emb-diversity 0.0.5

A package for measuring diversity in text and vector data

emb-diversity

A Python package for measuring data diversity on small- to medium-sized text datasets. All measures are calculating diversity based on embeddings, i.e., vector representations of your data. Depending on what embedding models you want to use, you are able to calculate semantic, stylistic and other types of diversity with our package.

This library is developed as part of the DataDivers project.

📖 Documentation: https://nlpsoc.github.io/Diversity-Measurement/

Install

Install the latest release from PyPI:

pip install emb-diversity

The first time you measure diversity, the default embedding model (all-mpnet-base-v2, ~420 MB) is downloaded from the Hugging Face Hub and cached locally, so later runs are fast and work offline.

Usage

Measuring the diversity of a dataset with our package is easy:

from emb_diversity import measure_diversity

# more style-diverse, more topic-uniform (music)
texts_a = [
    "I thoroughly enjoy the hair bands.",
    "songs of the 80's are the best.",
    "Hip Hop is going DOWNHILL!!!!!",
    "rock music just makes me feel good",
    "The 80's rocked!That generation had the best music!"
]

# Uses the default measures and semantic embeddings
print(measure_diversity(texts_a))
# Each result holds the score under 'value' and the configuration used under
# 'parameters' (shown as ... below):
# -> {'graph_entropy': {'value': 6.86..., ...},
#     'vendi_score':   {'value': 4.12..., ...},
#     'mean_pw_dist':  {'value': 0.69..., ...}}

Note that measuring the diversity of a dataset is usually only meaningful when comparing it to another datasets. The reason is that diversity values in isolation are not easily interpretable and are not bounded, sensitive to dataset size and sensitive to the used embedding space. Let's add another corpus.

# more style-uniform (formal), more topic-diverse
texts_b = [
    "I thoroughly enjoy the hair bands.",
    "They have not caused any harm to me.",
    "He has a very distinct walk.",
    "It depends on what they will pay.",
    "I would go out with the son of a preacher.",
]

print(measure_diversity(texts_a))
# -> {'graph_entropy': {'value': 6.86..., ...}, 'vendi_score': {'value': 4.12..., ...}, 'mean_pw_dist': {'value': 0.69..., ...}}

print(measure_diversity(texts_b))
# -> {'graph_entropy': {'value': 6.91..., ...}, 'vendi_score': {'value': 4.93..., ...}, 'mean_pw_dist': {'value': 0.98..., ...}}

When a measure considers a dataset to be more diverse, it will assign it a higher diversity value. Here, the three default measures consistently show that texts_b is more diverse than texts_a. This can change, when we change what diversity "axis" is considered, for example, "style" instead of "semantic".

# Use a different diversity axis, for style diversity AnnaWegmann/style-embeddings is the default
print(measure_diversity(texts_a, diversity_axis="style"))
# -> {'graph_entropy': {'value': 6.69..., ...}, 'vendi_score': {'value': 4.17..., ...}, 'mean_pw_dist': {'value': 0.93..., ...}}
print(measure_diversity(texts_b, diversity_axis="style"))
# -> {'graph_entropy': {'value': 6.32..., ...}, 'vendi_score': {'value': 2.24..., ...}, 'mean_pw_dist': {'value': 0.32..., ...}}

You can also specify a different embedding model with a HuggingFace identifier, for example, a model trained for Dutch. Be careful to use models that were trained on the diversity axis you are interested in, otherwise you might get some inconsistent results!

# Use a specific embedding model (here a small, fast SBERT model)
print(measure_diversity(texts_a, embedding_model="GroNLP/bert-base-dutch-cased"))
# -> {'graph_entropy': {'value': 6.61..., ...}, 'vendi_score': {'value': 1.89..., ...}, 'mean_pw_dist': {'value': 0.20..., ...}}
print(measure_diversity(texts_b, embedding_model="GroNLP/bert-base-dutch-cased"))
# -> {'graph_entropy': {'value': 6.80..., ...}, 'vendi_score': {'value': 1.52..., ...}, 'mean_pw_dist': {'value': 0.11..., ...}}

You can also use specific measures, see an overview here: https://nlpsoc.github.io/Diversity-Measurement/user-guide/measures.html. Use with caution. Some measures might be worse for your use case than others. We recommend to test whether your chosen measure and embedding space capture your diversity axis of interest.

# Run specific measures
print(measure_diversity(texts_a, measure=["diameter", "log_determinant"]))
# -> {'diameter': {'value': 0.94..., ...}, 'log_determinant': {'value': -0.93..., ...}}
print(measure_diversity(texts_b, measure=["diameter", "log_determinant"]))
# -> {'diameter': {'value': 1.0..., ...}, 'log_determinant': {'value': -0.06..., ...}}

Note that most measures return unbounded values that cannot be compared for datasets with differing sizes. Happy diversity measuring!

Table of Contents

• Install
• Usage
• Development
  • Development setup
  • Suggested Workflow for Collaboration
  • Working with uv
  • Docstring Style Guide
  • Adding New Measures
  • Adding New Diversity Axes
  • Building and publishing a release
• Funding
• Citation

Development

Development setup

To work on emb-diversity itself, install from a clone with uv:

git clone https://github.com/nlpsoc/Diversity-Measurement.git
cd Diversity-Measurement
uv sync --group dev          # runtime + dev tools (pytest, docs, ...)
source .venv/bin/activate

Use uv sync --no-group dev to install only the runtime dependencies.

