Data from: Ad hoc conventions generalize to new referents
Data files
Sep 16, 2026 version files 9.34 MB
-
README.md
26.51 KB
-
tangrams-ref-main.zip
9.32 MB
Abstract
Every day, humans effortlessly perform a remarkable communicative feat: we talk about things we have never discussed before. What happens in our minds when we agree on what to call something? It remains unclear whether such coordination merely attaches labels to objects, increases accessibility to certain vocabulary, or actively reshapes people's conceptual spaces. We test these competing accounts with a dyadic communication study using tangrams, a diverse set of abstract shapes. Through repeated interaction, participants established names for a set of images; after the interaction, they also described undiscussed images more similarly. Generalization decayed nonlinearly with image similarity (consistent with Shepard's law), and was robust across images of varying nameability. These results demonstrate that conversational interaction aligns conceptual representations. This process reveals how human language is continuously negotiated through social interaction, not fixed in isolation, with direct implications for theories of reference and language evolution.
All data and analysis code are in tangrams-ref-main.zip; this README documents its full contents. Unzipping produces data/, analyses/, and experiments/ as described below.
Two participants play a tangram reference game over abstract shapes. In each trial, the speaker sees a designated target and describes it; the listener reads the description and clicks the image they think is the target. Neither sees the other's screen.
Every game runs three phases:
| Phase | Name | Trials/game | Interaction | Measures |
|---|---|---|---|---|
| 1 | pre | 40 (20 per player, alone) | none | Baseline description per tangram, before partner contact |
| 2 | interactive | 30 (5 blocks × 6) | free bidirectional chat, roles swap | Convention formation |
| 3 | post | 20 | one speaker utterance, no reply | Whether conventions transfer to new referents |
A phase-3 target was either discussed in phase 2 (repeated) or is a new tangram from the same visual thread (control) — threads being families of visually similar shapes. Main experiment: 302 participants, 151 games, 13,590 trials. Two earlier pilots (60 and 59 games) are also included.
Files
data/ full_data.csv, demographics.csv, thread_contexts.csv
analyses/ analysis.qmd, pilot1.qmd, pilot2.qmd
data/ processed tables consumed by the notebooks
experiments/ web experiment source (Empirica/Meteor); see experiments/README.md
Conventions
Tangrams are named by their KiloGram file name, e.g. page7-40.svg; some columns drop the .svg. The 1,016 SVGs are in experiments/experiment/public/tangrams/.
Identifiers are arbitrary labels carrying no participant information: games G### (pilots P1G###/P2G###), rounds R##### (P1R#####/P2R#####), main-experiment participants integers 0–301, pilot participants P1-### / P2-###. The main experiment and the two pilots each recruited a distinct set of people, so the three participant namespaces never overlap and must not be joined to one another; demographics.csv covers the main experiment only.
Missing data is always an empty cell (NA in R, NaN in pandas) and is structural; the reason is noted per variable below. No other missing-data codes.
List-valued cells hold a Python list literal as text, e.g. "['page7-40.svg', 'page8-8.svg']"; parse with ast.literal_eval.
Similarity columns are cosine similarities unless stated. jaccard_index is word overlap (0–1). snd_value is Shape Naming Divergence (0–1, higher = harder to name), from KiloGram.
Several processed files carry an unnamed leading column (Unnamed: 0 or a blank header) — a pandas row index; ignore it.
data/full_data.csv
Main experiment: one row per trial per speaker. 13,590 rows = 151 games × 90 trials (40 + 30 + 20).
| Variable | Type / units | Description |
|---|---|---|
game_id |
G### |
Game (dyad). 151 unique. |
configuration_file |
string | JSON defining the game's trial sequence; see experiments/experiment/public/games/. |
phase |
1–3 | 1 = pre, 2 = interactive, 3 = post. |
round_id |
R##### |
Round. 7,550 unique — a phase-1 round is shared by both players. |
trial_index |
integer 0–51 | Position within the game. |
block |
0–4, else empty | Repetition block in phase 2. Empty in phase 1; always 0.0 in phase 3 (no block structure). |
target |
string | Tangram the speaker had to describe. |
snd_value |
0–1 | Shape Naming Divergence of the target. |
snd_class |
low/high |
low if snd_value <= 0.85. |
chat_mode |
string, else empty | Empty in phase 1; switch-bidirectional in phase 2 (both may write, roles alternate); single-utterance-unidirectional in phase 3 (one message, no reply). |
speaker_id |
0–301 | Speaker; joins to demographics.csv:participant_id. |
listener_id |
0–301, else empty | Listener; empty in phase 1, which has none. |
context_id |
integer | The 10-tangram context; joins to thread_contexts.csv. 150 of 155 contexts occur. |
context |
list of strings | The 10 tangrams displayed, including the target. |
description |
string or list | Speaker's text. Plain string in phases 1 and 3; in phase 2 a list holding the trial's full chat in order. |
time_to_message |
seconds, or list | Trial onset → message sent. A list in phase 2, one entry per message. |
sec_until_click |
seconds, else empty | Message → listener click. Empty in phase 1 and on timed-out trials. |
response |
string, else empty | Tangram clicked. Empty in phase 1 and on timed-out trials (75 in phase 2, 437 in phase 3). |
correct |
True/False, else empty |
response == target. Empty wherever response is empty. |
data/demographics.csv
The 302 main-experiment participants, one row each; reduced and generalized for public release (see Human subjects data). Pilot participants are not included.
