Knowledge workers spend an average of 28 hours in a typical work week on email and messaging, conversations and collaboration. Much of these communications are centered around getting organized and getting work done — define actions or tasks, make requests and commitments of each other, and exchange updates on the status of the work in progress. However, information overload remains a key obstruction to productivity. There’s a growing interest in AI-empowered collaboration tools to increase the productivity of workers in the enterprise.
The fact is that not all workers’ conversations are created equal. Some conversations individuals participate in involve actions that are directed to and of greater interest to them. Current collaboration tools are lacking in their ability to analyze conversations to identify such nuggets of information and present them to conversation participants as well as understanding nuances like who is talking to whom.
In IBM Research, we developed a novel AI method for identifying actions within conversational text, called Collaborative Language Engine (CLE). CLE can identify different classes of actionable statements such as requests, commitments and questions. It identifies some action types such as calendar and sharing, and it is extensible to learn other action types. For each action, it extracts key information such as people responsible for or involved in the action, location, time and other entities of interest. CLE is capable of being trained and learning from user interactions and feedback in a gradual, personalized and adaptive manner. This makes CLE customizable to different workers, domains or companies.
CLE was conceptualized and prototyped in IBM Research, then jointly developed by a team of researchers and IBM Watson Work engineers. This joint squad arrangement allowed us to move CLE’s innovation quickly to production as a cognitive capability for IBM Watson Workspace (Figure 1). IBM Watson Workspace offers a platform and enterprise-grade solutions for collaborations.
As a key characteristic and differentiator for the AI methods of CLE, the action identification algorithm is a white-box AI method. That is, the algorithm allows for learning complex forms of action statements as a set of independent patterns defined on a collection of syntactic and semantic level language features to create a model. The white-box approach makes the learned model inspectable, and the reasonings of CLE explainable to the person using the technology. By comparison, the so-called black-box AI methods prevalent in the industry today offer few clues as to how they arrive at a specific decision, making it difficult for workers to trust.
Figure 2 depicts the overall learning method of the CLE. The technical details of the algorithm can be found in our WWW 2017 paper, entitled “eAssistant: Cognitive Assistance for Identification and Auto-Triage of Actionable Conversations.” More broadly, CLE offers a learning framework that is capable of learning hierarchical models, which is the basis for customizing the learning for different domains, organizations and personalizing the learned models for different users.
To meet the scalability, cost and efficiency requirements of Watson Workspace, the cognitive pipeline of the CLE is architected for operational performance and can be dynamically scaled up and down. As such, it achieves an exceptionally low operational footprint with high scalability.
If you are interested to try the action identification in practice, you can experience it live in Watson Workspace. Action identification is included in the free offering of Watson Workspace. Application developers can also call the CLE directly through a published API in Watson Work, or through Watson Work apps that can register to be notified when new actions, questions, and commitments are identified in conversations.