A multilingual semi-supervised approach in deriving Singlish sentic patterns for polarity detection

Due to the huge volume and linguistic variation of data shared online, accurate detection of the sentiment of a message (polarity detection) can no longer rely on human assessors or through simple lexicon keyword matching. This paper presents a semi-supervised approach in constructing essential toolkits for analysing the polarity of a localised scarce-resource language, Singlish (Singaporean English). Corpusbased bootstrapping using a multilingual, multifaceted lexicon was applied to construct an annotated testing dataset, while unsupervised methods such as lexicon polarity detection, frequent item extraction through association rules and latent semantic analysis were used to identify the polarity of Singlish ngrams before human assessment was done to isolate misleading terms and remove concept ambiguity. The findings suggest that this multilingual approach outshines polarity analysis using only the English language. In addition, a hybrid combination of the Support Vector Machine and a proposed Singlish Polarity Detection algorithm, which incorporates unigram and n-gram Singlish sentic patterns with other multilingual polarity sentic patterns such as negation and adversative, is able to outperform other approaches in comparison. The promising results of a pooled testing dataset generated from the vast amount of unannotated Singlish data clearly show that our multilingual Singlish sentic pattern approach has the potential to be adopted in real-world polarity detection.