Kiki or Bouba? Sound Symbolism in Vision-and-Language Models

Although the mapping between sound and meaning in human language is assumed to be largely arbitrary, research in cognitive science has shown that there are non-trivial correlations between particular sounds and meanings across languages and demographic groups, a phenomenon known as sound symbolism. Among the many dimensions of meaning, sound symbolism is particularly salient and well-demonstrated with regards to cross-modal associations between language and the visual domain. In this work, we address the question of whether sound symbolism is reflected in vision-and-language models such as CLIP and Stable Diffusion. Using zero-shot knowledge probing to investigate the inherent knowledge of these models, we find strong evidence that they do show this pattern, paralleling the well-known kiki-bouba effect in psycholinguistics. Our work provides a novel method for demonstrating sound symbolism and understanding its nature using computational tools. Our code will be made publicly available.

Paper

Similar papers

Peer review

Reviewer DdD96/10 · confidence 3/52023-07-03

Summary

The Kiki-bouba effect is a well studied phenomena in humans to consistently associate sharp and smooth objects with certain phonetics. This paper explores whether such an effect is present within image and language models. Specifically it looks at Stable Diffusion (a text to image generative model) and CLIP (a discriminative model for text and images) to see whether these models form representations of made up "Kiki" and "Bouba" type words in similarly consistent ways as humans do. The paper is well written, and concludes that these models do indeed demonstrate this same tendency.

Strengths

The paper is well written. The methods used to investigate whether this effect is present in these different models appear valid. The observed effect is strong. It's interesting that these models demonstrate this as strongly as they do given they have no auditory dimension, but then again they are models of language. Nice observation though!

Weaknesses

It would have been nice to make use of a model that doesn't have have any exposure to images. For example taking BERT and seeing if the pseudo words also get separated with the same tendency or not? Something analogous to demonstrating kiki-bouba in blind people, which to my lay understanding I believe has been done but only to a limited extent compared to other validations of the effect in various human populations. I'm unsure if this paper opens up any further significant study, or unlocks anything in an engineering sense, so I think its impact is likely capped. Minor: citation on line 174 would be good, rather than just "some prior works". Nitpick: Line 24 -- "bœuf" is often translated as "beef" I thought, in which case there is a lot of overlap in sounds. This link is no doubt cultural, rather than saying anything about objects having similar sounding terms across languages. Maybe there is a better example to use though?

Questions

Did you try other ways of scoring? Does distance from centroids of the sharp/smooth groups also display the same effect for example? How important to observing this effect are the details in how scoring is done?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The authors do address the fact that these models don't have the same full set of inputs as humans do. They also note that nothing is discussed about why these models display these traits. These are important to highlight, but the paper has value as an observation that the effect is present.

Reviewer kkX77/10 · confidence 3/52023-07-04

Summary

The work's goal is to study whether sound symbolism is reflected in vision-and-language (VL) models like CLIP and Stable Diffusion (SD). The work proposed a method called **zero-shot knowledge probing**, and verified a sound symbolism phenomenon like **kiki-bouba effect** by evaluating the outputs from CLIP and SD when inputs were a set of predefined pseudowords, adjectives, and nouns, which are related to sharpness or roundness. In short, the study provides a simple paradigm for investigating sound symbolism in VL models from VL latent space and verifies the existence of sound symbolism in VL models.

Strengths

1. The work gives a clear explanation of motivation, methodology, and experimental results. 2. The work discusses the experimental results thoroughly and with some insights. 3. The work provides rich supplementary material, which answers a few questions raised when I was reading the main paper. 4. The work is well organized and tells a beautiful story with linguistic and cognitive backgrounds. In conclusion, I believe the work is solid and insightful.

Weaknesses

1. The work only studies open-sourced VL models CLIP and SD, which may relate to narrow computational resources. But studying models with different backbones like GAN or ViT would be more convincing. 2. The **zero-shot knowledge probing** is designed a bit intuitively without comparison with other baseline methods, such as training an individual classifier for text-text and text-image scores or human annotators. 3. The work did not discuss the inner results sound symbolism phenomenon in perspectives of the mechanism of machine learning but attributed it to the inherent knowledge in the models. I think the reason for the phenomenon might be a result of tokenization, i.e. a letter-level tokenized VL model / language model might produce different results. 4. The work only studies the phenomenon in English, given L41 saying the phenomenon is universal to different languages.

