Sentiment Analysis of Code-Mixed Malaysian Social Media Text: Translation Subsystem

Wong, Hoong Lik (2020) Sentiment Analysis of Code-Mixed Malaysian Social Media Text: Translation Subsystem. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Wong Hoong LIK_Fulltext.pdf Restricted to Registered users only Download (1MB)

Abstract

In multilingual countries like Malaysia, the use of code-mixed text (English-Malay-Slang) is increasing in social media. This subsystem will only focus on translation of code-mixed text. There are several elements to be translated including abbreviations, misspelt words, and slang words. Also, the most challenging part would be achieving translation without using parallel corpus. Therefore, the translation system will apply MUSE library to generate cross-lingual embedding in order to perform word-to-word translation by using only monolingual corpus. But the result is only able to perform with a single word translation and would risk losing the syntactic structure of the sentence. So, the proposed system will apply the Moses phrase table to randomly select a phrase from the monolingual corpus and uses the cross-lingual embedding to generate the translation result and scores the results in each trained phrase. Thus, the translator will be able to perform translation without losing its synthetic structure. However, the result generated from the phrase table is still not good enough according to the BLEU evaluation scores. Then, the enhancement will apply the iterative back-translation to perform translation by continuously reversing the source and target language to improve the translation quality. Hence, this proposed system benefits the sales and marketing department to further understand customer’s feedback, comments, and inquiries. Also, the translated results would be passable to the subjectivity and emotional analysis team for further evaluation.