Date

4-2024

Project Type

URC Presentation

Department

English

College or School

COLA

Class Year

Senior

Subject

Computational Linguistics

Major

Linguistics

Faculty Research Advisor

Rachel Burdin

Abstract

This study attempts to use three different kinds of n-gram text classification models to differentiate the standard forms of Croatian, Bosnian, and Serbian. These three languages, along with Montenegrin, were once considered one language, collectively termed “Serbo-Croatian”. These languages share a common South Slavic ancestry, and there is an argument to be made that their novel status as distinct languages is due to non-linguistic factors, such as culture and politics. This study uses 300,000 sentences from each language, sourced from Wikipedia pages written in the standard forms of each language. Three different classifiers were used: one unigram, one bigram, and one combined unigram-bigram model. The classifiers were evaluated by the industry-standard performance metrics for text classification, including accuracy, precision, recall, and F1. Overall, the classifiers differentiated the three languages relatively unreliably, with the unigram classifier achieving the highest accuracy rate at 74.05%. The bigram classifier only achieved a 51.30% accuracy rate, and the combined classifier achieved a 68.94% accuracy rate. All classifiers were able to predict Serbian sentences most accurately, followed by Croatian sentences, and Bosnian sentences, with the lowest rate of accuracy. The highest accuracy overall (87.07%) was achieved using the unigram classifier to predict Serbian sentences, higher than any other classifier used on any other language. The results suggest that Croatian, Bosnian and Serbian are too lexically similar to be reliably differentiated on the level of individual sentences. This may also suggest that the three languages, along with Montenegrin, should be categorized as one language for practical purposes.

Download

Share

COinS