The learning the sounds of Japanese: Experimental and computational approaches

Silva Fonseca, Marco Aurelio

Permalink

https://hdl.handle.net/2142/121446 Copy

Description

Title

The learning the sounds of Japanese: Experimental and computational approaches

Author(s)

Silva Fonseca, Marco Aurelio

Issue Date

2023-07-05

Director of Research (if dissertation) or Advisor (if thesis)

• Hualde, Jose Ignacio
• Scwhartz, Lane

Doctoral Committee Chair(s)

• Hualde, Jose Ignacio
• Scwhartz, Lane

Committee Member(s)

• Montrul, Silvina
• Shosted, Ryan

Department of Study

Linguistics

Discipline

Linguistics

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Second language acquisition
• phonetics
• phonology
• neural networks, machine learning
• Japanese.

Abstract

In this dissertation I analyze the learning of the sounds of Japanese from experimental and computational perspectives. More specifically, I conducted a perception experiment with second language learners who are native speakers of English (henceforth L2 learners) and modeling experiments using long-short term memory recurrent neural networks (LSTM RNNs). By using experimental machine learning to model sound phenomena and their acquisition, this dissertation provides a novel methodology to place second language acquisition and phonology research within an interdisciplinary field. In chapter 2, I present an experiment that I conducted with 22 advanced L2 learners and a control group of 17 native speakers of Japanese. In this experiment I analyzed the perception of three types of lexical contrasts: a) voiceless versus voiced stops: kakkou “outfit” versus gakkou “school”; b) long versus short vowels: biru “building” versus biiru “beer”; and c) words with lexical pitch accent on their last syllable versus first syllable: am ́e “candy” versus ́ame “rain”. The results of the experiment show that even though L2 learners perform as well as native speakers in the ABX discrimination, their performance is worse than native speakers in the lexical assignment task. Accuracy was particularly low for the pitch accent contrast. This indicates that even though L2 learners can hear the difference between contrasts that manifest phonetically differently in their language, it does not mean that they fully acquired such contrasts in their mental lexicon. In chapter 3, I investigate the role of distinctive features vs. phoneme embeddings to model the sounds of English and Japanese using long-short term memory recurrent neural networks (LSTM RNNs). Besides building English and Japanese networks, I additionally built two different types of bilingual networks: simultaneous (trained with English and Japanese at the same time) and consecutive (first trained in English and then in Japanese) with a variety of English/Japanese ratios. For both monolingual and bilingual models, feature-na ̈ıve networks outperformed feature-aware networks. These results corroborate previous research (Mirea and Bicknell 2019), and provide additional evidence using Japanese and bilingual networks. In chapter 4 I focus on the model’s performance related to two phonological phenomena of Japanese. By doing so, I propose a novel approach to model the phonology of Japanese. More specifically, I used LSTM RNNS to model nasal assimilation and loanword geminate devoicing. While nasal assimilation is a local phenomenon (i.e., it affects sounds adjacent to each other), loanword geminate devoicing is a long-distance phenomenon (i.e., it affects sounds across syllable boundaries). The results indicate that both bilingual and monolingual networks can learn nasal assimilation but not loanword geminate devoicing. This could be because LSTM RNNs are not efficient to model long-distance phenomena, but could also be because of the nature of the training data. In summary, this dissertation investigates how L2 learners and LSTM RNNs learn the sounds of Japanese. The results of chapter 2 revealed that L2 learners find it difficult to assign meaning to words that differ in pitch accent; chapter 3 reveals that LSTM RNNs without distinctive features in their architecture outperform those with them (in terms of cross-entropy values assigned to test data set); and chapter 4 revealed that LSTM RNNs can learn nasal assimilation but cannot learn geminate loanword devoicing. By employing both experimental and computational models, this dissertation provides evidence for the development of new techniques to conduct research on the acquisition of sounds by L2 learners.

Graduation Semester

2023-08

Type of Resource

Thesis

Copyright and License Information

Copyright 2023 Marco Aurelio Silva Fonseca

Owning Collections