GPT-3 Transformer-based NLP Ticket classification

Thesis (MA)

Advisor(s):  Simon Fuchs

Context

Providing technical support for own IT related products is an integral part of software developing or software providing companies. For this purpose, most companies provide their customers support ticket systems, in which users can create incident tickets describing their problem or request. In most state of the art support ticket systems (STSs) at least some key decisions in distributing these support tickets to the responsible support assistant or support team are still made by support staff members. This process would bind a lot of working time of technically skilled workers, wherefore big companies often use tier1 workers for support ticket distribution, which are often less-skilled and temporary; or increasingly outsource the whole support to a third-party company completely. This process of manually distributing emerging support tickets by often less-skilled human workers is on the one hand ineffective and expensive. On the other hand, it mostly increases the ticket resolution time and therefore lowers the satisfaction of the user initially creating the ticket.

In recent years, two trends can be identified in the field of support desks: First, support service becomes more broad, from only repairing and replacing broken products to maintaining a product as a service including also installation, online support, warranty, and upgrades. Second, support incident volume grows exponentially due to digitalization efforts made across all industries. This has led to support quality becoming one of the key-factors of customer satisfaction and customer loyalty towards companies. 

With Machine Learning (ML) algorithms becoming more and more common, the automation of support service desks has risen in interest. Technologies for automated ticket classification, improved data collection and (semi-)automated incident resolution open the possibility of automating basic day-to-day IT tasks replacing the first level support staff members while simultaneously accelerating the support process increasing customer satisfaction.

In recent months, so-called transformer models have largely risen in interest in the NLP community. The last and most famous example for transformer based Machine Learning is the Chatbot ChatGPT. Especially, pre-trained transformer models like GPT-3 have risen in interest and especially the question of their applicability to a wide area of use cases.

In this context, we want to explore the applicability of the GPT-3 model for automated support ticket classification in a small ticket system with around 7500 tickets as training data. A master thesis should try how far you can come with the GPT-3 model and our ticket data set.

Task(s)

  • Review the scientific literature in the field
  • Writing a theoretical section respective the functionality of transformer models
  • Built a ticket classifier based on the GPT-3 model
  • Evaluate the classifier and compare it with different classification efforts

Requirements

  • High degree of autonomy and responsibility
  • Interest in research on service desks and service desk automation
  • Ability to think “out of the box”
  • Interest in SAP Software
  • Basic ABAP Knowledge or the willingness to learn it
  • Good language skills in German and English

What We Offer

  • An opportunity to write your master thesis J
  • The possibility to work on a real world project using the “hype” model GPT-3
  • Insights into everyday research at a chair
  • A cool team :D to work in

 

Further Information

If you have further questions, please do not hesitate to contact me directly (s.t.fuchs@tum.de). Please send your application including a current transcript of records and your CV to s.t.fuchs@tum.de.