Bi-Phone: Modeling Inter Language Phonetic Influences in Text

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2580–2592 July 9-14, 2023 ©2023 Association for Computational Linguistics Bi-Phone: Modeling Inter Language Phonetic Influences in Text Abhirut Gupta1, Ananya B. Sai2, Richard Sproat1, Yuri Vasilevski1, James S. Ren1, Ambarish Jash1, Sukhdeep S. Sodhi1 and Aravindan Raghuveer1 1Google Research 2IIT Madras Corresponding author: abhirut@google.com Abstract A large number of people are forced to use the Web in a language they have low liter- acy in due to technology asymmetries. Writ- ten text in the second language (L2) from such users often contains a large number of errors that are influenced by their native lan- guage (L1). We propose a method to mine phoneme confusions (sounds in L2 that an L1 speaker is likely to conflate) for pairs of L1 and L2. These confusions are then plugged into a generative model (Bi-Phone) for synthetically producing corrupted L2 text. Through hu- man evaluations, we show that Bi-Phone gen- erates plausible corruptions that differ across L1s and also have widespread coverage on the Web. We also corrupt the popular language un- derstanding benchmark SuperGLUE with our technique (FunGLUE for Phonetically Noised GLUE) and show that SoTA language under- stating models perform poorly. We also intro- duce a new phoneme prediction pre-training task which helps byte models to recover per- formance close to SuperGLUE. Finally, we also release the FunGLUE benchmark to pro- mote further research in phonetically robust language models. To the best of our knowl- edge, FunGLUE is the first benchmark to in- troduce L1-L2 interactions in text. 1 Introduction We live in a multilingual world with over 7,000 languages spoken across the globe (Eberhard and Fennig, 2022). However, technology asymmetri- cally supports only a few specific languages. For instance, the internet is mostly in English with over 60% of websites using the language despite just around 16% share of its speaking population around the world1 (Grefenstette and Nioche, 2000). Increasingly, people are forced to navigate and pro- duce content on the web in languages they have not been formally trained on. The English text pro- duced by ESL (English as Second / L2 language) 1https://w3techs.com/technologies/overview/content_language writers is heavily influenced by their native lan- guage (L1). Research in the field of second-language acqui- sition has found evidence of phoneme-shift based misspellings stemming from L1 influence in L2 text for specific language pairs (Ibrahim, 1978; Cook, 1997; Bestgen and Granger, 2011; Sari, 2014; Ogneva, 2018; Motohashi-Saigo and Ishizawa, 2020). Studies in Natural Language Understand- ing (NLU) have been limited to spelling correc- tion Nagata et al. (2017); Flor et al. (2019) and native language identification Chen et al. (2017); Nicolai et al. (2013) in English learners. These studies predominantly use the TOEFL11 dataset (Blanchard et al., 2013) which deals with very spe- cific demographics such as test-takers who have formal training in the L2 language. We make the following four key observations about prior work in the study of L1-L2 influences in text and speech. First, …

Content truncated to 500 words