Complex word identification model for lexical simplification in the Malay language for non-native speakers

• Main Author: Salehah, Omar
• Format: Thesis
• Language: English; English
• Published: 2023
• Subjects: T Technology (General)
• Online Access: https://etd.uum.edu.my/10852/1/permission%20to%20deposit-grant%20the%20permission-s825765.pdf; https://etd.uum.edu.my/10852/2/s825765_01.pdf; https://etd.uum.edu.my/10852/

Text Simplification (TS) is the process of converting complex text into more easily understandable text. Lexical Simplification (LS), a method in TS, is the task of converting words into simpler words. Past studies have shown weaknesses in the LS first task, called Complex Word Identification (CWI), where simple and complex words have been misidentified in previous CWI model. The main objective of this study is to produce a Malay CWI model with three sub-objectives, i) To propose a dataset based on the state-of-the-art Malay corpus, ii) To produce a Malay CWI model, and iii) To perform an evaluation based on the standard statistical metrics; accuracy, precision, recall, F1-score, and G1-score. This model is constructed based on the development of the CWI model outlined by the previous researcher. This study consists of three modules, i) A Malay CWI dataset, ii) Malay CWI features with the new enhanced stemmer rules, and iii) A CWI model based on the Gradient Boosted Tree (GB) algorithm. The model is evaluated based on a state-of-the-art Malay corpus. This corpus is divided into training and testing data using k-fold cross-validation, where k=10. A series of tests were performed to ensure the best model was produced, including feature selection, generation of an improved stemmer algorithm, data imbalances, and classifier testing. The best model using the Gradient Boost algorithm showed an average accuracy of 92.55%, F1- score of 92.09% and G1-score of 89.7%. The F1-score was better than the English standard baseline score, with an increased difference of 16.3%. Three linguistic experts verified the results for 38 unseen sentences, and the results showed significantly positive results between the model built and the linguistic experts’ assessment. The proposed CWI model has improved the F1- score that has been obtained in second CWI shared task and positively affected non-native speakers and researchers.