Exploring Vector Representations for Phonological Similarity

Recent research has compared representation models of word meaning (Brown et al., 2023, Cognitive Science 47:e13291), however, less research has compared representation models of words' perceptual features. Thus, we compared vector-space word representation models that can be used to quantify words' phonological similarity. The main structure of the model was adapted from Cox et al.'s orthographic representation model (Behavior Research Methods 43:602-15, 2011). Variations of the model included phonetic mapping scheme, encoding scheme, the inclusion of lexical stress, and the combination of orthographic and phonological representations. We tested the model variants against human-rated phonological similarity and both phonological and orthographic Damerau-Levenshtein distance. Open n-gram encoding (1 ≤ n ≤ 2) performed better overall than terminal relative encoding across all phonological similarity metrics. Concatenated orthographic-phonological vectors improved the prediction of human ratings with terminal-relative encoding only. Using more fine-grained phonetic mapping or including lexical stress had minimal effects.