Questions

1. Could you give a result on how exactly $|w_{adj}|$ and $|v_{pw}|$ are in L206 and L217 since the distribution of the adjectives and pseudowords in latent space is quite a blur? 2. The caption of Table 2 is not clear, how is the list of words sorted and what does the `POS` line mean? 3. Why do you choose SD to map texts to images, rather than using retrieval methods with CLIP on the LAION-2B dataset? They have the exactly same inherent knowledge. 4. What are the probable applications of this study?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Although the work emphasizes that its study should be taken in context (L309), the inner reasons why VL models have the same sound symbolism effect as humans are not investigated. It should be addressed in future works.

Reviewer gvCK7/10 · confidence 4/52023-07-06

Summary

The authors present an investigation of the phenomenon of ‘sound symbolism’ in a pair of cutting edge Vision-and-Language Models. Sound symbolism is an intriguing phenomenon whereby the meaning of a word can be in part traced back to the way the word sounds. It is an important phenomenon in psychology and linguistics because it challenges a strict view of the mapping between form and meaning as arbitrary (stated most famously by Saussure). The authors probe CLIP and Stable Diffusion by creating a small dataset of pseudowords that were constructed to reflect phonetic features that generally map onto “sharp” vs. “round” speech sound categories. They project the embedding vectors in 1024-dimensional CLIP space onto a one-dimensional semantic dimension of interest, defined by two sets of antonym shape adjectives (one set corresponding to synonyms of ‘round’ and the other corresponding to synonyms of ‘sharp’). They also define a phonetic score to measure phonetic/graphemic associations of the shape adjectives with the pseudowords. They analyze the models using these geometric measures, as well as with a human evaluation where participants are asked to select which of two Stable Diffusion-generated images is best described by a pseudoword. Across these evaluations, they find relatively strong evidence for the presence of sound symbolism in these models.

Strengths

One of the strengths of this paper is that it demonstrates how ideas from Cogntive Science can drive investigations into modern AI models. The phenomenon of sound symbolism is particularly interesting both for its intuitive appeal and its relation to foundational ideas in Cognitive Science. The paper was well written and largely a pleasure to read. The methods were overall well described and I found the evidence relatively convincing.

Weaknesses

The weakest part of the paper for me were the motivations, hypotheses, and implications. Yes, Vision&Language Models are increasingly powerful and increasingly deployed, but why are they interesting objects of study for Cognitive Science? We are told that “these methods could provide new insights into the classic questions of what aspects of sound are tied to meaning”, but not explained how this might happen or what this could look like. As well, the conclusion about “cultural universality” seems like a very difficult and non-obvious question that these methods would provide insight into. And the flip-side, why is sound symbolism an important phenomenon to investigate in these powerful and deployable models? One specific weakness is that there were no hypotheses presented for the models, and you could imagine developing the hypotheses based on how these models were trained. I think there’s a lot to say about this, but all we got one sentence buried in the results section that mentioned that the models primarily saw valid English text during training and didn’t have access to the actual sound of the words. The second major weakness for me was the description and analysis of the User Study. The authors write “Over two hundred volunteers participated in this study”. The use of the word “volunteers” here suggests that the participants were not paid, which I hope is not the case. Who are these participants? How were they recruited? Was there any ethical review of this study? How many participants exactly were recruited (this is in the supplement, but should be presented in the main text)? On the analysis side, the use of a binomial test is not appropriate as the individual data points are not i.i.d. Participants gave multiple responses and, presumably, multiple items (pairs of images) were rated by multiple participants. Thus, something akin to a mixed-effects (multi-level / hierarchical) logistic regression model would be more appropriate to account for the participant- and item-wise variability, and provide a much more realistic analyses of the results (see Gelman & Hill, 2006 for background). I found the Phonetic Scores in 3.2 the most difficult to understand and still am not completely sure I get it. From the math, I would have thought this score was something akin to a control measurement, and a way to contextualize the Geometric scores, but the authors seem to have a different perspective, which I didn’t fully get. Finally, Figures 2 & 3 took up a lot of space for not that much information. They could easily be condensed to make room for some of the context in the response.

