@masterthesis{Sobotta, type = {Bachelor Thesis}, author = {Sobotta, Leon Niklas}, title = {Investigating the cross-linguistic performance of pre-trained transformer-based language models in binary sentence classification tasks}, school = {Hochschule f{\"u}r Technik und Wirtschaft Berlin}, pages = {62}, abstract = {This study evaluates the effectiveness of AI models in multilingual in-car voice assistant systems, utilizing DistilBERT [Devlin et al., 2018] as a feature extractor to train a rudimentary feedforward neural network to classify voice commands. It explores the possibility of minimizing the number of required models for the Indo-European language family to enhance accuracy and efficiency with a refined model approach. By evaluating language-specific AI models using MBition's test data, the research examines classifier training for each language and linguistic heritage group. It assesses the performance of these groupings, especially within the Indo-European family, comparing it to monolingual models emphasizing Germanic and Romance languages to verify if model unification can maintain efficiency. The results of this study indicate that language family-based classification with DistilBERT significantly enhances binary sentence classification, merging several language models into a single entity without degrading performance.}, language = {en} }