Exploring How Generative Adversarial Networks Learn Phonological Representations

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3115–3129 July 9-14, 2023 ©2023 Association for Computational Linguistics Exploring How Generative Adversarial Networks Learn Phonological Representations Jingyi Chen Department of Linguistics The Ohio State University chen.9220@osu.edu Micha Elsner Department of Linguistics The Ohio State University elsner.14@osu.edu Abstract This paper explores how Generative Adversar- ial Networks (GANs) learn representations of phonological phenomena. We analyze how GANs encode contrastive and non-contrastive nasality in French and English vowels by apply- ing the ciwGAN architecture (Beguˇs, 2021a). Beguˇs claims that ciwGAN encodes linguis- tically meaningful representations with cate- gorical variables in its latent space and manip- ulating the latent variables shows an almost one to one corresponding control of the phono- logical features in ciwGAN’s generated out- puts. However, our results show an interac- tive effect of latent variables on the features in the generated outputs, which suggests the learned representations in neural networks are different from the phonological representations proposed by linguists. On the other hand, ci- wGAN is able to distinguish contrastive and noncontrastive features in English and French by encoding them differently. Comparing the performance of GANs learning from different languages results in a better understanding of what language specific features contribute to developing language specific phonological rep- resentations. We also discuss the role of train- ing data frequencies in phonological feature learning. 1 Introduction Recent studies in natural language processing (NLP) have demonstrated two generic trends: neu- ral networks dominate language-specific machine learning models; the common practice of model training (pre-training and fine-tuning) outperforms many traditional training methods and is partic- ularly suitable for the development of language models used for various downstream tasks. These language models, however, are of black-box nature. The interpretability of these models is limited that the language representation they learned might not align to human language. How, then, to understand the opaque and complex learned representation of language models is an important question in recent studies. Phonology, the study of the sound system of human languages, plays an important role in understanding models’ inherent biases and their ability to make human-like generalizations. The sound systems of human languages are not organized arbitrarily, but contain structural gener- alizations and interdependence. Thus, learning a sound system involves not only learning to acousti- cally realize or recognize segments (phonetics), but also mapping them to an inventory characterized by distinctive features, and learning distributional con- straints on segment sequences (phonology). Just as computational psycholinguists have investigated the degree to which neural network language mod- els learn linguistically motivated features like syn- tax (Linzen et al., 2016; Lau et al., 2017; Gulor- dava et al., 2018; Marvin and Linzen, 2018; Futrell et al., 2019), they have also investigated the degree to which phonological organization emerges from neural models trained on acoustics (Gelderloos and Chrupała, 2016; Chrupała et al., 2017). The degree to which these models learn phono- logical features is still debatable. Recently, a neural network autoencoder seems to successfully learn phoneme-like …

Content truncated to 500 words