The challenge of POS tagging and lemmatization in morphologically rich languages is examined by comparing German and Latin. We start by defining an NLP evaluation roadmap to model the combination of tools and resources guiding our experiments. We focus on what a practitioner can expect when using state-of-the-art solutions. These solutions are then compared with old(er) methods and implementations for coarse-grained POS tagging, as well as fine-grained (morphological) POS tagging (e.g. case, number, mood). We examine to what degree recent advances in tagger development have improved accuracy – and at what cost, in terms of training and processing time. We also conduct in-domain vs. out-of-domain evaluation. Out-of-domain evaluation is particularly pertinent because the distribution of data to be tagged will typically differ from the distribution of data used to train the tagger. Pipeline tagging is then compared with a tagging approach that acknowledges dependencies between inflectional categories. Finally, we evaluate three lemmatization techniques.

No abstract

No abstract

Transformer-based models are now predominant in NLP. They outperform approaches based on static models in many respects. This success has in turn prompted research that reveals a number of biases in the language models generated by transformers. In this paper we utilize this research on biases to investigate to what extent transformer-based language models allow for extracting knowledge about object relations (X occurs in Y ; X consists of Z; action A involves using X). To this end, we compare contextualized models with their static counterparts. We make this comparison dependent on the application of a number of similarity measures and classifiers. Our results are threefold: Firstly, we show that the models combined with the different similarity measures differ greatly in terms of the amount of knowledge they allow for extracting. Secondly, our results suggest that similarity measures perform much worse than classifier-based approaches. Thirdly, we show that, surprisingly, static models perform almost as well as contextualized models -in some cases even better.