Automated Alignment between Elicitation Interviews and Requirements

A research team has formalized a method to automatically measure how well software requirements align with the stakeholder interviews that produced them, proposing two heuristic metrics to quantify the match between transcripts and user stories [1]. The work, led by Francesco Dente, addresses a persistent bottleneck in software engineering: evaluating whether documented requirements faithfully capture what stakeholders said during elicitation sessions remains a largely manual task [1][2]. The researchers introduce a framework called TEXT2STORIES, which casts the alignment as a matching problem between transcript chunks and user stories [3]. Two interpretable measures are proposed: requirements faithfulness, defined as the proportion of stories supported by the transcript, and interview coverage, the proportion of transcript segments supported by at least one story [1][2]. Experiments across four datasets show that a large language model-based solution achieves a macro-F1 score of 0.86 on manually labeled chunk-story pairs [1]. To make the approach more scalable, the team also demonstrates how embedding models can serve as blockers, reducing the number of candidate pairs that the more expensive LLM must evaluate [1][3]. The pipeline segments transcripts into chunks, then applies a pairwise matcher — which can be a cross-encoder or an LLM judge — only to the top-K candidates identified by embedding similarity [3]. The paper builds on a growing body of work applying language models to requirements elicitation. A separate study published on arXiv describes an LLM-based approach to extract goals from interview transcripts and construct goal models, reporting 72.2% accuracy in finding refinement relationships [4]. Another recent system, called RECOVER, classifies individual conversation turns as requirements-relevant and then uses an LLM to generate system requirements from those turns, rather than summarizing entire conversations [5]. RECOVER operates through three steps: classification of turns, filtering and contextual enhancement, and automatic generation of requirements [5]. The TEXT2STORIES authors note that their formal framework and automated matching techniques are basic components that can support emerging tasks such as tracing requirements back to interviews and generating requirements directly from conversations [1][2]. The submission history shows an initial version of the paper uploaded on October 8, 2025, at 1,218 KB, with a revised version following on June 9, 2026, at 630 KB [1].