Computational Analysis of Chat Transcripts

Florida Gulf Coast University (FGCU) is a large public university with the basic Carnegie classification of “Doctoral/Professional Universities” that serves a very high undergraduate student body. The University Library is limited to a main library and is supported by eight in-unit faculty librarians who perform both liaison and functional responsibilities. One of those responsibilities is staffing the online chat “desk.”

The University Library subscribes to Springshare and uses the LibChat platform. The service was launched in 2015 and served as a supplement to a physical reference desk consultation model. The COVID-19 pandemic and associated closure of the FGCU campus led to the subsequent closure of the physical reference desk and a substantial increase of chat questions.

The authors of this project are Rachel Tait-Ripperdan, Digital Humanities Librarian and liaison to History and Humanities, and Regina M. Beard, Associate Dean of Research and Engagement and liaison to Business, Resort & Hospitality, and School of Entrepreneurship.

The purpose of the current investigation is to answer the following questions:

• What was the change in quantity and type of chat questions before COVID-19, during COVID-19, and after COVID-19. In other words, what impact did the pandemic have on patrons’ use and expectations of library chat service at Florida Gulf Coast University?

Additional questions we hope to answer are:

• What interventions are needed to improve our online chat service?
• Should chat continue to be staffed by faculty librarians?
• How can student success be measured and improved upon?
• What role should artificial intelligence play in the future of the service and how should that be implemented?

Our goals for the overall project are as follows: 

• Download and clean up the data for nine years of chats. 
• Gather the quantitative datasets in one place for comparison and visualization of statistics over the years.
• Pre-process the chat questions and transcripts.
• Use natural language processing (NLP) and/or other tools to anonymize the chat questions and transcripts.
• Find significant terms in the chat questions and transcripts.
• Create a seed list from the significant terms.
• Use the seed list to categorize all chat questions and transcripts using GuidedLDA machine learning or some other method, if a better one is discovered.