Unsupervised Multiword Expression Extraction on Chinese-English Code-Mixed Malaysian Expressions on E-commerce Platform

Multiword Expression (MWE) extraction has been popular research in Natural Language Processing (NLP) area. This research aims to extract different MWE types, including Chinese (ZH), English (EN), Chinese-English (ZH-EN) code-mixed, and new word MWEs, from Malaysian Expressions (MSEs) in the e-commerce platform. MSEs are multilingual and contain code-mixed text as well as Out-Of-Vocabulary (OOV). Thus, they are highly unstructured. A comparative study is conducted using various statistical models such as Branch Entropy (BE), Normalized Branch Entropy (NBE) and Accessor Variety (AV). Each statistical model's performance is assessed using Top N precision. The study reveals the characteristics of ZH-EN code-mixed MSEs in e-commerce platforms and the contribution of punctuation towards the Top N precision of statistical models. The experiment results show that AV outperforms BE variations, whereas NBE outperforms conventional BE.