SummVis is an open-source visualization tool that supports fine-grained analysis of summarization models, data, and evaluation metrics. Through its lexical and semantic visualizations, SummVis enables in-depth exploration across important dimensions such as factual consistency and abstractiveness.

Authors: Jesse Vig¹, Wojciech Kryściński¹, Karan Goel², Nazneen Fatema Rajani^1
1Salesforce Research ²Stanford Hazy Research

📖 Paper 🎥 Demo

Note: SummVis is under active development, so expect continued updates in the coming weeks and months. Feel free to raise issues for questions, suggestions, requests or bug reports.

• User guide
• Installation
• Quickstart
• Running with pre-loaded datasets
• Get your data into SummVis
• Deploying SummVis remotely
• Citation
• Acknowledgements

SummVis is a tool for analyzing abstractive summarization systems. It provides fine-grained insights on summarization models, data, and evaluation metrics by visualizing the relationships between source documents, reference summaries, and generated summaries, as illustrated in the figure below.

The SummVis interface is shown below. The example displayed is the first record from the CNN / Daily Mail validation set.

(a) Configuration panel
(b) Source document (or reference summary, depending on configuration)
(c) Generated summaries (and/or reference summary, depending on configuration)
(d) Scroll bar with global view of annotations

N-gram overlap: Word sequences that overlap between the document on the left and the selected summary on the right. Underlines are color-coded by index of summary sentence.
Semantic overlap: Words in the summary that are semantically close to one or more words in document on the left.
Novel words: Words in the summary that do not appear in the document on the left.
Novel entities: Entity words in the summary that do not appear in the document on the left.

Currently only English text is supported.

The following will create a conda environment named summvis

git clone https://github.com/robustness-gym/summvis.git
cd summvis
conda env create -f environment.yml
conda activate summvis # OR "source activate summvis" on older conda versions

Installation takes around 2 minutes on a Macbook Pro.

IMPORTANT: Please use python>=3.8 since some dependencies require that for installation.

# Requires python>=3.8
git clone https://github.com/robustness-gym/summvis.git
cd summvis
pip install --upgrade pip
pip install -r requirements.txt

Start using SummVis immediately on a small sample from the CNN / Daily Mail validation set.

conda activate summvis # OR "source activate summvis" on older conda versions

This may take several minutes if you haven't previously loaded CNN / DailyMail from the Datasets library.

sh quickstart.sh

This will create the file data/cnn_dailymail_10.validation, which contains the first 10 pre-cached examples from the CNN / Daily Mail validation set. For more details on the loading process, see Running with pre-loaded datasets.

streamlit run summvis.py

The tool will automatically load the dataset created in the previous step. If you have other files in the data directory, you will need to select the appropriate file from the dropdown in the upper-right corner of the interface.

Download our pre-cached dataset that contains predictions for state-of-the-art models such as PEGASUS and BART on CNN / Daily Mail and XSum validation sets.

We recommend that you choose the smallest dataset that fits your need in order to minimize download / preprocessing time.

mkdir data
mkdir preprocessing
curl https://storage.googleapis.com/sfr-summvis-data-research/cnn_dailymail_1000.validation.anonymized.zip --output preprocessing/cnn_dailymail_1000.validation.anonymized.zip
unzip preprocessing/cnn_dailymail_1000.validation.anonymized.zip -d preprocessing/

mkdir data
mkdir preprocessing
curl https://storage.googleapis.com/sfr-summvis-data-research/xsum_1000.validation.anonymized.zip --output preprocessing/xsum_1000.validation.anonymized.zip
unzip preprocessing/xsum_1000.validation.anonymized.zip -d preprocessing/

Next, we'll need to add the original examples from the CNN / Daily Mail or XSum datasets to deanonymize the data (this information is omitted for copyright reasons). The preprocessing.py script can be used for this with the --deanonymize flag.

Set the --n_samples argument and name the --processed_dataset_path output file accordingly.

python preprocessing.py \
--deanonymize \
--dataset_rg preprocessing/cnn_dailymail_1000.validation.anonymized \
--dataset cnn_dailymail \
--version 3.0.0 \
--split validation \
--processed_dataset_path data/cnn_dailymail_100.validation \
--n_samples 100

python preprocessing.py \
--deanonymize \
--dataset_rg preprocessing/cnn_dailymail_1000.validation.anonymized \
--dataset cnn_dailymail \
--version 3.0.0 \
--split validation \
--processed_dataset_path data/cnn_dailymail_1000.validation \
--n_samples 1000

python preprocessing.py \
--deanonymize \
--dataset_rg preprocessing/cnn_dailymail.validation.anonymized \
--dataset cnn_dailymail \
--version 3.0.0 \
--split validation \
--processed_dataset_path data/cnn_dailymail.validation

python preprocessing.py \
--deanonymize \
--dataset_rg preprocessing/xsum_1000.validation.anonymized \
--dataset xsum \
--split validation \
--processed_dataset_path data/xsum_1000.validation \
--n_samples 1000

Once the app loads, make sure it's pointing to the right File at the top of the interface.

streamlit run summvis.py

Alternatively, you may point SummVis to a folder where your data is stored:

streamlit run summvis.py -- --path your/path/to/data

You may further specify the filename within the folder:

streamlit run summvis.py -- --path your/path/to/data --file filename

Note that the additional -- is not a mistake, and is required to pass command-line arguments in streamlit.

