Advancement: WellBe: A Conversational Agent for Well-Being

Speaker Name: 
Jiaqi Wu
Speaker Title: 
PhD Candidate (Advisor: Marilyn Walker)
Speaker Organization: 
Computer Science and Engineering
Start Time: 
Wednesday, September 18, 2019 - 2:00pm
Location: 
Engineering 2 Building, Room 280
Organizer: 
Professor Marilyn Walker

Abstract:
A conversational agent for well-being, deployed on a user’s phone, or on devices in the home such as Google Assistant or Amazon Alexa could bring benefits to many people. Developing such a conversational agent requires advances in two distinct technology areas: (1) analysis and categorization of user affect, user activity types and their relationship; and (2) design of dialogue management strategies that can utilize these categories, along with theories of well-being, to successfully interact with the user and, over time, improve the user’s well being.

In our work to date, we have tackled the analysis and categorization of user affect, by first showing that standard sentiment-analysis tools, such as Stanford sentiment, do poorly at recognizing users’ affect, perhaps largely because they have been trained on datasets, such as product reviews, that are a poor match to users’ descriptions of everyday activities and affectual states. We developed three datasets that are more representative of first-person statements of affect and activities, and conducted a set of experiments on classifying user affect, both within and across datasets. Our final results show that a hierarchical attention model based on BERT can achieve F1 measures as high as .88 within-domain, and as high as .84 across domain suggesting that the models trained on one type of data generalize well. We also present a set of experiments on the AffCon dataset on classifying user activity types, and show that we achieve F1 measures ranging from .64 to .94.

We have also developed a pilot version of WellBe, a conversational agent for well-being, implemented on the Amazon Alexa platform. The design of WellBe involves many open questions. We plan to build on previous work on phone-based applications for well-being to develop a set of dialogue strategies that are conditioned on user-affect, and which recommend activities that can improve the users' well-being. We plan to develop and evaluate two different dialogue strategies: one based on affect alone and the other based on affect plus user activity classification, and explore methods for automatically learning to improve WellBe’s dialogue strategies.