Research

ModuleQ Human-Centered AI Research Update

Presentation at the Stanford University AI symposium shows the power of combining machine learning and deep domain knowledge.


Last month, ModuleQ founders Dr. Anupriya Ankolekar and Dr. David Brunner presented our People-Facing AI for business professionals at the AAAI-MAKE symposium. Held at Stanford University, the symposium is dedicated to novel hybrid AI systems that combine machine learning and knowledge engineering.

ModuleQ’s People-Facing AI demonstrates the value of domain-tuned hybrid AI for augmenting customer-facing professionals by recommending timely, relevant business news through proactive alerts in Microsoft Teams.

Cutting through information overload

We designed People-Facing AI to surface all kinds of business insights—from sales collateral to internal research to predictive analytics—but we focused first on the news because news provides crucial context to inform important business decisions. Every day, every professional turns to the news to orient themselves and re-evaluate their priorities. Unfortunately, keeping up with the immense volume, complexity, and interdependence of relevant news and information under time pressure creates intense information overload. AI augmentation can relieve professionals by surfacing meaningful news from the long tail of 50,000+ news articles published globally every day. Intelligent news recommenders working on behalf of the user can evaluate every single article, in real-time, in relation to a dynamic, continuously-updated AI model of each user’s current work priorities and business relationships.

Recommending valuable news to business professionals is a hard problem in general, but well-suited to the application of hybrid AI. Business professionals need a concise set of highly targeted, easily understood recommendations, delivered where they work. These users have little time to configure or train a system and low tolerance for irrelevant news. In addition, the confidential and sensitive nature of their work data creates a scarcity of training data for pure machine learning approaches. Most existing news recommendation systems are not targeted enough to deliver valuable recommendations for the highly specific, fast-changing needs of B2B professionals.

Domain tuning enables explainable AI with instant value

By adopting a domain-tuned, knowledge-engineering approach, ModuleQ’s People-Facing AI can deliver highly targeted and explainable recommendations without the need for burdensome configuration, providing value to users within minutes.

Domain tuning is essential for successful AI augmentation. Domain tuning refers to deep machine understanding of the user’s domain in terms of key entities and engagement patterns, and the application of hand-crafted, domain-specific rules and heuristics. Beyond the explainability, logic, and debuggability provided by a knowledge-engineering approach, domain-tuning enables the creation of intelligent support in professional contexts where the training data is sensitive and/or limited. Extending and customizing the system to support new professional roles and industries is possible with the infusion of domain-specific entities and heuristics.

Traditional, pure knowledge engineering approaches often encounter scalability challenges. Conversely, pure machine learning approaches need enormous volumes of training data and lack explainability. Domain tuning offers the best of both worlds: judicious injection of domain knowledge can bootstrap and accelerate machine learning, while helping provide an overarching context in which to situate and interpret the output of ML models.

Secure & compliant human-centered AI augmentation

ModuleQ’s People-Facing AI uses domain tuning in two primary ways: (1) to understand people’s interests and priorities from their professional work data (via our patented Personal Data Fusion technology) and (2) to deliver recommendations that are highly targeted to the professional’s interests and work domain. Our system employs a federated architecture with a central knowledge graph that captures domain-relevant entities and relationships. Machine learning and NLP techniques are used alongside domain heuristics to identify domain-relevant entities and relationships in the raw user data and content, and to recommend the most professionally relevant content for users. Decentralized learning components of the platform are deployed into each organization’s cloud estate, so sensitive data stays within established security and compliance regimes.

The results provide strong validation of our approach to human-centered AI augmentation. Daily and monthly active user metrics show high, sustained engagement. The usefulness of the news recommendations, measured by actual user feedback on specific news articles, clearly demonstrated the power of our domain-tuning techniques. Furthermore, our simple chatbot UI in Microsoft Teams proved to be an effective means of integrating AI into the workflow of busy professionals.

To learn more, our paper is published in the symposium proceedings here: Hybrid AI System Delivering Highly Targeted News to Business Professionals

We gratefully acknowledge the contributions LSEG Labs (formerly Refinitiv Labs) in helping tune our AI for finance professionals.