Q: An intelligent conversational interface for personal data labeling

Mostafa Mohsenvand


With the exponential increase of personal data in the forms of images, videos, emails, and social media posts, the time is ripe for building personal AIs that utilize these data to enhance the productivity and creativity of the users. Training AI algorithms require labeled and processed data.

However, annotating data is time-consuming and often regarded as the bottleneck of supervised learning. Most tools used for data labeling are tailored for the needs of data-scientists and researchers and are far from being useful for general users.  The users of these systems need to know the ontology of possible labels beforehand and use complex interfaces and workflows to maintain the consistency and quality of the resulting dataset. "Q" aims to reformulate data annotation as an engaging conversation by asking appropriate questions and automatically highlighting possible regions of interest. To come up with relevant questions, Q learns from the Wikidata public knowledge graph by computing the probable properties and relationships of entities. It also utilizes the previously annotated pieces of data to speed up the process.