Knowledge Network
Background
The company’s AI technology composed a knowledge network for each search query it received, providing a visual representation of how common themes, ideas and concepts in that field relate. The knowledge network could provide a new way to experience information that supports multi-modal learning. Here is the gist of how it worked:
The AI recognizes the main concepts and supporting ideas from each document.
The AI had an inventory of cognitively significant relationships that it could identify between concepts by using linguistic analysis. It would generate a knowledge network of this information, which was both machine and human readable.
The AI connects information from different sources across texts using this network.
The network contains knowledge and can be queried.
The AI learns from everything it reads, continuously building a bigger and smarter network.
Approach and Challenges
We wanted build a user experience around the network that would make it useful to researchers. Designing solutions that addressed their unique pain points while showcasing our AI’s unique power required us to conduct research and surveys, and to iteratively seek feedback on features and versions of features we produced.
As our work progressed, we encountered challenges conveying the significance of the analysis. This was because we were building an altogether new way to experience information and the earliest versions of our network were static. We would continue to design and refine features to address user pain points while meeting these challenges.
The features I describe below fall into two classes: features which helped users navigate the network, and features which showcased our AI’s analytical power.
1: Features that improved network navigation
My product team built features that would help users make sense of the new format of information including guides and drop-downs, which infused context about the AI’s analytical framework into the user experience. This would help make the significance of the AI’s analysis clearer. We would continue to update and improve our inventory and micro-interactions as we encountered new data and obtained user feedback on what was insightful and/or useful.
A high-level guide contains the list of the relationships present in the network. Relationships were organized into categories based on the type of information they provided users.
We developed a drop-down accessible for each instance of a relationship in the network which would display the relation’s definition and any details about its occurrence.
We developed a summary drop-down which would appear for any concept that the engine compiled a summary for.
2: Features that showcased our analytical tools
My product team built features that showcased the analytical capabilities our AI could provide to researchers. These features would make the network and its contents useful and usable in a workflow that supported a research journey.
Users can apply a filter on any concept to keep only relations of a select type in view. Here, the filter has been applied to the concept 'anxiety' to display only the concepts it has a 'causal' relationship to. Users were able to use the filter to obtain insights by producing unique views of the network data.
Users could opt to filter for relationships across the full network, to cast different views on the information contained within it.
We provided the complete list of articles that mention any given concept in the network, providing users quick access to the sources indicating its significance.