Automatic extraction of UML class diagrams from textual specifications is an important advance in software engineering, enabling the transformation of textual descriptions into visual models useful for system design. This study is part of the Model-Driven Architecture (MDA) approach, aiming at automating the generation of UML diagrams from specifications. To evaluate the performance of this approach, we compare six pre-trained language models: BERT, RoBERTa, SpanBERT, XLNet, MiniLM and Electra, by applying them to the task of extracting UML classes, attributes, methods and relationships from an annotated corpus. Each model is successively used to measure their effectiveness in capturing UML elements and generating class diagrams. This comparative evaluation highlights the specific characteristics of each model and helps guide the choice of models suitable for text extraction in an MDA framework. Our work thus constitutes a reference basis for the selection of language models in the context of automatic generation of UML diagrams.