README.md

---
dataset_info:
  features:
  - name: question_key
    dtype: string
  - name: question
    dtype: string
  - name: temporal_offset_days
    dtype: int64
  - name: audience
    dtype: string
  - name: canvas
    dtype: image
  - name: responses_count
    dtype: int64
  - name: detailed_results
    list:
    - name: x
      dtype: float64
    - name: y
      dtype: float64
    - name: country
      dtype: string
    - name: language
      dtype: string
    - name: age
      dtype: string
    - name: gender
      dtype: string
    - name: occupation
      dtype: string
    - name: userScore
      dtype: float64
  splits:
  - name: train
    num_bytes: 5510676.0
    num_examples: 16
  download_size: 1146089
  dataset_size: 5510676.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# 🧭 The Mental Timeline Atlas (84k Taps Across the World)

**Where in space do people place the past and the future?** 84,000 respondents worldwide
saw a blank canvas with a single dot marked *"today"* and were asked to tap where *tomorrow*,
*yesterday*, *10 years from now*, and five other moments in time belong. One tap per person.
This dataset contains every tap, with respondent-level language, country, and demographic metadata.

Companion datasets: [kiki–bouba text](https://huggingface.co/datasets/Rapidata/psychology-association-kiki-bouba-etc)
and [kiki–bouba audio](https://huggingface.co/datasets/Rapidata/psychology-association-kiki-bouba-audio).
Collected with the [**Rapidata API**](https://www.rapidata.ai/) in August 2026.
Please consider leaving a heart if this dataset is useful to you.

---

## Abstract

Psycholinguistics has long held that literate humans carry a *mental timeline* — a spatial
mapping of time whose direction follows the writing system they read (Boroditsky 2001;
Casasanto & Boroditsky 2008). Almost all evidence comes from small laboratory samples.
We elicited single-tap spatial placements of eight temporal expressions from a worldwide
respondent pool, ~10,000 responses per question, plus a *"tap anywhere"* control condition.
Three results stand out. **(1)** The timeline is real, horizontal, and metric: taps for past
expressions cluster left of "today", future expressions right, and median tap distance grows
monotonically — approximately logarithmically — with temporal distance in both directions.
**(2)** Timeline direction tracks the writing system: speakers of left-to-right languages show
strong future-right mappings (UK English viewers: **+57.4** points), while Arabic viewers
— bidirectional readers of a right-to-left script — show **no horizontal timeline at all**
(+2.5 ± 4.4, n=1,799);
an axis *cancellation*, not the reversal a naive script account predicts.
**(3)** The far future drifts *upward*: the share of taps above center climbs monotonically
from past to future (42% → 65% in the
familiarized audience), suggesting the mental time axis is a diagonal, not a flat line.

---

## 1. Method

**Task.** A locate task: respondents see a neutral 1024×1024 canvas with a black dot at the
exact center and an instruction such as *"The dot is today. Tap where tomorrow is."* The
interface accepts exactly **one tap**. No other marks, labels, or hints appear on the canvas.

**Questions.** Eight conditions, ~10,000 responses each (global audience):

| key | instruction | temporal offset |
|---|---|---|
| `ten-years-ago` | The dot is today. Tap where 10 years ago is. | −10 y |
| `childhood` | The dot is today. Tap where your childhood is. | autobiographical past |
| `yesterday` | The dot is today. Tap where yesterday is. | −1 d |
| `control` | Tap anywhere on the image. | — |
| `tomorrow` | The dot is today. Tap where tomorrow is. | +1 d |
| `day-after-tomorrow` | The dot is today. Tap where the day after tomorrow is. | +2 d |
| `next-year` | The dot is today. Tap where next year is. | +1 y |
| `ten-years-from-now` | The dot is today. Tap where 10 years from now is. | +10 y |

**Control condition.** The *"tap anywhere"* question measures where people tap in the absence
of any temporal content. This matters: free taps skew strongly rightward
(73.8% right of center), presumably reflecting handedness and thumb ergonomics on
phones. All directional results below are read against this baseline, and analysts using this
dataset should do the same.

**Audiences.** Every question ran on two respondent pools:

* **`global`** (~10,000/question) — Rapidata's worldwide audience, translated automatically
  into each respondent's language. The Arabic translation was additionally verified by two
  native Arabic speakers.
* **`familiarized`** (~500/question) — respondents who were first made familiar with the
  interface by solving a short series of example tasks asking them to tap in various parts of
  the image ("tap to the left of the dot", "tap as far right of the dot as possible", …)
  before contributing. Two properties of this pool are worth keeping in mind when comparing
  the two audiences: its respondents had already practiced the exact tap interaction on the
  exact canvas, and — being a smaller pool answering eight related questions — they overlap
  far more across questions, so many of them saw several temporal questions and therefore had
  more context about the task family than a typical global respondent, who often saw only one.
  Both factors give their taps sharper spatial structure (see §3.4).

