Annotating Manzoni: Challenges in the Annotation of Lemmas, POS and Features in "I Promessi Sposi"

Slides presentated at CLiC-it 2025 (Eleventh Italian Conference on Computational Linguistic). In this paper we introduce a dataset of I Promessi Sposi annotated with lemmas, UPOS tags, and features aligned with Universal Dependencies (UD). Three representative chapters from Manzoni’s 1840 edition (791 sentences, almost 26 K tokens) were automatically tagged with UDPipe and fully manually corrected. Tailored guidelines extended standard UD practice with: (i) a double lemmatization approach, one that maintains archaic spellings and altered forms and one that normalizes lemmas, (ii) novel features that capture specific important characteristics of the novel, such as the use of apocopated and altered forms. Using the resulting dataset, we retrained the Stanza pipeline to obtain an in-domain model. Augmenting training data with ISDT sentences yielded further, although smaller, gains. Finally, a CRF sequence tagger was developed to identify apocopated forms.