Combined Machine-Learning Approach to PoS-Tagging of Middle English Corpora

Abstract

This paper considers the problem of part-of-speech tagging in Middle English corpora (as well as historical corpora in general). Whereas PoS-tagging in general is now considered a solved problem for Modern English and is mainly achieved via hidden Markov models (HMM) and matrix-based word-to-vector conversions with every word in the dictionary being embedded into a single dimension, this approach relies on recurrent syntactic structures and context-free generative grammars and is therefore not applicable to older iterations of the English language due to irregular word order. As such, we believe that Middle English could be better handled by a morphographemic encoding and instance-based machine learning algorithms like SVM, random forests, kNN, etc. Using a moving-average method to generate multidimensional vectors giving a reliable numeric representation of character composition and sequences, we have achieved a precision and recall of 87.5% in classifying Middle English words by their part of speech while using a simplistic combined voting-based binary classifier. This result could be deemed satisfactory and encourages further research in the area.