Singlish Message Paraphrasing: A Joint Task of Creole Translation and Text Normalization

Proceedings of the 29th International Conference on Computational Linguistics, pages 3924–3936 October 12–17, 2022. 3924 Singlish Message Paraphrasing: A Joint Task of Creole Translation and Text Normalization Zhengyuan Liu†, Shikang Ni‡∗, Ai Ti Aw†, Nancy F. Chen† †Institute for Infocomm Research, A*STAR, Singapore ‡University of Cambridge, UK {liu_zhengyuan,nfychen}@i2r.a-star.edu.sg Abstract Within the natural language processing com- munity, English is by far the most resource- rich language. There is emerging interest in conducting translation via computational ap- proaches to conform its dialects or creole lan- guages back to standard English. This com- putational approach paves the way to leverage generic English language backbones, which are beneficial for various downstream tasks. How- ever, in practical online communication sce- narios, the use of language varieties is often accompanied by noisy user-generated content, making this translation task more challenging. In this work, we introduce a joint paraphrasing task of creole translation and text normaliza- tion of Singlish messages, which can shed light on how to process other language varieties and dialects. We formulate the task in three differ- ent linguistic dimensions: lexical level normal- ization, syntactic level editing, and semantic level rewriting. We build an annotated dataset of Singlish-to-Standard English messages, and report performance on a perturbation-resilient sequence-to-sequence model. Experimental re- sults show that the model produces reasonable generation results, and can improve the perfor- mance of downstream tasks like stance detec- tion. 1 Introduction While the development of natural language process- ing (NLP) has been focused on major languages such as English, Chinese, and French, there is emerging research interest in similar languages, varieties, and dialects (Zampieri et al., 2020). The distinction of language variations generally comes from geographical, historic, communicative set- tings, and social group dimensions (Coseriu, 1981). In particular, creole languages are formed in condi- tions when major languages (e.g., English, French) ∗ The contribution of Shikang Ni is conducted during his student internship in Institute for Infocomm Research, A*STAR, Singapore. are adopted in another culture or region, and they often mix with existing languages and evolve into other varieties in their own right. Such examples include the French-based Haitian (Degraff, 1992), English-based Australian Kriol (Harris et al., 1993), and Colloquial Singaporean English (Singlish) (Ho et al., 1993). To adopt computational NLP solu- tions on dialects or creoles, applying existing mod- els trained with major language resources will re- sult in degraded performance, and collecting and annotating sufficient data for task-specified domain adaptation is time-consuming and labor-intensive. Similar to studies on multilingual scenarios (Bal- ahur and Turchi, 2012; Eriguchi et al., 2018), one straightforward and effective approach is to con- form the varieties to their base languages by ma- chine translation (Zbib et al., 2012), then other NLP systems (e.g., sentiment analysis, information retrieval, reading comprehension) that take base languages as input could be applied. However, it would be challenging to apply this approach to di- alects or creoles that include certain deviated gram- mar and local vocabularies with systems trained on the corresponding standard languages, especially to the under-resourced language varieties …

Content truncated to 500 words