| Variable | Type | Description |
|---|---|---|
participant_id |
0–301 | Arbitrary number, not derived from personal information. Joins to speaker_id/listener_id in full_data.csv and player_id in shepard_pairs*.csv. |
age_range |
categorical | 18-24, 25-34, 35-44, 45-54, 55+, or NA. Banded and top-coded, not exact age. |
sex |
Male/Female/NA |
NA covers both "prefer not to say" and unavailable values. |
country_of_residence |
United Kingdom/United States |
Residence during the study. |
Counts: 164 UK / 138 US; 163 male, 138 female, 1 NA; ages 18-24 (37), 25-34 (95), 35-44 (79),
45-54 (46), 55+ (43), NA (2). All participants were English speakers.
data/thread_contexts.csv
Stimulus design. 1,550 rows = 155 contexts × 10 tangrams.
| Variable | Type | Description |
|---|---|---|
context_id |
0–154 | Joins to full_data.csv:context_id. |
thread_id |
0–9 | The visual-similarity thread the context is drawn from. |
pre_post_A |
string | First member of this slot's pre/post comparison pair. |
pre_post_B |
string | Second member of that pair. |
interactive |
string | Tangram used in the interactive phase for this slot. |
Read the three tangram columns across a context's 10 rows to reconstruct which shapes were paired for the pre/post measure and which were discussed interactively.
analyses/data/ — processed tables
Derived from data/full_data.csv and used directly by the notebooks. Supplied ready-made so the analyses re-run without recomputing embeddings.
clean_no_discourse.csv — interactive trials, cleaned
4,449 rows. Phase-2 trials, minus those whose chat was purely discourse management.
| Variable | Type / units | Description |
|---|---|---|
game_id |
G### |
Game. |
block |
0–4 | Repetition block. |
target, response |
string | Target tangram; tangram clicked. |
correct |
True/False, else empty |
Empty on 4 timed-out trials. |
controlled |
True/False |
True = control target (not repeated across phases), False = repeated. analysis.qmd recodes this to condition (control/repeated). |
snd_value, snd_class |
0–1; low/high |
SND of the target, and its binarization. |
sec_until_click |
seconds | Message → listener click. |
speaker_desc |
list of strings | All speaker messages this trial. |
listener_desc |
list of strings | All listener messages (empty list if none). |
speaker_len, listener_len, total_len |
words | Word counts of the above, and their sum. |
chat_len |
messages | Messages exchanged this trial. |
redacted_speaker_len, redacted_listener_len, redacted_total_len, redacted_chat_len |
words; messages | The four counts above, recomputed on the redacted text (see Human subjects data). Equal to the originals except on trials whose text changed. |
pre_post.csv — pre/post description similarity
2,948 rows = 1,474 tangram pairs × 2 phases. The central test of generalization.
| Variable | Type / units | Description |
|---|---|---|
phase |
pre/post |
Descriptions from phase 1 or phase 3. |
pair-id |
integer | Tangram pair; appears once as pre, once as post. |
game |
G### |
Game. |
tangrams |
list of 2 strings | The pair, with .svg. |
tangram_A, tangram_B |
string | The two tangrams, without extension. |
p1_description, p2_description |
string | Descriptions given for tangram_A and tangram_B this phase. |
similarity |
cosine | Text-embedding similarity of the two descriptions. Main dependent measure. |
is_control |
True/False |
True = control (non-repeated) pair. |
tangram1_snd, tangram2_snd |
0–1 | SND of each pair member. |
prepost_avg_snd |
0–1 | Mean SND of the pair. |
interactive_phase_tangram |
string | Same-thread tangram actually discussed in phase 2. |
interactive_phase_snd |
0–1 | Its SND. |
thread_avg_snd |
0–1 | Mean SND across the thread. |
thread_id |
0–9 | Visual-similarity thread. |
bin |
0–9 | Design bin the context was sampled into (equals thread_id here). |
context_id |
integer | Joins to thread_contexts.csv. |
sim_A_interactive, sim_B_interactive |
cosine | Image similarity of tangram_A / tangram_B to the interactive-phase tangram. |
avg_sim_AB_interactive |
cosine | Mean of the two. |
jaccard_index |
0–1 | Word overlap between the two descriptions. |
shepard_pairs.csv — phase-2 vs. phase-3 pairs
2,988 rows. Relates description similarity to visual similarity (a Shepard-style generalization gradient).
