Getting the Null Subject Right with Neural Machine Translation

Ferlito, Federico (2021) Getting the Null Subject Right with Neural Machine Translation. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text Thesis_Final_Federico_fixed.pdf Download (1MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (99kB)

Abstract

Null subjects are non overtly expressed subject pronouns found in pro-drop languages, such as Italian, Greek and Spanish. In the past, translating null subjects into a non-pro drop language, where the subject must be explicit, had shown to be problematic for older MT systems.The current state-of-the-art of MT offers many benefits compared to the previous methods, however there is limited research that investigates their quality during null-subject translation. In this project, we quantify and compare the occurrence of the null-subject for several languages in the Europarl corpus. Next, we evaluate null subjects’ translation into English, a “non pro-drop” language. We do so by training various NMT methods which are compared on their ability to generate the correct subjects during the null-subject translation, and their ability to produce quality translations. With the results, we determine the improvement compared to the previous research on the topic, explaining which mechanism allowed the models to overcome the difficulties in this task. Finally, we measure the bias of generated subjects with regard to gender, and we propose a novel method to alter the training data with the aim of reducing the bias

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Spenader, J.K.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	15 Sep 2021 11:42
Last Modified:	15 Sep 2021 11:42
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/26094

Actions (login required)

View Item