From Harmony to Probability: The Problem of Identifiability and a Bayesian Inference Perspective on Greek Nominal Stress

Maximum Entropy and Gradient Harmonic Grammar are well-established grammatical models for the analysis of linguistic phenomena, but we demonstrate that their probabilistic versions are inherently unidentifiable, employing parameters that cannot be uniquely determined from empirical data. Through examining Greek nominal stress patterns, we propose a reparameterization approach that introduces two identifiable, phonologically interpretable, parameters. We employ both point estimation and Bayesian inference to calculate their values. The latter approach yields reliable parameter estimates with quantified uncertainty that point estimation cannot offer. Our findings shed light on lesser-known aspects of Greek stress and offer a robust methodological framework for probabilistic phonological modeling across languages.