Detection of wordplay generated by reproduction of letters in social media texts

Wordplay generated by letters of its original word being repeated is commonly found in social network texts. Most of the time, wordplay items of this type are ambiguous to machines in language processing tasks such as Text-to-Speech. This paper shows some statistics on the number of letters from 102,586 real social network text items and proposes a set of classification features together with a few classification frameworks to detect repeated-letter wordplay tokens from Thai social network texts, which were tokenized by CRF-based Thai word segmentation. Evaluation on 48,949 text items shows that the proposed method achieves the detection accuracy of 98.45% which is an improvement over simple rule-based and some previously proposed methods.