Improving Answer Extraction in Context-based Question Answering Systems Using LLMs

A new question-answering system built on a fine-tuned large language model has demonstrated high accuracy in extracting answers from text, according to research submitted to arXiv on 4 June 2026 [1]. The system uses the Roberta-base model trained on the Stanford Question Answering Dataset. The work targets a persistent shortcoming in current QA technology: systems often produce irrelevant or imprecise responses even when the correct information is present in the source text [1]. Question answering, a discipline spanning information retrieval and natural language processing, typically constructs answers by querying structured knowledge bases or pulling information from unstructured document collections such as Wikipedia pages and newswire reports [3]. The proposed system accepts a textual context and a question as input and outputs a concise answer [2]. The researchers fine-tuned a pre-trained Roberta-base model on SQuAD1.1, a benchmark dataset that provides high-quality context-question-answer triplets for supervised training and evaluation [2]. High-quality labeled training datasets of this kind are usually difficult and expensive to produce because of the time required to label data [6]. The fine-tuned model achieved a ROUGE-L score of 86.84%, a BLEU score of 28.24%, and a BERTScore of 95.38% [1]. These metrics are part of a broader ecosystem of standardized tests designed to evaluate language model performance on tasks such as answering questions, text classification, and machine translation [4]. The results indicate that targeted fine-tuning substantially improves the reliability and precision of QA systems [2]. The findings arrive as large language models continue to advance rapidly; OpenAI’s GPT-4, for instance, was integrated into Microsoft’s Bing Chat in February 2023 and released in ChatGPT the following month [5]. The authors of the new study conclude that their approach demonstrates strong accuracy and answer relevance for context-based question answering tasks [1].