Natural Language Processing Model for Maltese Syntax

Abstract

The objective of this thesis is to create a Natural Language Processing Model for the Maltese Language. The ultimate goal is that the model would be able to recognise syntactical features, that is the linguistic features and the relationship of a sequence of words, in Maltese. The performance and accuracy of the Maltese model is compared with the models of languages that have great influence on the Maltese language. The results outputted by the dependency parser were linguistically analysed to provide in depth analysis of the results outputted during training and testing. The model is tested on unseen text to provide a further understanding of the level of accuracy of the machine learning algorithm. For this syntax annotator, the model created is trained on manually annotated data and then the output is syntax data that is processed by the dependency parser and part-of- speech tagger. This model is made using the Python package spaCy. Since every language is unique, the linguistic rules are evaluated, to teach the model the rules of the language being researched. The MUDTv1 corpus developed by Slavomír Céplö for his Phd Thesis is used to train this model. The results show that the Maltese syntax model had a 91% part-of-speech tag accuracy, 74% unlabelled attachment score and 66% labelled attachment score. The model is further tested on unseen non-annotated text, the tag accuracy is 75% and the tokeniser accuracy is 99%.