Stanford Alexa Prize Team Shares Insights on Improving Socialbot User Experience
Stanford University's AI team, developers of the Chirpy Cardinal socialbot that secured second place in the 2019 Alexa Prize Socialbot Grand Challenge, has released findings from three research papers presented at SIGDIAL 2021. The study leverages real-world interaction data from US-based Alexa users to address common pain points in open-domain conversational AI. The researchers focused on three critical areas: understanding and predicting user dissatisfaction with neurally generated dialogue, identifying effective strategies for handling offensive user behavior, and increasing user initiative to create a more balanced conversational dynamic. Unlike controlled laboratory settings, the Alexa Prize environment offers insights into how bots perform amidst varying user expectations, background noise, and diverse topics ranging from current events to personal interests. The team utilized a modular design combining neural generation, specifically a GPT-2 medium model fine-tuned on empathetic dialogues, with scripted responses. These findings provide practical guidance for chatbot developers and researchers aiming to enhance the naturalness and engagement of socialbots in unstructured, real-life scenarios.
Wire timeline
Stanford Alexa Prize Team Shares Insights on Improving Socialbot User Experience
Stanford University's AI team, developers of the Chirpy Cardinal socialbot that secured second place in the 2019 Alexa Prize Socialbot Grand Challenge, has released findings from three research papers presented at SIGDIAL 2021. The study leverages real-world interaction data from US-based Alexa users to address common pain points in open-domain conversational AI. The researchers focused on three critical areas: understanding and predicting user dissatisfaction with neurally generated dialogue, identifying effective strategies for handling offensive user behavior, and increasing user initiative to create a more balanced conversational dynamic. Unlike controlled laboratory settings, the Alexa Prize environment offers insights into how bots perform amidst varying user expectations, background noise, and diverse topics ranging from current events to personal interests. The team utilized a modular design combining neural generation, specifically a GPT-2 medium model fine-tuned on empathetic dialogues, with scripted responses. These findings provide practical guidance for chatbot developers and researchers aiming to enhance the naturalness and engagement of socialbots in unstructured, real-life scenarios.
The Stanford AI Lab Blog