Using AI to Extract Value from Unstructured Information

Organizations are collecting and generating information at an astonishing rate. At least 80% of all data is unstructured, in the form of documents, images, and video files. Unlocking the vast intellectual capital contained within this unstructured text, imagery and voice data is a challenge, and while new analytics techniques have helped, large gaps still exist.

PARC’s AI business focuses on solving high-value problems that require understanding this unstructured content. Targeting specific business needs that have large impact for organizations, we deliver toolsets via an AI-as-a Service platform model, enabling a range of deployments designed to augment people in their work.

Knowledge workers spend an estimated 15-35% of their time searching through unstructured data. But when AI makes relevant information available at the time of need, employees work more efficiently, saving cost, and experience greater satisfaction and less frustration. Our solutions help people find content and leverage insights in remote field service, customer support, document analysis, research inquiries, and basic knowledge work.

We take a differentiated hybrid approach to our solutions, combining machine-learning methods with domain specific models, and applying our core technical capabilities:

• Natural language processing (NLP) for document understanding
• Computer vision (CV) for specialized image and video recognition
• Augmented reality (AR) for visualizing information in context
• Human computer teaming with an expertise in explainability

Combining AI and AR to improve operational problem solving

We are creating solutions that empower field workers with intelligence at their fingertips to complete tasks on their own with greater accuracy. By applying AI models in concert with AR, we are augmenting the intelligence of service teams to uniquely up-skill knowledge gaps, while greatly improving competency and efficiency levels. For example, contextual data and interactive instructions can be overlaid on 3D objects or assets, displaying information needed to diagnose and solve problems at the site, without support from more highly trained technicians and without delay as problems occur.

 

A more human understanding of documents

In NLP, we are solving problems requiring higher levels of relevance that also support better generalizability. Some language models have quirky failures due to their massive, non-human-like training, limiting their application in business-critical functions. Other failures indicate that the systems are not generalizing from the data in the way people do. We are working on strategies to adjust the training regime to force these types of systems to structure their analyses in a more appropriate way. We are also considering alternative architectures for learning linguistic abilities.

No large labeled dataset? No problem.

Large labeled datasets can be impractical to acquire, especially in specialized, high-value domains like augmented reality assistance, medical imaging, video mining, and remote visual coaching. Our models and processes aim to assist knowledge workers or to automate tasks with limited datasets.

AI as a collaborative partner

Our work is designed to improve the lives of individuals and teams involved in specific tasks or multifaceted workflows. Technologies that use NLP, image/video analytics, augmented reality, machine learning, and knowledge graphs can answer timely questions, discover insights, and provide recommendations on demand. Our systems hypothesize and formulate possible answers based on available evidence. Our models are trained and can adapt from data, as well as learn from their errors through re-training or human supervision.

These approaches are enabling us to tackle an array of high-return use cases that are too narrow for machine learning-only models to solve accurately.

Use cases:

• Advanced computer vision technologies can be used to train AI systems to understand images and video, search across video clips, and provide automated guidance to field service technicians using augmented reality.
• Our NLP-powered document understanding helps knowledge workers understand, analyze, and create critical documents such as contracts and proposals through intelligent search and easy reuse of existing information. 
• In a new direction, we are currently exploring greenhouse gas monitoring by jointly optimizing ground sensor placement and machine learning-based reconstruction of greenhouse gases maps. This line of research requires our joint hardware and software expertise at PARC.

From research to market – a platform for partners

PARC has been developing the science of AI for more than 40 years and is now working with dozens of organizations to apply sophisticated AI systems to a variety of problems.

Our platform approach enables commercial customers to deliver multiple innovative solutions in enterprise applications, customer service operations, and an array of other areas.

We have been awarded numerous DARPA contracts, most recently to develop Explainable AI technologies in partnership with several leading universities.

The future of AI is about a deep, transparent understanding between humans and machines, and at PARC, we’re excited to keep charting the course to this collaborative future.