Old Norse as a problem of automatic PoS-tagging

Abstract

This article dwells upon automatic PoS-tagging of Old Norse by computational means, including machine learning. It analyzes the available language material in diachrony from the standpoint of how language evolution might have affected the quality of automatic PoS-tagging. This article further describes the phonetic traits that have assumingly led to any classification errors. The research material is an Old Norwegian educational text titled Konungs skuggsjá, or “King’s Mirror”, vectorized by the moving average method and then used to train an Ada-Boosted random forest model. The resulting classification accuracy is about 97%. However, being non-contextual, this vectorization method enables no complete differentiation of morphologically similar parts of speech: verbs, nouns, adjectives, and adverbs. This becomes evident when digging into the identified high-weight classification features, each being a vectoral dimension corresponding to a specific alphabet character; another indicative factor comprises Morfessor-identified high-rank morphs, analyzing which reveals the morphogrammatic units that cause the most classification errors. Historical consideration of these morphs shows that their collision is due to them being inherited from Proto-Germanic (PG) while undergoing rhotacism, or conversion from PG /z/ to ON /r/. However, the same process effectively prevents the collision of rhotacized finite verbal forms with the genitive case that inherits the PG suffix -s. The key finding is that such morphological collision being unavoidable, character-based vectorization might not suffice when using a small training set or when trying to classify not only by parts of speech, but also by specific forms in the paradigm.