explainaboard 0.8.8

Explainable Leaderboards for Natural Language Processing

ExplainaBoard: An Explainable Leaderboard for NLP

Introduction | Web Tool | API Tool | Download | Paper | Video | Bib

What is ExplainaBoard?

When developing a natural language processing (NLP or AI) system, often one of the hardest things is to understand where your system is working and where it is failing, and deciding what to do next. ExplainaBoard is a tool that inspects your system outputs, identifies what is working and what is not working, and helps inspire you with ideas of where to go next.

It offers a number of different ways with which you can evaluate and understand your data:

1. Single-system Analysis: What is a system good or bad at?
2. Pairwise Analysis: Where is one system better (worse) than another?
3. Data Bias Analysis: What are the characteristics of different evaluated datasets?
4. Common Errors: What are common mistakes that top-5 systems made?
5. Fine-grained Error Analysis: where do errors occur?
6. System Combination: Is there potential complementarity between different systems?

How do you get started?

1. Browse Outputs Online: We provide an online interface where you can browse and understand outputs from various state-of-the-art systems.
2. Upload Your Own Outputs: You can also upload your own system outputs through the online interace, and we will process them so you can view them alongside other systems.
3. Run Your Own Analyses Offline: You can follow the installation instructions below to run analyses on your own system data offline. We also plan to support command-line upload to the web interface soon!

Web-based Toolkit

We have deployed ExplainaBoard as a Web toolkit, which includes 9 NLP tasks and will be more, 40 datasets and 300 systems. Detailed information is as follows.

Installation

Method 1 - Standard Use: Simple installation from PyPI (Python 3 only)

pip install --upgrade pip  # recommending the newest version of pip.
pip install explainaboard
python -m spacy download en_core_web_sm  # if you plan to use the TextClassificationProcessor

Method 2 - Development: Install from the source and develop locally (Python 3 only)

# Clone current repo
git clone https://github.com/neulab/ExplainaBoard.git
cd ExplainaBoard

# Install the required dependencies and dev dependencies
pip install ."[dev]"
pre-commit install

• Testing: To run tests, you can run python -m unittest.
• Linting and Code Style: This project uses flake8 (linter) and black (formatter). They are enforced in the pre-commit hook and in the CI pipeline.
  • run python -m black . to format code
  • run flake8 to lint code
  • You can also configure your IDE to automatically format and lint the files as you are writing code.
  • There are various ways to skip formatting or linting if you find it necessary. The following are some examples (please refer to official documentation for more options)
    • # noqa signals flake8 to ignore the current line. (flake8: noqa ignores the current file.)
    • lines between # fmt: on and # fmt: off are skipped by black.

Then, you can run following examples via bash

Example for CLI

• text-classification:

explainaboard --task text-classification --system_outputs ./data/system_outputs/sst2/sst2-lstm.tsv

See detailed description and more supported tasks.

Example for Python SDK

from explainaboard import TaskType, get_loader, get_processor

path_data = "./explainaboard/tests/artifacts/test-summ.tsv"
loader = get_loader(TaskType.summarization, data=path_data)
data = list(loader.load())
processor = get_processor(TaskType.summarization)
analysis = processor.process(metadata={}, sys_output=data)
analysis.write_to_directory("./")

Acknowledgement

We thanks all authors who share their system outputs with us: Ikuya Yamada, Stefan Schweter, Colin Raffel, Yang Liu, Li Dong. We also thank Vijay Viswanathan, Yiran Chen, Hiroaki Hayashi for useful discussion and feedback about ExplainaBoard.