Supervised Grapheme-to-Phoneme Conversion of Orthographic Schwas in Hindi and Punjabi

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7791–7795 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 7791 Supervised Grapheme-to-Phoneme Conversion of Orthographic Schwas in Hindi and Punjabi Aryaman Arora*, † Luke Gessler † Nathan Schneider † *School Without Walls High School †Georgetown University {aa2190, lg876, nathan.schneider}@georgetown.edu Abstract Hindi grapheme-to-phoneme (G2P) conver- sion is mostly trivial, with one exception: whether a schwa represented in the orthogra- phy is pronounced or unpronounced (deleted). Previous work has attempted to predict schwa deletion in a rule-based fashion using prosodic or phonetic analysis. We present the first statistical schwa deletion classifier for Hindi, which relies solely on the orthography as the input and outperforms previous approaches. We trained our model on a newly-compiled pronunciation lexicon extracted from various online dictionaries. Our best Hindi model achieves state of the art performance, and also achieves good performance on a closely re- lated language, Punjabi, without modification. 1 Introduction Hindi is written in the Devanagari script, which is an abugida, an orthographic system where the basic unit consists of a consonant and an optional vowel diacritic or a single vowel. Devanagari is fairly regular, but a Hindi word’s actual pronun- ciation can differ from what is literally written in the Devanagari script.1 For instance, in the Hindi word ��� ⟨pep@R@⟩ ‘paper’, there are three units � ⟨pe⟩, � ⟨p@⟩, and � ⟨R@⟩, corresponding to the pronounced forms [pe], [p@], and [r]. The second unit’s inherent schwa is retained in the pronounced form, but the third unit’s inherent schwa is deleted. Predicting whether a schwa will be deleted from a word’s orthographic form is generally difficult. Some reliable rules can be stated, e.g. ‘delete any schwa at the end of the word’, but these do not perform well enough for use in an application that requires schwa deletion, like a text-to-speech syn- thesis system. 1Throughout this paper, we will adopt the convention of us- ing ⟨angle brackets⟩ to describe how a word is literally spelled, and [square brackets] to describe how a word is actually pro- nounced. This work approaches the problem of predict- ing schwa deletion in Hindi with machine learning techniques, achieving high accuracy with minimal human intervention. We also successfully apply our Hindi schwa deletion model to a related lan- guage, Punjabi. Our scripts for obtaining machine- readable versions of the Hindi and Punjabi pronun- ciation datasets are published to facilitate future comparisons.2 2 Previous Work Previous approaches to schwa deletion in Hindi broadly fall into two classes. The first class is characterized by its use of rules given in the formalism of The Sound Pattern of En- glish (Chomsky and Halle, 1968). Looking to an- alyses of schwa deletion produced by linguists (e.g., Ohala, 1983) in this framework, others built schwa deletion systems by implementing their rules. For example, this is a rule used by Narasimhan et al. (2004), describing schwa deletion for words like ���� ⟨dZ@Ng@li:⟩: C V C C a C V C V C …

Content truncated to 500 words