**Respondent metadata.** Each tap carries country, viewing language, age bracket, gender,
occupation (where available), and `userScore` — Rapidata's per-respondent reliability estimate,
derived from performance on known-answer tasks across the platform. It is included so analysts
can weight responses, exactly as the platform itself does when aggregating.

---

## 2. Results

### 2.1 The timeline is real and horizontal

![heatmaps](img/heatmaps.png)

Past questions form a dense horizontal band left of "today"; future questions mirror it on the
right. Taps concentrate on the horizontal midline far above chance. Relative to the control
baseline, *yesterday* pulls -23.4 points leftward and *next year* pushes
+5.1 points rightward (global audience).

### 2.2 Distance in time becomes distance in space

![distance](img/distance_scaling.png)

Median tap distance from "today" grows monotonically with temporal distance, in both
directions, on an approximately logarithmic scale — in the familiarized audience from
21.9 (yesterday) to 44.8
(10 years from now, in canvas-percent units). Notably, *childhood*
(43.9) lands at the same distance as *10 years ago*
(43.4): autobiographical time appears to share the metric of
calendar time.

### 2.3 Direction follows the writing system — and Arabic cancels the axis

![atlas](img/language_atlas.png)

Timeline strength (tomorrow %right − yesterday %right) by viewing language: Japanese
+44.2, Spanish +38.8, English (all viewers)
+24.6 — and Arabic +2.5 ±
4.4: statistically indistinguishable from zero, with n=1,799.
The null persists at the 10-year horizon, within every Arabic-speaking country separately
(Egypt, Iraq, Algeria), in every age bracket, and among the highest-`userScore` respondents —
so it is not explained by sample composition. Arabic speakers scale *distance* normally; they
simply assign no consistent left–right *direction*. Since most Arabic readers also routinely
read left-to-right material (numbers, Latin-script content), the natural interpretation is that
exposure to both directions cancels the horizontal axis rather than reversing it.

The right panel decomposes English — a viewing language spanning many native scripts — by
country: UK +57.4 (the strongest group in the study) down to Pakistan +10.5,
where the dominant native script (Urdu) runs right-to-left. The gradient follows the script of
the respondent's likely native language, reinforcing the writing-system account and
illustrating why viewing language is only a proxy for native language.

### 2.4 The future rises

![rises](img/future_rises.png)

The share of taps landing *above* center climbs monotonically from the deep past to the far
future. The effect survives a geometry check (gains in the upper-right octant are about twice
those in the lower-right, so it is not an artifact of corners simply affording larger
distances). The mental time axis, at least for far time, appears to tilt diagonally upward.

### 2.5 Interface familiarization sharpens every effect

![familiarization](img/familiarization.png)

The familiarized audience — practiced on the interaction and with more cross-question context
(§1) — shows the same qualitative structure as the global audience with substantially sharper
spatial statistics: fewer taps landing on the reference dot itself, a timeline strength of
+59.9 (vs +27.8 global), and more taps on the horizontal axis. Consistent
with this, timeline strength in the global audience increases monotonically across `userScore`
quartiles — the platform's reliability weighting is well-calibrated for this task, and analysts
should regard score-weighted (or familiarized-audience) estimates as the best estimate of the
population effect.

---

## 3. Limitations

* **Viewing language is a proxy** for native language and writing habits (see §2.3's English
  decomposition for how much this can matter).
* **The canvas is finite.** For the ±10-year questions a large share of taps saturate near the
  canvas edge, so far-lag distances are compressed; the childhood ≈ 10-years-ago equality
  should be read with this in mind.
* **The rightward free-tap baseline** (73.8% right) must be subtracted for any
  directional claim; raw percentages overstate future-right and understate past-left effects.
* **Between-subject purity is partial**: a respondent may answer more than one question
  (by design more often so in the familiarized audience).
* One tap records direction and magnitude but not confidence or reaction time.

---

## 4. Dataset structure

One row per question × audience (16 rows):

* `question_key` / `question`: condition id and the instruction shown
* `temporal_offset_days`: signed offset (−3650 … +3650); null for `control` and for the
  autobiographical `childhood` question
* `audience`: `global` or `familiarized`
* `canvas`: the exact image shown (dot center at 50%, 50%)
* `responses_count`, and `detailed_results`: one entry per respondent —
  `x`, `y` (0–100, percent of canvas, y grows downward; the dot is at 50/50),
  `country`, `language`, `age`, `gender`, `occupation`, `userScore`

```python
from datasets import load_dataset
import pandas as pd

ds = load_dataset("Rapidata/mental-timeline-atlas")
df = pd.DataFrame(ds["train"])

row = df[(df.question_key == "tomorrow") & (df.audience == "global")].iloc[0]
taps = pd.DataFrame(row["detailed_results"])
taps["dx"] = taps.x - 50   # >0 = right of "today"
print(taps.groupby("language").dx.apply(lambda s: (s > 0).mean()))
```

---

## 5. Citation

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