A toolkit for understanding factuality & consistency errors in summarization models.
A harness for generating text summaries with automated factuality evaluations
An interactive query interface for exploring generated summaries (i.e. XSum or custom dataset)
An interactive query interface for ngram lookup
Setup (python 3.8):
pip install -r requirements.txt
pip install .
streamlit run interface/app.py
You can also run interfaces individually, i.e.
streamlit run interface/summary_interface.py
Setup (python 3.8):
pip install -r requirements.dev.txt
pip install -Ue .
Before commiting:
black sumtool/ interface/ scripts/
flake8 sumtool/ interface/ scripts/
Create a Github token to access your private repositories. Follow these steps here: Github: Creating a Personal Access Token
Create a new Colab notebook and set the runtime type to GPU
Add the following commands in the first cell to clone the repository and install the requirements
!git clone https://[your-git-token]@github.com/cs6741/summary-analysis.git
!pip install -r /content/summary-analysis/requirements.txt
Add the following command to run the text generation script
!python /content/generate_xsum_summary.py --bbc_ids [idx1,idx2] --data_split [train|test]
Pipeline for storage:
/data/<dataset>/<model-id>-summaries.json
<document_id>:
summary: the generated summary,
metadata: ...metadata for the generated summary, i.e. annotations / score / entropy
/data/<dataset>/<model-id>-metrics.json
<document_id>:
...metrics for a stored summary, i.e. rouge-score, bert-score