This dataset explores mental associations, the intuitive connections people make between shapes, objects, sounds, or concepts that have no inherent relation. The most famous example is the Bouba–Kiki effect, where people reliably associate a spiky shape with the sound “kiki” and a rounded shape with “bouba.” (See: https://en.wikipedia.org/wiki/Bouba/kiki_effect)

Our dataset expands this experiment to 20 different association questions, each presenting respondents with two options and asking them to choose one based on a specific prompt (e.g., “Which one is called 'Kiki'?”).

In total, the dataset contains:

~200,000 responses
~10,000 responses per question
Respondents from all over the world
Demographic metadata (where available)

The data was collected using the Rapidata API, which enables fast global collection of human-generated judgments. If there is community interest, we may expand this dataset in the future. Please consider leaving a ❤️ if you enjoy this dataset.

Noteworthy Findings

We encourage the community to explore the data and share insights. See below for details on the dataset structure and how to load it. Below are some observed patterns uncovered so far.

1. The Kiki–Bouba effect… reversed?

When asking the question such as: “Which one is ‘Kiki’ and which one is ‘Bouba’?" the following trend emerges:

Spiky shape → “Kiki”
Blob shape → “Bouba”

However, in our dataset we find that globally the blob-like shape is selected more often (although marginally) when asking the question: “Which one is called ‘Kiki’?”. What is interesting however, is that, as shown below, when aggregating by language, Arabic speakers overwhelmingly choose the blob as “Kiki”, whereas Japanese speakers overwhelmingly choose the spiky shape.

While the global aggregate does not match the classic findings, some possible explanations may be:

Our phrasing (“Which one is called ‘Kiki’?”) removes the paired Bouba context. Typically respondents are informed that one is Bouba and the other is Kiki adding additional constraints.
We have observed that Arabic speakers may be overrepresented in the data.

2. “Which one is reckless?” Bowling ball vs rubber duck

In this question, we ask respondents "Which one is reckless?", and present them with an image of

A rubber duck
A bowling ball

Globally, the bowling ball is the clear winner, likely due to destructive potential.

However, Polish speakers strongly diverge, with a much higher proportion selecting the duck.

Dataset Structure

Each row corresponds to one of the 20 association questions, containing:

question: The association question presented (e.g., “Which one is ‘Kiki’?”)
option_1 / option_2: Image files representing the two choices
option_1_selections / option_2_selections: Number of respondents choosing each option
detailed_results: A list of respondent-level dictionaries Each dictionary includes:
- selection: which option the respondent chose
- country
- language
- age
- gender
- occupation
- userScore: a respondent-level quality score used by Rapidata

Example structure of `detailed_results`

{
  "selection": "option_1",
  "country": "PL",
  "language": "pl",
  "age": "25-34",
  "gender": "Female",
  "occupation": "Graphic Designer",
  "userScore": 0.87
}

Loading the Dataset

Below is an example of how to load all individual responses for a given question into a Pandas DataFrame:

from datasets import load_dataset
import pandas as pd

# Login using e.g. `huggingface-cli login` to access this dataset
ds = load_dataset("Rapidata/psychology-association-kiki-bouba-etc")

df_from_hf = pd.DataFrame(ds['train'])

question_idx = 0
q = pd.DataFrame(df_from_hf["detailed_results"][question_idx])

Data Collection

All responses were collected using the Rapidata API, which enables:

fast global human feedback
targeted or global reach
image, video, audio, and text-based tasks
real-time evaluation at scale

🔗 More about Rapidata: (https://www.rapidata.ai/)

Intended Use

Cognitive science experiments
Cross-cultural analysis
Linguistic influence studies
Human perception & semantic association research
Fun explorations of collective intuition

Limitations

The dataset is primarily for exploration and entertainment
No guarantees of scientific rigor
Responses may be skewed by:
- demographic imbalances
- translation differences
- question phrasing
Some demographics are missing or self-reported

Citation

If you use this dataset, please refer back to this page, and consider leaving a like!

Downloads last month: 31

Total file size:

4.44 MB

🧠 Mental Association Dataset (200k Global Responses)

Dataset Summary