In this thesis, we investigate the automatic accentuation of words in Slovenian
texts, since the stress position generally cannot be reliably determined from
the written form alone due to Slovenian’s free word stress, while existing
resources are either content-limited or too small. We therefore build a pipeline
for automatic accentuation based on the Sloleks lexicon and the CLASSLAStanza tool, which correctly accents more than 90 % of words in transcriptions
of diverse spoken Slovenian, with shortcomings mainly for rarer and foreign
words and homographs, since it does not distinguish word meaning. We
used the pipeline to build a set of accented words, which, together with their
pronunciations, we used to train an XGBoost classifier for detecting vowel
stress: acoustic features alone reach only 62 % accuracy, while the full feature
set reaches 89,5 %, indicating a genuine but relatively weak predictive signal
from acoustic features. On our data, we also evaluated a model developed
for Croatian that predicts word stress position from pronunciation, showing
that it reaches 67,7 % accuracy on our data and that accuracy declines as the
number of vowels in a word increases.
|