The simplest way to use SummVis with your own data is to create a jsonl file of the following format:

{"document":  "This is the first source document", "summary:reference": "This is the reference summary", "summary:testmodel1": "This is the summary for testmodel1", "summary:testmodel2": "This is the summary for testmodel2"}
{"document":  "This is the second source document", "summary:reference": "This is the reference summary", "summary:testmodel1": "This is the summary for testmodel1", "summary:testmodel2": "This is the summary for testmodel2"}

The key for the reference summary must equal summary:reference and the key for any other summary must be of the form summary:<summary_name>, e.g. summary:BART. The document and at least one summary (reference, other, or both) are required.

NOTE: SummVis is optimized for documents that are approximately the length of a news article. Documents that are significantly longer than this may render slower in the tool, especially if many summaries are displayed.

The following additional install step is required.:

python -m spacy download en_core_web_lg

You have two options to load this jsonl file into the tool:

This is the simplest approach for loading your data into SummVis. The disadvantage of this approach is that all computations are performed in real-time, which is particularly expensive for the semantic similarity computation. As a result, each example will be slow to load (~5-15 seconds on a Macbook Pro).

1. Place the jsonl file in the data directory. Note that the file must be named with a .jsonl extension.
2. Start SummVis: streamlit run summvis.py
3. Select your jsonl file from the File dropdown at the top of the interface.

Note: To avoid additional text cleaning that may remove newlines, etc., use streamlit run summvis.py -- --no_clean

You may run preprocessing.py to precompute all data required in the interface (running spaCy, lexical and semantic aligners) and save a cache file, which can be read directly into the tool. Note that this script may run for a while (~5-15 seconds per example on a MacBook Pro for documents of typical length found in CNN/DailyMail or XSum), and will be greatly expedited by running on a GPU.

1. Run preprocessing script to generate cache file

   python preprocessing.py \
   --workflow \
   --dataset_jsonl path/to/my_dataset.jsonl \
   --processed_dataset_path path/to/my_cache_file

   Additional options:
     --n_samples <number_of_samples>: Process the first number_of_samples samples only (recommended).
     --no_clean: Do not perform additional text cleaning that may remove newlines, etc.

2. Copy output cache file to the data directory

3. Start SummVis: streamlit run summvis.py

4. Select your file from the File dropdown at the top of the interface.

As an alternative to steps 2-4, you may point SummVis directly to the cache file:

streamlit run summvis.py -- --path <parent_directory_of_cache_file> --file <cache_filename>

The instructions in the previous section assume access to model predictions. We also provide tools to load predictions, either by downloading datasets with precomputed predictions or running a script to generate predictions for HuggingFace-compatible models. In this section we describe an end-to-end pipeline for using these tools.

Prior to running the following, an additional install step is required:

python -m spacy download en_core_web_lg

Loads in a dataset from HF, or any dataset that you have and stores it in a standardized format with columns for document and summary:reference.

python preprocessing.py \
--standardize \
--dataset cnn_dailymail \
--version 3.0.0 \
--split validation \
--save_jsonl_path preprocessing/cnn_dailymail.validation.jsonl

python preprocessing.py \
--standardize \
--dataset_jsonl path/to/my_dataset.jsonl \
--save_jsonl_path preprocessing/my_dataset.jsonl

Expected format of my_dataset.jsonl:

{"document":  "This is the first source document", "summary:reference": "This is the reference summary"}
{"document":  "This is the second source document", "summary:reference": "This is the reference summary"}

If you wish to use column names other than document and summary:reference, you may specify custom column names using the doc_column and reference_column command-line arguments.

Takes a saved dataset that has already been standardized and adds predictions to it from prediction jsonl files. Cached predictions for several models available here: https://storage.googleapis.com/sfr-summvis-data-research/predictions.zip

You may also generate your own predictions using this this script.

python preprocessing.py \
--join_predictions \
--dataset_jsonl preprocessing/cnn_dailymail.validation.jsonl \
--prediction_jsonls \
predictions/bart-cnndm.cnndm.validation.results.anonymized \
predictions/bart-xsum.cnndm.validation.results.anonymized \
predictions/pegasus-cnndm.cnndm.validation.results.anonymized \
predictions/pegasus-multinews.cnndm.validation.results.anonymized \
predictions/pegasus-newsroom.cnndm.validation.results.anonymized \
predictions/pegasus-xsum.cnndm.validation.results.anonymized \
--save_jsonl_path preprocessing/cnn_dailymail.validation.jsonl

Takes a saved dataset that has been standardized, and predictions already added. Applies all the preprocessing steps to it (running spaCy, lexical and semantic aligners), and stores the processed dataset back to disk.

python preprocessing.py \
--workflow \
--dataset_jsonl preprocessing/cnn_dailymail.validation.jsonl \
--processed_dataset_path data/cnn_dailymail.validation \
--try_it

python preprocessing.py \
--workflow \
--dataset_jsonl preprocessing/cnn_dailymail.validation.jsonl \
--processed_dataset_path data/cnn_dailymail

See these tutorials on deploying a Streamlit app to various cloud services (from Streamlit docs):

• How to Deploy Streamlit to a Free Amazon EC2 instance, by Rahul Agarwal
• Host Streamlit on Heroku, by Maarten Grootendorst
• Host Streamlit on Azure, by Richard Peterson
• Host Streamlit on 21YunBox, by Toby Lei

When referencing this repository, please cite this paper:

@misc{vig2021summvis,
      title={SummVis: Interactive Visual Analysis of Models, Data, and Evaluation for Text Summarization}, 
      author={Jesse Vig and Wojciech Kry{\'s}ci{\'n}ski and Karan Goel and Nazneen Fatema Rajani},
      year={2021},
      eprint={2104.07605},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2104.07605}
}

We thank Michael Correll for his valuable feedback.