Authors:

Christina Joslin

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

The creation of forums for frequently asked questions began in the 1980s to provide users with answers to common problems without requiring direct assistance from technical support staff. However, the growing scale and complexity of user inquiries in high-performance computing environments has surpassed the capacity of traditional support methods. To address this challenge, we developed an automated solution that generates frequently asked questions from historical technical support ticket data. This multi-stage process began by summarizing raw support tickets into issue-resolution pairs using a fine-tuned large language model enhanced by Low-Rank Adaptation adapters and eight-bit quantization. The model was trained on synthetically generated examples to improve performance on domain- specific inputs. These structured summaries were then embedded and grouped using K-Means clustering, with a custom scoring mechanism based on cluster size, cohesion, and separation guiding the selection of high-quality clusters. Finally, representative examples were passed to an instruction-tuned large language model using few-shot learning techniques for frequently asked question generation. Early results indicated that this pipeline improves the relevance, clarity, and consistency of generated content. By transforming unstructured ticket logs into actionable knowledge, our approach reduces redundant inquiries, accelerates resolution, and supports scalable technical support infrastructure. This work contributes to broader efforts to modernize documentation and assistance systems using artificial intelligence in complex computing environments. Keywords: Frequently Asked Questions; Large Language Models; Low-Rank Adaptation; Semantic Clustering; Few-Shot Learning

Source:

Purdue University / 2025

Topics:

No topics listed

Co-authors:

Christina Joslin