Clustering in a Phonotactic Maximum Entropy Model
Summary
How learners of a language, such as children, acquire knowledge of a language has always been a big question within linguistics. Often, to model this, a maximum entropy model is used. Nazarov (2016) and Mayer (2018) showed that it was possible to learn phonotactic phenomena without a priori knowledge. These studies, however, did not consider the effect of clustering on accuracy.
The goal of this research is to measure the effect of clustering on a phonotactic maximum entropy model. The main question of this research is whether clustering in a phonotactic maximum entropy model, which learns a language, improves the performance of the model, compared to a model without clustering.
To answer this question, a model is created. This model first creates constraints using clusters of classes of phones. Next, a maximum entropy model is used to weigh these constraints. Finally, using these weights, the model predicts the probability of words in the language. This is compared to the actual probability and evaluated. The model also compares different models with and without clustering.
The models with clustering turned out to perform better than the models without clustering. From this, it could be inferred that clustering improved the performance of the model.
This model was run using a made-up 'toy' language. Possible follow up research could take a look at a real language, with a bigger dataset.