I am a fourth year undergraduate studying Computer Science at the University of Michigan. I'm advised by Dr. Walter S. Lasecki in the Crowds + Machines (CROMA) Lab. I'm interested in building human-centered Artificial Intellifence to help humans better interact with AI systems.
Our work aims to generate clarifying questions that require less effort for the end user to answer. We use machine intelligence to calculate informativeness of each question, to find questions that could more evenly separate the candidate dataset. We use human intelligence to balance the informativeness with answerability, so that the question was easy to recall and say. We introduce a novel hybrid workflow that combine machine knowledge and human knowledge.
Traditional AI systems need structured input, which keeps users from fully utilizing their functionality, so we aim to build a conversational AI, which can infer what the end-users want through natural conversations. I led a team of 8 undergraduates where we built a conversational AI that could assist a user to generate a travel itinerary. We indentified possible intention logical transition based on a finite state machine. We trained an intent classifier to infer user intent, making use of an End-to-End Recurrent and Bi-Direction Long Short-Term Memory (LSTM) Neural Network.
Our work in this project aims to reduce annotation effort for large text dataset. We aim to elicit and leverage high-level human intelligence by making them write labeling rules instead of merely having them label each instance. We introduce a workflow where machine suggests possible keywords to humans by clustering the data into keyword-based patterns using Latent Dirichlet allocation (LDA). And then humans identified a label of each pattern to form a rule, which can be encoded into a labeling function implemented by computing cosine similarity using Glove word embedding.