UC Irvine

Quantifying the complexity of a natural language is a difficult task on its own and comparing two or more languages typically requires establishing a reference point and determining the biases and context of the languages being compared. I propose a new metric for unbiasedly quantifying the complexity of a language in a way that allows for easy comparison between languages. I use a variety of common machine learning solutions for tasks such as part- of-speech tagging and language modeling, then analyze the learning ability of these models as parameters are adjusted. I then use the evaluation metrics from these tasks to compare similar models trained on different languages. I find that the evaluation metrics accuracy and perplexity mimic the behavior of four metrics found in linguistics literature and can be used to compare relative complexities.