Intent Recognition on Low-Resource Language Messages in a Health Marketplace Chatbot
David Tresner‐Kirsch, Amanda Azari Mikkelson, Chika Yinka-Banjo, Mary Akinyemi, Siddhartha Goyal · 2023
In this paper we explore intent recognition models of user messages sent to a chatbot that primarily uses system-initiated navigation. The chatbot (askNivi) discusses sexual and reproductive health topics for educational purposes and to facilitate healthcare access. It is deployed in four languages, three of which are considered low-resource languages (Hausa, Hindi, and Swahili; the fourth language is English). Although the primary navigation mode is system-initiated, many users attempt to take initiative by expressing various intents with natural language messages. This paper describes a multi-lingual corpus of those initiative attempts manually annotated with intent labels as well as results of modeling experiments on the classification of these intents using context-dependent supervised machine learning.