Transformer-based AI model tackles hidden conflicts in software requirements

Identifying conflicting or redundant requirements in software specifications remains a labour-intensive task. A research team has developed a machine learning system designed to automate this process, using transformer models fine-tuned to catch inconsistencies that often derail development efforts.

The Problem#

Clear, non-overlapping requirements are foundational to successful software projects. When specifications conflict or duplicate effort, teams waste resources and introduce bugs. Manual review becomes increasingly impractical as requirement sets grow larger. The challenge is that spotting these issues demands both semantic understanding and domain knowledge that traditional rule-based systems struggle to replicate.

A Transformer-Based Solution#

The researchers designed SR-BERT, a neural architecture combining Sentence-BERT and Bi-encoders to classify pairs of requirements as either conflicting, duplicate, or compatible. Rather than starting from scratch, they applied transfer learning—leveraging pre-trained transformer models and refining them on requirement-specific data through multi-stage fine-tuning.

The approach recasts the problem as sentence-pair classification, allowing the model to learn relationships between two requirement specifications and determine whether they align or contradict.

Testing and Results#

Performance was evaluated across four different datasets. The fine-tuned transformer models demonstrated strong consistency across diverse datasets, with SR-BERT showing superior performance on larger requirement collections. The researchers also tested how well conflict detection generalizes across different domains and integrated rule-based filtering to validate model predictions, improving reliability.

Why It Matters#

Automating conflict and duplicate detection could compress a time-consuming manual step into seconds, freeing teams to focus on substantive design work. The use of transformer-based NLP strategies suggests that general-purpose language models can be effectively adapted for specialized software engineering tasks with appropriate training.