M.S. AAI Capstone Chronicles 2024
A.S.LINGUIST
8
and text outputs (Pardasani et al., 2018), and the type of conversations one can have with the
chatbot varies with the project scope, spanning from providing psychotherapy and psychological
counseling (Huang et al., 2021) to assisting users in retail environments (Berkley School of
Information). In terms of machine learning methods, different approaches can be adopted to
build well functioning sign language interpreters and chatbots, based on the characteristics of
inputs and outputs, the size of the datasets used to build the models, and other constraints like
time and machine capacity.
The project we present leverages supervised learning with neural networks, similar to
other notable projects, including a MIDS Capstone project from UC Berkeley (American Sign
Language Bot, n.d.) and a project developed by GitHub user Sophgdnfor (Sophgdn, n.d.). The
MIDS Capstone project from UC Berkeley focused on assisting ASL speakers in retail
environments using real-time video intelligence to interpret ASL. The team provided an
overview of data processing, model implementation and error analysis in their invention.
SignBot, an initial prototype developed by Sophgdn, uses LUIS, a Microsoft's machine learning
based service, to translate natural language speech into Australian Sign Language. The bot
operates by translating messages into videos that depict how they would be communicated in
sign language. The developers use LUIS to simplify complex sentences, making them
compatible with the Australian Sign Language vocabulary. According to their README file, the
application records a live feed of the user communicating via sign language, processes this feed
using LUIS, and presents interpretation options for the user to choose from.
Our project incorporates similar approaches to both mentioned projects but with some
notable differences. Like the UC Berkeley project, we used a live feed in our application, but
while their primary goal was to focus on ASL in retail, our vision was to create a more
188
Made with FlippingBook - professional solution for displaying marketing and sales documents online