psychology-association-kiki-bouba-etc

# 🧠 Mental Association Dataset (200k Global Responses) ## **Dataset Summary** 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](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**](https://www.rapidata.ai/), 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` ```json { "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: ```python 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!