• KJ Ha

ML4Web - Assignment 9

Conversation Generator


I trained my own model using the basic generative text model on the Runway website.

I took Topical-Chat dataset as a source and created a new dataset only composed of the messages. Topical-Chat is a knowledge-grounded human-human conversation dataset where the underlying knowledge spans 8 broad topics and conversation partners don’t have explicitly defined roles.


Original dataset example:


Preprocessed dataset example:


Training setting:


My motivation was to create a model that functions as a chatbot; however, the conversation didn't go as smoothly as I expected. It might be because of the nature of the original dataset, which consists of a set of conversations about very specific topics. Many topics, such as NFL games or football, weren't familiar to me to keep the conversation going. If I trained the model on a more general conversational dataset, I might get better results.


After I hosted the model over HTTP, I used p5.js sketch to design the interface.

(Since I haven't made the model public, it won't be visible to others besides myself.)



Examples



Demo video


© 2020 Kyungjoo Ha. All Rights Reserved