Personalization and Relevance in NLG

Despite the recent advances in language modeling techniques, personalization remains a challenge for many NLP tasks. In this talk, we will explore personalization through several different lens to understand how we can make progress on this front, and emphasize why human-centered approach is a crucial part of the solution. First, I will challenge the ground-truth assumption in the context of user or situation sensitive language tasks. In other words, I will argue that the same question might be addressed differently by a system, depending on the user or the situation they are currently facing. Then we’ll discuss what are our user needs, and how can we design these tasks to produce useful and relevant responses, but also what potential harms we should be aware of working on personalization [5]. Next, we will look into personalization in the augmentative and alternative communication (AAC) world. Specifically, how through an icon-based language, individuals with compromised language abilities (that may arise due to Traumatic Brain injury (TBI) or Cerebral Palsy (CP)), we can accommodate their needs and what are the challenges in developing icon-based language models [3]. Third, in the process of personalization, models are expected to accommodate and adapt to the specific language and jargon spoken by the user. What are remaining challenges for deep learning architectures in the process of adapting user data or new domains [2, 4]. Finally, I will share work-in-progress where through Wizard-of-Oz (WoZ) experiments [1] we identify and learn useful actions of social conversational systems in classroom setting.