≡ Menu

#IBMWoW: Simplifying and Scaling Cognitive Computing with Watson


Continuing in the analyst program at IBM’s World of Watson event with Beth Smith, GM Offerings and Technology for IBM Watson, introducing some Watson elements for Conversation – one of the four C’s of Watson (Cloud, Content, Compute and Conversation).

Watson, at its core, is about finding knowledge in noisy data at enormous scale. Watson listens to signals, uses machine learning and deep learning to find patterns and then makes recommendations that it can explain. For instance, in the Conversation piece, there is Watson Conversation for developers (designed to be used with other content and services) and a configurable apps – Watson Virtual Agent for customer service (built on top of the developer service).

Both kinds of products – developer services and configurable apps – are delivered continuously as cloud solutions with new capabilities being added. This kind of development is increasingly done with the interactions the deployed services are managing.

Watson Conversation is a service that is free for developers to engage with. It comes with the tone analysis service integrated. The service has four main pillars:

  • The intents
    Each intent can have lots of strings representing different ways to say the intent. The system uses these as a base line set of ways to identify the intent but will also learn other ways to identify the intent.
  • The entities
    Can define new ones and can use system entities like date and time, percentage etc.
  • The dialog
    A flow can be defined for a dialog using steps and links
  • Improvement
    Can see interactions by intent, entities etc so can rapidly see what would help it be better.

It’s worth noting that there is no additional training step – as you add things to the definitions they are part of the system’s behavior. At any point the developer can use a try panel to see how a particular string is handled.

Watson Virtual Agent is preconfigured on top of the Conversation service with pre-defined intent, entities etc. The service is configured using a more business-oriented UI based on tiles. There are various handlers for each tile – redirect, invoke your own workspace, escalate to a human agent, give a text response. This allows the business user to configure one of the 90+ predefined cross-industry intents (there are also a number of Telco-specific ones) with the option to link to a custom solution.

The app has some lightweight metrics and reporting built in around intents and entities – for instance what are the intents that lead to human interaction most often. In addition all the data can be exported for analysis elsewhere.

Another key piece of tooling is the Watson Knowledge Studio, designed to support exploration and discovery. This is designed to move beyond these kinds of structured conversations to a more general understanding of a domain. This allows a Watson service to apply a more domain-specific view of the content. Examples are viewed and mapped to defined entities. Relations can be defined simply by linking them graphically – allowing the organization that manufactures a product to be shown, for instance. The engine will then use patterns in the example to find similar patterns and so identify additional entities. It also uses the patterns to avoid over-classifying things based on simple format or data type. Domains defined in this way can then be applied using the Watson Alchemy Language Service or Watson Explorer.

In summary, the ability to merge and layer these services, using a rich set of APIs, as well as the ability to customize their behavior and apply customer domain knowledge, are critical to scaling Watson services.

Watson also processes in a wide variety of languages and more are being added – these are not translations but learning done in a different language. In addition, there are starter kits, demos, sample code etc on the Watson Developer Cloud.