Parsing-based approach for entity extraction from natural language software requirements

Software requirements are written in natural language, so they are inherently ambiguous. This ambiguity results in the misinterpretation of the requirements and leads to an incorrect implementation of the system that does not meet the user's needs. So, it is important for all the stakeholders to comprehend the requirements. The requirement statements revolve around the requirement entities. Identifying requirement entities is essential to understand requirements, and they are used in identifying coreferent entities, creating glossaries, producing summaries of documents, feature extraction, requirements classification, and domain understanding. Manual entity extraction is a labour-intensive task. Recently, transformer-based models have played an important role in improving entity extraction results, but they require extensive pre-training of models for domain-specific tasks. To overcome these challenges, the proposed automated approach is used to extract entities from software requirements without the need for manual efforts and a large dataset for pre-training. In this study, two widely used parsing techniques, constituency parsing and dependency parsing, are applied to extract noun phrase-based entities. In addition, a hybrid approach comprising the combined extractions of both parsers is developed. The dataset comprises the software requirements from the healthcare domain and is manually annotated to establish a gold standard for evaluation. Experimental results indicate that the hybrid approach achieves the best overall extraction performance among the evaluated methods. While the constituency parser achieves the highest precision, it exhibits comparatively lower recall. In contrast, the hybrid approach achieves the highest recall and F1-score while maintaining precision comparable to that of the constituency parser. These results demonstrate that the hybrid method provides a more effective solution for automatically extracting entities from software requirements by leveraging syntactic structures of sentences rather than relying on domain-specific model training.