| Variable | Type / units | Description |
|---|---|---|
player_id |
0–301 | Participant; 301 of 302 appear. |
is_same_thread |
True/False |
Whether the two tangrams share a thread. Balanced, 1,494 each. |
phase2_image, phase2_text |
string | Tangram described in phase 2, and its description. |
phase3_image, phase3_text |
string | Tangram described in phase 3, and its description. |
phase2_thread, phase3_thread |
list of strings | The thread each image belongs to. |
image_sim |
cosine | Visual-embedding similarity of the two images. |
text_sim |
cosine | Text-embedding similarity of the two descriptions. |
shepard_pairs_pixsim.csv
Same 2,988 rows and columns as shepard_pairs.csv (player_id uses the same numbering), plus three model-free image measures for comparison against image_sim:
| Variable | Type / units | Description |
|---|---|---|
jaccard_index |
0–1 | Word overlap between phase2_text and phase3_text. |
pixel_sim |
0–1 | Pixel-overlap similarity of the two rendered tangrams. |
pixel_corr |
−1–1 | Pearson correlation of the two images' pixel values. |
block_sim_ada_jaccard.csv — convergence across blocks
596 rows = 149 games × 4 consecutive block pairs.
| Variable | Type / units | Description |
|---|---|---|
game_id |
G### |
Game. |
blocks |
(n,n+1) |
Blocks compared: (0,1), (1,2), (2,3), (3,4). |
ada_cos_sim |
cosine | Text-embedding similarity between the two blocks' descriptions. |
jaccard_index |
0–1 | Word overlap between them. |
voting_text_embed_similarity.csv — human validation of text similarity
1,474 rows. Annotators judged which of two description pairs was more similar; compared against three embedding models.
| Variable | Type / units | Description |
|---|---|---|
key |
string | <game_id>-<tangram_A>-<tangram_B>, e.g. G133-page2-21-page7-10. Tangram names themselves contain hyphens. |
tangram_A, tangram_B |
string | The pair judged. |
ori_pre, ori_post |
string | The two phase-1 / phase-3 descriptions, joined by |. |
controlled_thread |
True/False |
Whether this is a control thread. |
ada_pre_sim, ada_post_sim |
cosine | Similarity from OpenAI ada text embeddings. |
clip_pre_sim, clip_post_sim |
cosine | Same, CLIP text embeddings. |
sbert_pre_sim, sbert_post_sim |
cosine | Same, Sentence-BERT. |
pre_sim, post_sim |
cosine | Values used downstream (ada). |
gt, eq, lt |
vote counts | Annotators judging the post pair more / equally / less similar than the pre pair. Blank for 2 items with no votes. |
human_judgement |
−1/0/1, else empty | Majority verdict: 1 = post more similar, 0 = equal, -1 = less. Empty for 27 items with no majority. Modeled as an ordered factor. |
ada_diff, clip_diff, sbert_diff |
numeric | post_sim − pre_sim per model — the model's analogue of human_judgement. |
voting_image_embed_similarity.csv — human validation of image similarity
1,494 rows. Annotators compared two image pairs on visual similarity.
| Variable | Type / units | Description |
|---|---|---|
key |
string, 20 chars | Arbitrary annotation-item key. Not linked to any participant. |
pair0_phase2_image, pair0_phase3_image |
string | The first image pair shown. |
pair1_phase2_image, pair1_phase3_image |
string | The second image pair shown. |
pair0_image_sim, pair1_image_sim |
cosine | Image-embedding similarity of each pair. |
gt, eq, lt |
vote counts | Votes that pair 0 was more / equally / less similar than pair 1. |
human_judgement |
−1/0/1, else empty | Majority verdict. Empty for 36 items with no majority. |
sim_diff |
numeric | pair0_image_sim − pair1_image_sim. |
pilot1_data.csv, pilot2_data.csv
1,800 rows (60 games) and 2,360 rows (59 games). Single-phase pilots, both
single-utterance-unidirectional only, each run with its own distinct set of participants.
