Watson Signal Services

Bricks in the AI Lego set

Year

2015-2016

Role

Design Lead

User experience

Current state

www.ibm.com/watson/developer/

My first assignment when I joined Watson Design in 2015 was to revamp a series of interactive demos to show the developer world what could be done with Watson technology. Each Signal Service is good at doing one or two things. AI happens when a few of them are used together and set loose among loads of unstructured data.

The business looked to these demos to drive customer acquisition and retention. But in context, thanks to the design team's push for user research, and developer colleagues who corroborated our insights with their daily experience, we uncovered a broader strategic need to provide a solid developer user experience first. Furthermore, in some cases, services would need to become products in order to set up customers for success.

In total, I ended up working on 7 Signal Services, which can be organized into these [human] cognitive capabilities:

Each service involved working with a new, remote team of research scientists, engineers, product managers, and salespeople. I led small teams of designers and front-end developers locally, out of IBM Design in Austin. Sometimes, I facilitated design thinking workshops between product teams and special customers (we call them sponsor users), to break through alignment plateaus, or pivot.

Emotion

I worked on the beta and general availability of Tone Analyzer, a service that analyzes human affect, but takes personality and context into account.

Better examples

The previous Tone Analyzer demo used a sales email as its use case. While this might make sense to IBM employees, to reach a broader audience, I pushed for examples that would be relatable. Demonstrating the value of a new technology with esoteric examples wastes the opportunity for an emotional connection.

I channelled Plutchnik's wheel of emotions in using color for these visualizations.

Visualizations that provide focus

Sometimes the best visualization is a very simple one, or none at all.

Particularly with emotion, personality, and behavior Watson services, people try to make connections in the data that do not exist, which could have real consequences in a business setting. Instead of showing the full array of things the service can detect for an impressive effect, I prefer to go for something authentic and focus on the things it is certain about.

Roadmapping

This is one of the more popular Signal Services, because of the way it could transform certain industries in the near future. I facilitated a design thinking workshop between customers and the research and product teams to come up with "recipes" of tone, emotion, and behaviors associated with personality traits for customers to leverage in their businesses.

Product strategy

I designed developer resources meant to show the breadth of what could be done with multiple Watson services. With this one product in particular I influenced the broader strategy for the developer experience, strategy on Conversation, and even signalled ethical risks in providing "agnostic" developer resources such as these. This was the subject of a talk I gave at the 2016 O'Reilly Design Conference.

The below demo used Tone Analyzer, Personality Insights, and Natural Language Understanding within the context of a support bot handling complaints people tweet at a company account. It demonstrated how Watson could use people's tone of voice to decide whether to respond with an automated (but personalized) message or not, and use personality for advertising purposes.

Personality and behavior

The Personality Insights service predicts personality characteristics, needs, and values from user-generated text. It is geared towards businesses interested in understanding customer habits and preferences individually and at scale. In conjunction with one or two other data points, they can even start predicting behavior. This service is also fairly popular, and as such it was important for me to be accurate, particularly due to a few things that the service measures.

Good looking but misleading

A sunburst visualization looks cool, but in this case misinterprets the data it represents by representing it proportionately. The research team had worked in a text description for the profile in response to people not knowing how to begin digesting the data.

Better examples

I used social media profiles of public figures to give users examples that they could "corroborate," showcasing an intended source of content for the technology.

I also offered other types of content (interviews, speeches, and diaries) that lend themselves to inferring personality.

Actual business value

Consumer behavior is at the heart of this service. Extrapolating meaning about each personality factor doesn't offer the same level of impact and is best left to psychologists anyway.

Remove distractions

I focused on statistically-significant information to discourage users from jumping to conclusions (e.g., justifying bias or discrimination).

2016 Presidential Election

Working with the Federal Government, I ran into people who had worked on presidential campaigns. They described their work similarly to the marketing customers we targeted with this service. In my clinical psychology work, I had also encountered research looking at personality, values, and political leaning.

A couple of friends and I threw an app together for the election. It analyzed a person's Twitter posts, generated a personality, values, and behavior profile, and attempted to predict who they would vote for. We created a profile for presidential candidates, based on everything from stump speeches to debates, to emails. But after we started seeing the results, we decided not to pursue the concept further because we thought it wasn't working. We kept getting too many Trump vote predictions. Yep.

Knowledge Acquisition

These services enable users to start a system for processing unstructured data, give it some guidelines on how to form connections, and allow the system to run with it.

Concept Insights

Concept Insights was my very first project with Watson. I completely re-hauled a demo, turning it from an IBM personnel directory to something that used TED Talks to show nuanced connections between concepts. Users were able to input keywords or paste in a body of text to get TED Talk recommendations.

Using TED Talks as the example domain even helped the research scientists themselves explain what their work was about.

Concept Expansion

To demonstrate how a machine expands its dictionary, I rehauled a demo to use more obvious examples. The value of this service is the ability to scale work without sacrificing nuance, not gaining insights per se.

Using slang as the example domain made it easier for even the research scientists to explain the technology. The previous demo used drug names as its example use case.

Decision Making

These services enable end-users to deal with information overload.

Retrieve and Rank

This demo showing how Watson's long-tail search technology could change the way people find information. We used Stack Exchange Travel as the relatable example, which was one of the services that evolved into a product called Watson Discovery.

Tradeoff Analytics

This service explored several visualizations for choosing between several choices.

Takeaways

Looking back on my work on this team, I noticed a push and pull in my work between trying to demonstrate technology from the machine's viewpoint, and trying to relate it to how humans acquire and exercise those capabilities themselves. I enjoyed using my psychology background to form analogies between the signal services and how people (including developers) understand them. It not only helped me give direction to other designers on the team, but it helped forge a relationship between the academic R&D teams and design.