In this paper, we describe experiments on part-of-speech tagging of spoken Italian that we conducted in the context of the EVALITA 2020 KIPoS shared task (Bosco et al. 2020). Our submission to the shared task is based on SoMeWeTa (Proisl 2018), a tagger which supports domain adaptation and is designed to flexibly incorporate external resources. We document our approach and discuss our results in the shared task along with a statistical analysis of the factors which impact performance the most. Additionally, we report on a set of additional experiments involving the combination of neural language models with unsupervised HMMs, and compare its performance to that of our system.