Columns match full_data.csv except:
| Variable | Type / units | Description |
|---|---|---|
game_id, round_id |
P1G###/P2G###, P1R#####/P2R##### |
Pilot-specific identifiers. |
speaker_id, listener_id |
P1-### / P2-### |
Pilot participants (120 in pilot 1, 118 in pilot 2), numbered separately per pilot. No pilot participant took part in the other pilot or in the main experiment. |
sec_until_press |
seconds | Time until the speaker began typing. |
sec_until_click |
(n,) |
Seconds until the listener clicked, as a 1-tuple literal — strip parentheses and trailing comma. |
description_len |
words | Word count of description. |
redacted_description_len |
words | description_len recomputed on the redacted text (see Human subjects data). |
controlled |
True/False |
Control vs. repeated target. |
Software and workflow
Analyses are Quarto notebooks in R. Render with quarto render analyses/analysis.qmd or run the chunks in RStudio; paths resolve via the here package, so run from the repository root.
| Notebook | Reads | Covers |
|---|---|---|
analysis.qmd |
clean_no_discourse, pre_post, shepard_pairs, shepard_pairs_pixsim, block_sim_ada_jaccard, voting_text_embed_similarity |
All main results: convention formation (accuracy, description length, response time, utterance similarity), pre/post generalization, similarity-gradient models, human-vs-embedding validation |
pilot1.qmd, pilot2.qmd |
pilot1_data, pilot2_data |
The two pilots |
R packages: tidyverse, here, ggthemes, tidyboot, lme4, lmerTest, brms. brms needs a Stan toolchain and is the slowest step. Embedding-derived columns (ada_*, clip_*, sbert_*, image_sim, similarity) were computed ahead of time, so re-running needs no API access or GPU.
The experiment is a Meteor app on the Empirica framework —
experiments/experiment/ (main) and experiments/pilot/, each with client/, server/, and public/. Setup and deployment are in experiments/README.md. Under public/: tangrams/ holds the 1,016 stimulus SVGs; games/ holds 310 JSON files (ref2_pilot5_binned_threads_<context_id>_<phase>.json, matching full_data.csv:configuration_file) that predefine each trial's images, target, block, per-player ordering, roles, and chat mode; games/old_games/ holds superseded configurations not used by the released data; packages/meteor-empirica-core/ is a vendored third-party copy of Empirica, included so the app builds reproducibly.
Tangram stimuli and the snd_value norms come from the KiloGram dataset.
Human subjects data
Participants were adults recruited online. All consented to take part and to public release of their de-identified data. No minors took part (minimum age 18). This release contains no direct identifiers. The de-identification procedure was:
- Platform identifiers removed. Recruitment-platform accounts were never carried into the release; each participant has an arbitrary label not derived from personal information and not reversible (
0–301for the main experiment,P1-###/P2-###for the pilots, numbered separately because the three studies recruited distinct people). - Internal database identifiers replaced. The Empirica/Meteor server's record ids for every game, round, and player were replaced with arbitrary labels (
G001,R00001, …), so nothing traces back to the original experiment server. - Indirect identifiers limited to three —
age_range,sex,country_of_residence. Ethnicity, country of birth, nationality, language, student status, and employment status were removed. - Age generalized into bands, top-coded at
55+so no older participant is isolated. - Free text screened for names, contact details, locations, ages, and self-descriptions. One message where a participant volunteered their age was found; it is not part of this release.
- Free text redacted for public release. Utterances carrying crude or sexual content, or that were wholly conversational rather than descriptive — greetings and sign-offs, pleasantries, task and platform commentary, demographic probes, requests for a description, and filler — were removed; where such a remark shared an utterance with a real shape description, only the remark was cut and the description left as typed. One direct identifier, a personal name, is marked
[redacted].REDACTION_LOG.mdgives the counts by category andredaction_changelog.tsvthe cell-by-cell audit trail.
Retained free-text fields (description, speaker_desc, listener_desc, phase2_text, p1_description, and similar) are short referring expressions for abstract shapes — "weird 6", "an unlocked padlock". They are the dataset's primary scientific content and, having been screened as above, carry no information about participants.
Access to the full dataset. What is published here is the de-identified public release. The full dataset may be available to qualified researchers upon request to the corresponding author, subject to appropriate review and any applicable data-use requirements.