Automated Morphemic Analysis of Russian Words

The paper considers formal-structural and formal-semantic principles of morphemic analysis and illustrates that difficulties in morphemic analysis can be viewed as the deviations from the additive model of morphology. The work compares several classifications of the degrees of morphemic segmentability. The study examines machine learning approaches to automated morphemic analysis which utilise the predictability of the segment, a unigram probability model, conditional random fields and convolutional neural networks. In addition to the available test sample, a new sample of rare words (e.g. neologisms and recent loan words) was manually created and annotated for morphemic structure. Finally, a linguistic classification of errors was proposed, according to which the models make mistakes due to the influence of more frequent morphs, morphs with low frequency, de-etymologisation, morphological alternation and rebracketing.