This paper wishes to explore to what extent lemmatisation and morphosyntactic annotation are important for deep learning predictions and textual motif detection. A broader research on the style of Peter Damian's letters (11th century) was the occasion to explore this question. After having trained two deep learning models on a selection of 12 classical authors using the Hyperdeep platform, one on lexical forms alone and the other on lemmatised and annotated texts, we introduced to them the medieval letters of Peter Damian in order to not only examine which authors are deemed to be stylistically close to Peter according to both models, but also to compare whether the results are similar and whether the same linguistic structures receive a high activation rate. The results suggest that a dialogue between both methods could be an interesting path to explore in the search for textual motifs, as the first "lexical" model may indicate rough outlines of these motifs, whereas the second model can offer concrete examples and/or variants of the first motifs identified.