PANCETTA: Phoneme Aware Neural Completion to Elicit Tongue Twisters Automatically

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 491–504 May 2-6, 2023 ©2023 Association for Computational Linguistics PANCETTA: Phoneme Aware Neural Completion to Elicit Tongue Twisters Automatically Sedrick Scott Keh 1, Steven Y. Feng∗†2, Varun Gangal†1, Malihe Alikhani 3, Eduard Hovy 1 1Carnegie Mellon University, 2Stanford University, 3University of Pittsburgh {skeh,vgangal,hovy}@cs.cmu.edu syfeng@stanford.edu, malihe@pitt.edu Abstract Tongue twisters are meaningful sentences that are difficult to pronounce. The process of au- tomatically generating tongue twisters is chal- lenging since the generated utterance must sat- isfy two conditions at once: phonetic diffi- culty and semantic meaning. Furthermore, pho- netic difficulty is itself hard to characterize and is expressed in tongue twisters through a heterogeneous mix of phenomena such as alliteration and homophony. In this paper, we propose PANCETTA: Phoneme Aware Neural Completion to Elicit Tongue Twisters Automatically. We leverage phoneme represen- tations to capture the notion of phonetic diffi- culty, and we train language models to gener- ate original tongue twisters on two proposed task settings. To do this, we curate a dataset called TT-Corp, consisting of existing English tongue twisters. Through automatic and hu- man evaluation, as well as qualitative analysis, we show that PANCETTA generates novel, phonetically difficult, fluent, and semantically meaningful tongue twisters. 1 Introduction A tongue twister is a sentence which is both artic- ulatorily difficult (i.e. colloquially speaking, hard to say or "twisting") while at the same time being meaningful and fluent. Some examples of tongue twisters are shown in Table 1. Together with riddles, rhymes, fables, and other such creative artifacts, tongue twisters were histor- ically often employed as a vehicle for early trans- mission of native language diction, grammar, and vocabulary to children, through parent-child inter- action, playtime activity, and kindergarten instruc- tion (Akinyemi, 2003; Mcgovern, 2021). Tongue twisters have also been used as experimental aids for research studies of speech production in cog- nitive science and related disciplines, both among ∗Work done while at CMU. † Equal contribution by Steven and Varun Figure 1: Overview of the phoneme-aware training in the PANCETTA model. healthy speakers and those with speech and audi- tory disorders such as dysarthria (Kember et al., 2017). They are also used as pedagogic aids in speech therapy, as well as for treatment of speech disorders and psychological disorders relating to public speaking and elocution (Revathy and Ravin- dran, 2016). An example of this was in a scene1 from The King’s Speech (2010), where George VI repeats a tongue twister during therapy to reduce his stutter. Lastly, tongue twisters find use as teach- ing aids for English diction in EFL (English as a Foreign Language) instructional settings (Proši´c- Santovac, 2009). The coining of a novel, unique tongue twister which spreads sufficiently to become normative and well-recognized is rare, hence characterizing them as long-tailed linguistic phenomena (Naik, 2022). However, they are not limited to English, and are found across the world’s languages, e.g., Persian ("Shish sikh jigar sikhi shi shezar.") (Jam, 2018) and French ("Cinq chiens chassent …

Content truncated to 500 words