TechBot: A Computer Science Focused QA System

Abstract

This paper presents Tech Bot: A Computer Science Focused QA System designed for the computer science domain. Since there is a rapid increase in the field of computer science, a system that can provide accurate, context-aware responses specific to computer science is desperately needed. TechBot addresses this need by combining the advanced capabilities of Natural Language Processing techniques and Deep Learning Models. It includes a strong preprocessing pipeline, a query comprehension module that is capable of decoding technical terms and an answer retrieval framework optimized for computer science content. It performs better in the computer science domain than in general-purpose systems. Using the Hugging Face Dataset Library with extra web scraping, 20,076 question-answer pairs were collected. Secondly, the QA pairs were split into training and validation sets. Two state-of-the-art transformer-based sequence-to-sequence models, T5 and BART, which are designed for natural language generation, were employed. After fine-tuning on the domain-specific dataset, BART outperformed T5 based on the ROUGE score evaluated on the validation set. Initially, the BART model with 140 million parameters was trained for 21 epochs, achieving a ROUGE score of 0.2867. Finally, an average ROUGE score of 0.2890 was obtained on the test dataset.