Authors:

David Burns

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

High-performance computing centers handle a large number of support tickets, many of which are repetitive, which is a hindrance on support staff who must respond to similar issues repeatedly. One solution is to automatically generate answers to frequently asked questions (FAQs) by analyzing support tickets. The goal was to automate this process using a pipeline that integrates natural language processing tools and large language models. Although the tickets were already organized by topic, the existing categories were too broad, requiring further refinement. Tickets were first filtered by frequency and recency of the issue. A large language model was then used to summarize the ticket contents into concise, structured issue-resolution pairs. This model was fine-tuned using low rank adaptation using legacy tickets. Manual inspection of the summaries led to the choice of summarization approach, as quantitative metrics did not perform well in comparison. The tickets were then clustered using K-Means. This decision was based on several quantitative metrics, such as silhouette score. After clustering, the top- ranked sub-clusters were selected for FAQ generation. Two methods were tested: generating a single Q&A per cluster, and splitting the cluster content into multiple pairs. While both approaches worked, the one-step approach appeared to produce better results according to empirical testing. Initial reviews suggest that the new FAQs are more consistent and readable, helping to reduce ticket volume, improve documentation, and assist in training new staff. Keywords: Natural Language Processing; Large Language Models; High Performance Computing; Question Generation

Source:

Purdue University / 2025

Topics:

No topics listed

Co-authors:

David Burns