Published papers

A previous study observed that the association between utterances and their ascribed meaning depends, among other factors, on phonation types. Specifically, the perceived size and shape of the referents of pseudo-words varied depending on whether they were pronounced with modal, creaky, falsetto or whisper phonation. In this study, we report a re-analysis of the same shape and size subjective ratings in terms of the four psychoacoustic measures: loudness, roughness, sharpness and pitch. We found that phonation types were positively associated with psychoacoustic features-falsetto with pitch, whisper with sharpness, and creakiness with loudness and roughness. Psychoacoustic features were also associated with subjective ratings: loudness and sharpness with subjective size and pitch and roughness with subjective shape. These findings indicate that psychoacoustic features (closer to the auditory representation of the heard sounds) may be good predictors of sound symbolism judgments.

This study constructs machine learning algorithms that are trained to classify samples using sound symbolism, and then it reports on an experiment designed to measure their understanding against human participants. Random forests are trained using the names of Pokémon, which are fictional video game characters, and their evolutionary status. Pokémon undergo evolution when certain in-game conditions are met. Evolution changes the appearance, abilities, and names of Pokémon. In the first experiment, we train three random forests using the sounds that make up the names of Japanese, Chinese, and Korean Pokémon to classify Pokémon into pre-evolution and post-evolution categories. We then train a fourth random forest using the results of an elicitation experiment whereby Japanese participants named previously unseen Pokémon. In Experiment 2, we reproduce those random forests with name length as a feature and compare the performance of the random forests against humans in a classification experiment whereby Japanese participants classified the names elicited in Experiment 1 into pre-and post-evolution categories. Experiment 2 reveals an issue pertaining to overfitting in Experiment 1 which we resolve using a novel cross-validation method. The results show that the random forests are efficient learners of systematic sound-meaning correspondence patterns and can classify samples with greater accuracy than the human participants.

Introduction: This paper presents a cross-linguistic study of sound symbolism, analysing a six-language corpus of all Pokémon names available as of January 2022. It tests the effects of labial consonants and voiced plosives on a Pokémon attribute known as friendship. Friendship is a mechanic in the core series of Pokémon video games that arguably reflects how friendly each Pokémon is. Method: Poisson regression is used to examine the relationship between the friendship mechanic and the number of times /p/, /b/, /d/, /m/, /g/, and /w/ occur in the names of English, Japanese, Korean, Chinese, German, and French Pokémon. Results: Bilabial plosives, /p/ and /b/, typically represent high friendship values in Pokémon names while /m/, /d/, and /g/ typically represent low friendship values. No association is found for /w/ in any language. Discussion: Many of the previously known cases of cross-linguistic sound symbolic patterns can be explained by the relationship between how sounds in words are articulated and the physical qualities of the referents. This study, however, builds upon the underexplored relationship between sound symbolism and abstract qualities.

The social categories that characterize a speaker frequently correlate with the use of specific linguistic variables. Research suggests that such correlations are sometimes recognized as socially-indexed meanings. This study examines Japanese individuals' attitudes toward variables that have been shown to correlate with the social category of gender in production. In particular, we contrast patterns of gendered variation that (i) have been prescriptively associated with speaker sex and (ii) tend to correlate with gender in speech production but are outside of the set of prescriptive "women's language". We found that individuals have formed associations between the gender of the speaker and prescriptive variables but not other patterns of variation. Additionally, knowledge of the speech context of the variables had no significant effect on individuals' judgments. The results indicate that not all social information available from patterns of language use is recovered by listeners. More broadly, examining the transmission of social meaning through linguistic variation requires a combination of production- and perception-based research methods.

Sound symbolism, systematic and iconic relationships between sounds and meanings, is now a topic that is very actively explored by linguists, psychologists and cognitive scientists. As a new research strategy to study the nature of sound symbolic connections across different languages, a number of scholars have started using Pokémon names, a research paradigm that is now dubbed "Pokémonastics."The previous Pokémonastics studies have experimentally explored how the evolution status is symbolically expressed by native speakers of English, Japanese and Brazilian Portuguese. Building on these studies, the current experiment examined the sound symbolic knowledge of Russian speakers, and found that they are more likely to associate large, post-evolution Pokémon characters with names containing voiced obstruents than with names containing voiceless obstruents, and that they are also more likely to associate post-evolution characters with names containing [a] than with names containing [i]. The experiment also revealed that Russian speakers are less likely to associate post-evolution characters with names having labial consonants than with names having coronal or dorsal consonants. Overall, the current results show that Russian speakers generally have knowledge of sound symbolic associations that is similar to that of English, Japanese and Brazilian Portuguese speakers, suggesting that some sound symbolic patterns hold robustly across multiple languages.

