Zero-shot Learning for Grapheme to Phoneme Conversion with Language Ensemble

Findings of the Association for Computational Linguistics: ACL 2022, pages 2106 - 2115 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Zero-shot Learning for Grapheme to Phoneme Conversion with Language Ensemble Xinjian Li and Florian Metze and David R Mortensen Shinji Watanabe and Alan W Black Language Technologies Institute, Carnegie Mellon University {xinjianl, fmetze, dmortens, swatanab, awb}@cs.cmu.edu Abstract Grapheme-to-Phoneme (G2P) has many appli- cations in NLP and speech fields. Most existing work focuses heavily on languages with abun- dant training datasets, which limits the scope of target languages to less than 100 languages. This work attempts to apply zero-shot learn- ing to approximate G2P models for all low- resource and endangered languages in Glot- tolog (about 8k languages). For any unseen target language, we first build the phylogenetic tree (i.e. language family tree) to identify top-k nearest languages for which we have training sets. Then we run models of those languages to obtain a hypothesis set, which we combine into a confusion network to propose a most likely hypothesis as an approximation to the target language. We test our approach on over 600 un- seen languages and demonstrate it significantly outperforms baselines.1 1 Introduction Grapheme-to-Phoneme (G2P) plays a crucial role in many NLP tasks. In particular, it is used heav- ily in many speech-related tasks such as speech recognition and speech synthesis (Arık et al., 2017; Miao et al., 2015). Even in the latest end-to-end systems, it still has a strong impact on the speech performance (Hayashi et al., 2021). Typically, the G2P task is language-dependent—many language- specific factors affect the G2P process such as the general characteristics of scripts (Ager, 2008), phonotactic constraints (Hayes and Wilson, 2008) and other orthography factors (Frost and Katz, 1992). For example, in Table 1, Mandarin and Japanese are not using the Latin script, therefore they cannot share their G2P models with English. As a consequence, to develop a G2P model, we need either to create a training set for the target language, like (CMU, 2000), or to ask linguists to 1Our code would be available at https://github. com/xinjli/transphone Language Grapheme Phoneme English hello /h@l@U/ Mandarin 你好 /nixAU/ French bonjour /bOZuK/ German hallo /halo/ Japanese こんにちは /konnichiwa/ Spanish hola /ola/ Table 1: A small sample of G2P examples from high- resource languages in our training set. explicitly define a set of orthographic rules to map from graphemes to phonemes (Mortensen et al., 2018). Both approaches have achieved success for high-resource languages; however, they can only account for a small number of the world’s languages. The majority still do not have access to G2P due to limited training resources. A good G2P model would be beneficial to many speech tasks in low-resource languages (Li et al., 2020a,b; Yan et al., 2021) In this work, we attempt to tackle this chal- lenging problem by using the language ensem- ble approach. Our approach allows us to propose an approximated G2P baseline to all languages present in the GlottoLog database: around 8000 of them (Nordhoff and Hammarström, 2011). …

Content truncated to 500 words