Transforming Model-to-Model Interactions with Advanced Natural Language Processing: A Deep Learning Approach

The following discusses a feasibility review being carried out for the use of natural language processing in the context of capturing Model-to-Model transformations in visual models, in particular, the extraction of information from text labels in BPMN and UML notations. Three significant tasks that the paper aims at taking care of include the extraction of verb/noun relationships, parsing conjunctive/disjunctive statements, and distinguishing abbreviations and acronyms. The current state of the art of NLP tools—such as CoreNLP, Stanford Stanza, Flair, Spacy, AllenNLP, BERT, and ELECTRA—coupled with some custom models, such as BERT-BiLSTM-CRF and ELMo, has resulted in these benchmarks. This system uses a BiLSTM-CRF tagging model trained on augmented data to do well on classes where the other models are confused. The model's performance is good in extracting noun and verb phrases irrespective of the recognition of words present in the shorttext labels, which are used in most of the UML and BPMN models. This study also raises challenges in the detection of abbreviations and acronyms and in processing conjunctive and disjunctive statements. Experimental results note that although Stanza-based implementation excels in the extraction of phrases and conjunctive phrase processing, the BERT-BiLSTM-CRF model is outdone in the detection of verb phrases. This goes a long way to show that, as promising as deep learning techniques are, they still bear their flaws but can be improved upon for application in different domains in NLP.