Modern Japanese has a set of morphophonemic alternations known collectively as rendaku that involve initial consonants in second elements of compounds, as in /jama+dera/ ‘mountain temple’ (cf. /tera/ ‘temple’). An alternating element like /tera/ ~ /dera/ has an initial voiced obstruent in its rendaku allomorph and an initial voiceless obstruent in its non-rendaku allomorph. Lyman's Law blocks rendaku in a second element containing a medial voiced obstruent. This paper gives three arguments that Lyman's Law originated as a constraint prohibiting prenasalisation in consecutive syllables. First, constraints on similar consonants in close proximity generally apply not to voicing but to features with phonetic cues that are more spread out, such as prenasalisation. Second, in some Japanese dialects with prenasalised voiced obstruents, rendaku cannot occur if it would result in adjacent syllables containing these marked consonants. Third, phonographically attested Old Japanese compounds are consistent with a constraint on adjacent syllables.

An issue that is actively explored in the contemporary linguistics literature is how to account for probabilistic generalizations, for which there are currently various competing theories. To bear on this debate, Hayes (2021) proposes that we examine a grammatical framework in terms of its quantitative signature, typical probabilistic patterns that the framework is predicted to generate. In this research report, I zoom in on the quantitative signature of maximum entropy harmonic grammar (MaxEnt HG), because this framework has proven useful in modeling probabilistic generalizations across different linguistic domains. Given a linear scale of violations of one constraint, MaxEnt yields a sigmoid curve. When another constraint is relevant and can be violated multiple times, this sigmoid curve can be shifted, yielding multiple sigmoid curves, which results in a stripy wug-shaped curve. Expanding upon a previous study (Kawahara 2020b), the current experiment demonstrates that we observe a stripy wug-shaped curve in a particularly clear manner in patterns of sound symbolism, systematic associations between sounds and meanings. Concretely, the experiment with Japanese speakers shows that the judgment of Pokémons’ evolution status is affected by the mora counts of nonce names, resulting in a sigmoid curve, and that this sigmoid curve is shifted according to the number of voiced obstruents contained in the names. The overall results suggest that MaxEnt is a useful tool for modeling systematic sound-meaning correspondences.*.

The bouba/kiki effect - the association of the nonce word bouba with a round shape and kiki with a spiky shape - is a type of correspondence between speech sounds and visual properties with potentially deep implications for the evolution of spoken language. However, there is debate over the robustness of the effect across cultures and the influence of orthography. We report an online experiment that tested the bouba/kiki effect across speakers of 25 languages representing nine language families and 10 writing systems. Overall, we found strong evidence for the effect across languages, with bouba eliciting more congruent responses than kiki. Participants who spoke languages with Roman scripts were only marginally more likely to show the effect, and analysis of the orthographic shape of the words in different scripts showed that the effect was no stronger for scripts that use rounder forms for bouba and spikier forms for kiki. These results confirm that the bouba/kiki phenomenon is rooted in crossmodal correspondence between aspects of the voice and visual shape, largely independent of orthography. They provide the strongest demonstration to date that the bouba/kiki effect is robust across cultures and writing systems. This article is part of the theme issue 'Voice modulation: from origin and mechanism to social impact (Part II)'.

© 2017 Elsevier Ltd In Tokyo Japanese, /u/ is typically devoiced between two voiceless consonants. Whether the lingual vowel gesture is influenced by devoicing or present at all in devoiced vowels remains an open debate, largely because relevant articulatory data has not been available. We report ElectroMagnetic Articulography (EMA) data that addresses this question. We analyzed both the trajectory of the tongue dorsum across VC 1 uC 2 V sequences as well as the timing of C 1 and C 2 . These analyses provide converging evidence that /u/ in devoicing contexts is optionally targetless—the lingual gesture is either categorically present or absent but seldom reduced. When present, the magnitude of the lingual gesture in devoiced /u/ is comparable to voiced vowel counterparts. Although all speakers produced words with and without a vowel height target for /u/, the frequency of targetlessness varied across speakers and items. The timing between C 1 and C 2 , the consonants flanking /u/ was also effected by devoicing but to varying degrees across items. The items with the greatest effect of devoicing on this inter-consonantal interval were also the items with the highest frequency of vowel height targetlessness for devoiced /u/.