Since vowel length in Classical Latin is phonemic, Latin vowel length annotation efforts to date rely on dictionary lookup, with the exception of a small number of rules of thumb. In this squib, I explore the possibility of predicting the length of Latin syllable nuclei based only on phonological features available from the orthography, without recourse to a dictionary. I show that a simple machine learning classifier is able to score remarkably well on this task, showing that Latin vowel length is in fact predictable for 95% of previous seen and 89% of unseen syllables. Further, by using interpretable machine learning architectures in the form of decision trees, random forests and (structured) perceptron, we can extract the linguistic features judged most predictive for the task.
Special thanks to Kevin Ryan for sharing his compiled dataset of Latin text annotated with macrons.
This project uses Anaconda to manage dependencies. Clone this repository, then use
conda env create -f environment.yml
to install the dependencies. Activate your new conda environment with
conda activate latin-vowel-classifier
If the /data folder only contains Ryan_Latin_master.txt, run
python preprocess.py
to create the two corpus files Latin_words_preprocessed.txt and Latin_words_preprocessed_unique.txt.
Then run model.py to train and evaluate a model:
python model.py
Various variables can be set in model.py,
including the type of classifier, its
hyperparameters, which features to use, and whether or not to plot the tree
(only applicable for a decision tree). These are set in the main body of the script.
Tree plotting options can be changed in the plot_fitted_tree() method.
Logs will be output both to the console and to a log file under /experiments
with the current timestamp.
Unit tests for syllabification are contained in test_syllabify.py.
The file test_syllabify_corpus.py contains an integration test which checks
that the entire corpus can be syllabified without errors.