Technical Summary: LENS – LLM-Based Discovery of Latent Requirements from Stakeholder Conversations

Problem Statement

Stakeholder interviews are a primary source for requirements elicitation, yet a significant portion of relevant requirements remains implicit. Stakeholders frequently describe workflows, operational challenges, and inefficiencies without explicitly articulating the software capabilities needed to address them. While recent research has explored using Large Language Models (LLMs) to extract requirements from conversational data, existing approaches primarily focus on identifying explicitly stated requirements. This limitation overlooks latent requirements: plausible opportunities for software automation that become apparent only when stakeholder statements are interpreted in the context of existing organizational tools, workflows, and infrastructure. The challenge lies in automating the inference of these latent requirements while maintaining traceability to the original conversation and ensuring the generated requirements are actionable.

Methodology: The LENS Approach

The authors propose LENS (LLM-Enabled Needs Discovery from Stakeholder Interviews), a two-step framework designed to analyze stakeholder interview transcripts to extract both explicit and latent requirements.

Core Workflow

1. Step 1: Explicit Requirements Extraction:

   • The LLM analyzes interview transcripts to identify requirements expressly stated by stakeholders.
   • It extracts these as structured user stories linked to specific transcript excerpts and timestamps to ensure traceability.
   • The process utilizes few-shot prompting with specific rules to distinguish between actual requirements and mere descriptions of current manual processes or existing automation.

2. Step 2: Latent Requirements Inference:

   • The LLM analyzes the same transcripts alongside organizational context, specifically descriptions of available software tools, infrastructure, and automation capabilities.
   • By reasoning over stakeholder descriptions of pain points and inefficiencies in light of these tools, the LLM infers latent requirements.
   • For each inferred requirement, LENS generates a structured rationale explaining:
     • The underlying problem or inefficiency.
     • How the requirement addresses the problem using specific organizational tools.
     • The expected impact on organizational practices (e.g., time savings).

Technical Implementation

• Model: The study utilized Claude Sonnet 4.5, chosen for its strong reasoning capabilities and large context window.
• Prompt Engineering: The system employs prompts containing:
  • Role definitions (extractor vs. inferencer).
  • Thirteen Quality User Story (QUS) criteria to ensure clarity, completeness, and consistency.
  • Specific rules for distinguishing explicit intent from conversational noise.
  • Few-shot examples to guide generation.
  • Structured output formats (JSON) for downstream integration.
• Context Grounding: Step 2 explicitly inputs a list of organizational tools and their capabilities to ground the inference in technical feasibility.

Empirical Evaluation

The authors conducted a preliminary evaluation in an industrial setting with eSentire, a global cybersecurity company.

• Dataset: Twelve semi-structured interview transcripts from a Security Operations Centre (SOC), covering analyst performance assessment and reporting workflows. The transcripts averaged ~60 minutes, containing ~8,900 words and 154 speaker turns.
• Evaluation Metrics:
  • RQ1 (Accuracy of Explicit Extraction): Measured via Precision, Recall, and F1-score against a ground-truth set constructed by a researcher and validated by domain experts.
  • RQ2 (Usefulness of Latent Requirements): Evaluated by two independent domain experts at eSentire. They assessed requirements based on time gain (comparing automated vs. manual execution), novelty, and feasibility.

Key Results

• Explicit Requirements (RQ1):
  • At a temperature setting of 0.7, LENS achieved the best performance:
    • Precision: 73%
    • Recall: 100%
    • F1-Score: 84.4%
  • False positives were primarily caused by the LLM misinterpreting mentions of manual work (without automation intent), discussions of existing automation, or confirmation-seeking language (e.g., "right?") as new requirements.
• Latent Requirements (RQ2):
  • Usefulness: On average, 75% of the inferred latent requirements were rated by experts as providing Moderate or High Time Gain (i.e., useful automation or time-saving potential).
  • Novelty: Experts noted that the inferred requirements often provided insights and opportunities stakeholders had not initially considered.
  • Feasibility Challenges: Approximately 49% of expert feedback highlighted a lack of concrete technical details in the rationales. Other issues included dependencies on specific data sources/APIs not fully described and occasional misalignment with the actual capabilities of proposed tools.

Key Contributions

1. LENS Framework: A novel approach that extends LLM capabilities beyond extracting explicit requirements to inferring latent requirements by reasoning over organizational context (tools and infrastructure).
2. Structured Output: The generation of both explicit and latent requirements as traceable user stories with structured rationales, linking back to specific transcript timestamps.
3. Industrial Preliminary Evaluation: A realistic assessment of LLM performance in a cybersecurity domain, demonstrating high accuracy in extraction (84.4% F1) and significant perceived value in latent inference (75% time-saving potential).

Significance and Claims

The paper positions LENS not as a tool for producing immediate, implementation-ready features, but as a mechanism to stimulate further stakeholder discussion and iterative requirements elicitation.

• Early-Stage Artifacts: The authors claim that inferred latent requirements serve as valuable "early-stage elicitation artifacts." They help reframing conversations around mitigating inefficiencies and realizing untapped automation potential that stakeholders may not have articulated.
• Human-in-the-Loop: The study emphasizes that expert review remains critical. The moderate inter-rater agreement and feedback on feasibility suggest that LLM-generated latent requirements should be treated as starting points for ideas rather than finalized specifications.
• Contextual Necessity: The results highlight that the quality and feasibility of inferred requirements are highly dependent on rich descriptions of the organization's tools and infrastructure. Without this grounding, LLM suggestions may appear plausible but remain misaligned with system constraints.

The authors conclude that while their findings are preliminary (limited to one organization and domain), they provide early evidence that LLMs can effectively support both the extraction of explicit needs and the discovery of latent opportunities in requirements engineering.