VisText: A Benchmark for Semantically Rich Chart Captioning

VisText is a benchmark dataset of over 12,000 charts and semantically rich captions! In the VisText dataset, each chart is represented as its rasterized image, scene graph specification, and underlying datatable. Each chart is paired with a synthetic L1 caption that describes the chart's elemental and ecoded properties and a human-generated L2/L3 caption that describes trends and statistics about the chart.

In the VisText paper, we train text-based models (i.e, models that use the scene graph and data table chart representations) as well as image-guided models that include the chart image. We also include semantic prefix-tuning, allowing our models to customize the level of semantic content in the chat. Our models output verbose chart captions that contain varying levels of semantic content.

This repository contains code for training and evaluating the VisText models. For more info, see: VisText: A Benchmark for Semantically Rich Chart Captioning (ACL 2023)

Repository Structure

run.sh # Main script used to train and evaluate models

./code # Dataset generation, training, and evaluation code
    image_guided/ # image-guided models use chart images
    text_only/ # text_only models use chart scene graphs and data tables
    ...
    
./data # Stores the VisText dataset
    feature_extraction/ # Code to extract chart visual features
    ...

./metrics # Default directory prediction metrics will be written to
    
./models # Stores model outputs and pretrained model checkpoints

Set Up

Clone the VisText repo & set up environment

Clone this repo locally using git clone https://github.com/mitvis/vistext.git. To set up the conda environment with all dependencies, run conda env create -f environment.yml (for evaluation and text-only training) and conda env create -f environment-image_guided.yml (for image-guided models). This will create the conda environments with the correct package versions, which you can activate with conda activate vistext and conda activate vistext-image_guided respectively.

Download the raw data

Call download_data.sh to download and unzip the raw data (and optional image-guided features and weights). By default, the tabular data files are downloaded (containing splits for the reduced scenegraphs, linearized data tables, and captions). Options are:

--images                 # Download rasterized image files
--scenegraphs            # Download original/hierarchical scene graph files
--vl_spec                # Download Vega-Lite specs
--image_guided           # Download visual features and the weights for VLT5/VLBart necessary for fine-tuning

To only download the splits for text-based captioning, run:

bash download_data.sh

To download all of the raw data run:

bash download_data.sh --images --scenegraphs

Generate the VisText dataset from raw data

Run the code/dataset_generation.ipynb notebook from start to finish, which will generate the three split dataset files data/data_train.json, data/data_test.json, and data/data_validation.json.

Usage

Model training, evaluation, and predictions

Call run.sh with your desired parameters. Options are:

-c model_class           # Class of models to use ('text_only' or 'image_guided').
-b batch_size            # Batch size for training, validation, and testing.
-e num_epochs            # Number of epochs to train for.
-g num_gpus              # Number of gpus to parallelize across.
-i input_type            # Chart text representation ('scenegraph', 'datatable', or 'imageonly').
-m model_backbone        # Model architecture to finetune ('byt5', 't5', or 'bart').
-s seed                  # Seed number.
--prefix_tuning          # Flag to use sematic prefix-tuning.

For example, to train and evaluate the scene-graph model with prefix-tuning, run:

bash run.sh -c text_only -b 4 -e 50 -g 4 -i scenegraph -m byt5 -s 10 --prefix_tuning

If you are training image guided models, be sure to use the correct conda environment (vistext-image_guided).

Evaluating on test set

Call run_test.sh with your parameters. Options are:

-p predictions_path      # Path of predictions
-f results_filename      # Filename to save results under in metrics/
--split_eval             # Flag to use for evaluating L1 and L2/L3 captions separately; only works with prefix-tuning models.
--prefix_tuning          # Flag to use sematic prefix-tuning.

For example, to run evaluation on the above trained model, run:

bash run_test.sh -p models/vistext_scenegraph_byt5_prefixtuningtrue_seed10/generated_predictions.txt -f vistext_scenegraph_byt5_prefixtuningtrue_seed10.txt --split_eval --prefix_tuning

Mapping from Statista Dataset

VisText uses the Statista data scraped by Chart-to-Text: A Large-Scale Benchmark for Chart Summarization. We re-map the original IDs using the mappings found in data/statista_mappings.json.

Language Prompts

The models we trained on VisText, including the open-source pretrained VisText models, use the prompt structure: translate chart to {semantic level}: . For each training variation the prompt is:

• Full caption generation: translate chart to L1L2L3:
• Prefix-tuning L1 generation: translate chart to L1:
• Prefix-tuning L2/L3 generation: translate chart to L2L3:

Citation

For more information about VisText, check out VisText: A Benchmark for Semantically Rich Chart Captioning

@inproceedings{2023-vistext,
  title = {{VisText: A Benchmark for Semantically Rich Chart Captioning}},
  author = {Benny J. Tang AND Angie Boggust AND Arvind Satyanarayan},
  booktitle = {The Annual Meeting of the Association for Computational Linguistics (ACL)},
  year = {2023},
  url = {http://vis.csail.mit.edu/pubs/vistext}
}