How do people agree on ways of communicating visual concepts?

Note: This repo reflects the state of the codebase used for the journal article. To access the earlier version of the codebase as of CogSci 2019, please refer to: https://github.com/cogtoolslab/graphical_conventions_cogsci19.

This code has been tested on MacOS version 11.x.x (Big Sur) and Ubuntu 18.x.x. We do not recommend using Windows.

The easiest way to reproduce our Python code is to install the miniconda package manager and create a new environment from our specification:

conda create --name graphical_conventions --file requirements.txt

All data needed to reproduce our analyses are already included in this repo, so no additional steps are required to fetch data files once you clone this repo. However, for additional transparency we have additionally included the script that we use internally to fetch data files and organize them into the appropriate subdirectories expected by our analysis code (/data/download_data.py).

• experiment contains behavioral data from our original sample (run3run4) as well as our internal replication (run5_submitButton) of both the communication task and the recognition task. group_data and recog_data_raw mark the raw communication data and recognition data directly from our database, while bis marks the cleaned and pre-processed versions.
• diagnosticity contains annotations from our sketch mapping task and diagnosticity mapping task as well as pre-processed maps and masks for convenience.
• features contains pre-extracted CNN features for all sketches. To re-extract features (on a GPU machine with ), please follow the instructions in /analysis/preprocessing/.
• sketches contains the sketches from all experimental sessions, which we include in both PNG (bitmap) and SVG (vector graphics) formats. The PNG file format may be more appropriate for analysis pipelines that treat sketches as pixel-based inputs, while the SVG file format may be more appropriate for analyses that treat sketches as graphics programs (consisting of a sequence of commands to generate each pen stroke).

If you run into difficulties using our download script, feel free to click this link to directly download a zipped folder (approx. 214.3 MB) containing the full dataset. Once you extract the contents of this zipped file, please move its contents into the data/ directory in this repository before running the provided analysis code.

We ran four different experiments. To demo or replicate these experiments, please see the README inside the ./experiments subdirectory for installation instructions.

• initial repeated communication task (/experiments/reference-game/)
• sketch recognition task (/experiments/recognition/)
• sketch-mapping task linking strokes to regions of target object (/experiments/sketch-mapping/)
• diagnosticity mapping task (/experiments/object-diagnosticity/)

• Reproduce all model-based analyses of sketch image features by running XXX.
• Reproduce all statistical analyses by running /analysis/analysis.Rmd notebook.