Questions

My questions are the same as the weaknesses. I’m not sure all of questions 1-3 need to be answered, but at least 2 of them do. 1. How can these results inform theories in Cognitive Science? 2. What hypotheses can be derived from these models a priori about sound symbolism, presumably by thinking about their training data? 3. Why is sound symbolism important for evaluating sophisticated, deployable AI systems? 4. Can you unpack the Phonetic Scores in a little more detail, and explain their relevance? 5. Please provide more details about the User Study and analyze the results in a more appropriate manner.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

4 excellent

Limitations

The limitations were okay but could be improved by thinking more about the societal impact of these (human) biases getting incorporated into deployable generative models. I would bet some of the impacts of sound symbolism have been discussing in the psychological literature, and it would be important to mention them and discuss them in the context of these things getting baked into generative AI.

Reviewer NTGx8/10 · confidence 4/52023-07-07

Summary

This paper determines the extent to which pretrained vision and language models encode phonetic information associated with sharp or round objects. Prior research has shown a cross-lingual tendency to associate sounds with shapes in human studies. In this work, the authors investigate whether this holds for a discriminative (OpenCLIP) or generative (StableDiffusion) machine learning model. The methodology starts by constructing a set of 648 pseudowords based on a set of sharp / round Latin letters. Two prompts are used with the models, which either have the pseudoword used as a noun or an adjective. The prompts and pseudowords are used to show that both models do indeed exhibit a similar effect to humans. A human study shows a much stronger effect for the original kibi-bouba pair than the 648 pseudowords. The paper also claims that this finding cannot be attributed to the models seeing these examples during pretraining.

Strengths

* Interesting study that is well-motivated. * Claims are clearly supported in the experiments. * Human evaluation is used to further support the automatic evaluation metrics.

Weaknesses

* No major weaknesses that I could identify

Questions

* Do these results hold with a multilingual CLIP model or in a multilingual human evaluation?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

* The paper does a good job of discussing potential limitations of the research

Reviewer kkX72023-08-11

Thanks for your response. > the concern about multilingual results in our global response (item B) It is a reasonable reply. It will be better if the linguistic feature of different languages is considered in the additional experiments. > We note that these models differ in underlying architecture, training objectives, and with respect to training data. "training objectives": I do not think their training objectives differ a lot. It might still be worth experimenting with GAN-style models. > While SD and (Open)CLIP were both trained on LAION, the intrinsic knowledge learned by SD’s denoising goal (which requires understanding local regions within images) is likely to be different from knowledge gained from image-level retrieval. From this perspective, you should define "intrinsic knowledge" more clearly that it is the knowledge learned by denoising training objective of SD in the revised paper.

Authorsrebuttal2023-08-12

Thank you for your reply and additional comments. Regarding linguistic features of languages tested, we will provide additional details, such as the language families of each language tested (Hungarian - Uralic/Ugric, Indonesian - Austronesian/Malayo-Polynesian, Finnish - Uralic/Finnic, Lithuanian - Indo-European/Baltic) and explaining how these differ from one another and from English, in a revised version. Regarding GAN-style models, we have also experimented with GALIP [1], a GAN-based text-to-image generation model. Using GALIP pretrained on the CC12M dataset with our pseudoword methodology, we see significant sound symbolic associations via our metrics and qualitatively - for example, phonetic score AUC 0.62 and geometric score AUC 0.98, significantly higher than 0.50 expected by chance for each. We will include these results in a revised version. Regarding intrinsic knowledge, we will clarify this wording in a revised version. [1] Tao et al. GALIP: Generative Adversarial CLIPs for Text-to-Image Synthesis. CVPR 2023

Reviewer NTGx2023-08-14

Thank you for running an additional experiment using a multilingual model. The results are interesting and it makes me more confident in my assessment of the work.

Reviewer gvCK2023-08-16

Thank you

I appreciate the very thoughtful rebuttal. Your explanation surrounding the motivation for the study (i.e., importance in the context of CogSci and a priori hypotheses about models) is well presented and thought provoking. I think this was the key thing missing for me with respect to motivation, and you’ve adequately addressed my concern here. So please put this in the paper. I also like your comments regarding the importance in evaluating AI systems with this perspective. So please include that in the paper as well. I also very much appreciate you implementing a more principled statistical analysis, so well done. Finally, regarding the user study: I don’t have strong feelings about whether to include or not (but you may want to consult the Ethical guidelines at NeurIPS about its inclusion). What I do feel strongly about is that you describe the protocol with more precise language (as you did in the rebuttal). Eg., in the paper you should write the population of participants you recruited (i.e., university graduate students) and that they were unpaid volunteers. Also, reporting the precise number that you collected in the main text, etc.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC