deep learning models to excel her learnings. Some of the key components to our models included building:

A marketing knowledge graph - Most of the out-of-box technologies have been trained on general Wikipedia or openly available datasets. They tend to do a reasonably good job when asking questions that are not domain specific. However, the traditional NLP models fail to understand different concepts such as guerrilla marketing, paid advertising or brand awareness. We created our own domain specific models to train Lucy to understand the vernacular of a marketer. Unlike most AI engines, we used a combination of deep learning models and domain specific SMEs to help us create the knowledge.

Parsers that can ingest all variety of data types - Often overlooked, ingestion parsers tend to be a critical component of an AI solution. Current ingestion parsers do not do a good job of parsing PDFs or PPTs well that could be ingested in a way to build an answer engine. Our team spent over a year doing research on different types of ingestion engines and created one that allows us to parse the data in a way that is most meaningful. This turned out to be a critical success factor for us in getting to answers.

A routing engine - Understanding the question's intent and which source is the best place to find the answer required our team to build an ontology of intents across structured and unstructured data sources. Now when a question is asked, Lucy has an intrinsic knowledge to understand the best place to find the answer and directs the question to be answered to the best possible repository.

As Lucy learns and matures, new use cases continue to be discovered. Recently we demoed Lucy’s AI-powered research capability to a Fortune 100 Aerospace Manufacturer. They were wowed by our technology and laid out a challenge for Lucy. Their maintenance workers were spending a significant amount of research time with their library of manuals in order to understand the reason for a fault code of an aeroplane. Every minute a plane was on the ground while they researched was leading to revenue loss for the airline. They wanted to engage our technology to speed up the discovery process. Could Lucy ingest all of their airline manuals and be trained to understand fault codes? Could she then return the description of code and the suggested remediation based on the information in the manuals?

In just six weeks, Lucy was able to train on the data and deliver the information in seconds with an accuracy of over 90 percent for over 92 percent of questions asked. How did she do it? This was possible by applying domain specific ontology and parsing the content, which made it easier for the system to enrich the data with meta data. Lucy’s learning models allowed her to step out of marketing research and become a Knowledge Management Platform for the maintenance workers. Her accuracy will continue to improve as she works with the maintenance team.

Manuals

It is amazing to see how quickly Lucy is learning. As she develops, new use cases continue to be explored. She is ready to take on other document discovery challenges such as RFP responses, training information, technical documentation, manuals, or sales decks. If these are document issues you are struggling with, or if your organization has a unique knowledge management challenge, reach out and let’s explore if Lucy can be of assistance.

 

Rahul Singhal, Chief Product Officer, Equals 3

Rahul Singhal is responsible for the overall product strategy and roadmap guiding the ongoing evolution of Lucy, Equals 3’s AI assistant to the marketing professional. Prior to joining Equals 3, Rahul was program director for the IBM Watson Platform where he was managed a portfolio of APIs that included Vision, Speech, Data and Language. During his three years in that role, he grew the usage of the services by over 100X and launched over 15 new services. Rahul also spent 12 years at IBM in variety of senior strategy, M&A and operational roles and working for Gartner where he helped startups on their go to market strategies.

 

 

 

 

 

Uploading is Obsolete! Automated Knowledge Management is the New Standard

AI-powered knowledge management systems now have the ability to remove time-consuming, manual processes of uploading, tagging and curation using automation. See what's possible in our latest white paper.

Get Your Copy