Suggested Workflow for Collaboration

1. Create a new branch for your feature or bug fix:

   git checkout -b feature/my-feature
   

2. Make your changes in the codebase.
3. Run tests to ensure everything works as expected:

   pytest
   

4. Commit your changes with a descriptive message:

   git add .
   git commit -m "Add feature X"
   

5. Push your branch to the remote repository:

   git push origin feature/my-feature
   

6. Create a pull request on GitHub to merge your changes into the main branch and request a review from your team members.
7. Address any feedback from the review process.
8. Once approved, merge your pull request into the main branch.
9. Delete your branch after merging to keep the repository clean:

   git branch -d feature/my-feature
   git push origin --delete feature/my-feature
   

Working with uv

Adding Packages with uv add

To add packages to your project, always use uv add rather than uv pip install. This ensures that your dependencies are properly managed and recorded in your pyproject.toml. For example:

uv add <package-name>

Adding Packages to a Dev Group

If you need to add a package specifically to your development environment, you can add it to the dev group like this:

uv add --group dev <package-name>

Switching Between Dev and Standard Mode

After you are done with testing and want to go back to standard mode, run:

uv sync --no-group dev

This will disable all additional groups and just load your main project dependencies.

Best Practice: Run uv lock -U

Whenever you upgrade, downgrade, or change versions of packages, it's a good practice to run:

uv lock -U

This updates your uv.lock file to ensure all versions are consistent and everything is in sync.

Docstring Style Guide

This project uses Google-style docstrings which are automatically parsed by the Sphinx Napoleon extension.

Functions and Methods

def calculate_diversity(vectors: np.ndarray, method: str = "vendi") -> MeasureResult:
    """Calculate diversity score for a set of vectors.

    This function computes various diversity metrics for vector representations.
    The default method uses the Vendi Score which is based on matrix entropy.

    References:
        Cox, Samuel Rhys, Yunlong Wang, Ashraf Abdul, Christian von der Weth, and Brian Y. Lim. "Directed Diversity: Leveraging Language Embedding Distances for Collective Creativity in Crowd Ideation." Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, May 6, 2021, 1–35. https://doi.org/10.1145/3411764.3445782.

    Args:
        vectors: Array of shape (n_samples, n_features) containing the vectors.
        method: Diversity calculation method. Options are "vendi", "entropy",
            or "distinctness". Defaults to "vendi".

    Returns:
        A dict ``{"value": float, "parameters": {...}}`` where ``value`` is the
        diversity score (higher means more diverse; scores are unbounded, not
        limited to [0, 1]) and ``parameters`` records the configuration used.

    Raises:
        ValueError: If vectors array is empty or method is not recognized.

    Example:
        >>> vectors = np.array([[1, 0], [0, 1], [1, 1]])
        >>> result = calculate_diversity(vectors)
        >>> print(f"Diversity: {result['value']:.2f}")
        Diversity: 1.73
    """
    pass

Key Points

• One-line summary: Start with a brief summary in imperative mood ("Calculate", not "Calculates")
• Blank line: After the summary, add a blank line before any detailed description
• References: Add related papers
• Args: Document each parameter with type information
• Returns: Describe what the function returns
• Raises: Document exceptions that might be raised
• Example: Include usage examples when helpful
• Type hints: Use type hints in function signatures AND document them in docstrings

Section Headers

Use these section headers in docstrings:

• References: Related papers
• Args: — Function/method parameters
• Returns: — Return value description
• Raises: — Exceptions that may be raised
• Yields: — For generators
• Attributes: — For class attributes
• Example: or Examples: — Usage examples
• Note: — Important notes
• Warning: — Warnings about usage

Further reading: Google Style Guide · Sphinx Napoleon docs

Adding New Measures

When you add a new measure to src/emb_diversity/measures/:

1. Create a new file. A measure is a plain function (no decorator) with the signature def name(data, <params>, *, diversity_axis="semantic", embedding_model=None) -> MeasureResult. Call data, embedding_model = resolve_embeddings(data, diversity_axis, embedding_model) first (it embeds text input and returns the resolved model id), then return {"value": <float>, "parameters": {<params>, "embedding_model": embedding_model}}. Add a complete docstring following the style guide above.
2. Export it from src/emb_diversity/__init__.py if it should be part of the public API.
3. Register it in src/emb_diversity/measures_registry.py with measures.register("name", func).
4. Update docs/source/user-guide/measures.md — add a row for the new measure in the appropriate table.

Adding New Diversity Axes

Register a new axis in src/emb_diversity/axes_registry.py:

from emb_diversity.axes_registry import DiversityAxis, axes

axes.register(
    "multilingual",
    DiversityAxis(
        name="multilingual",
        default_model="sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
        description="Cross-lingual semantic diversity",
    ),
)

Update docs/source/user-guide/axes.md with the new axis.

Building and publishing a release

Releases are published by CI via PyPI Trusted Publishing (no API token is stored), in two stages — a TestPyPI dry run, then production PyPI:

1. Bump version in pyproject.toml, commit, and merge to main. A version number can be uploaded only once, so every release needs a new number — you cannot re-publish or overwrite an existing version.

2. Tag and push → TestPyPI. Pushing a v* tag triggers publish-testpypi.yml (it checks the tag matches the pyproject.toml version):

   git tag v0.0.1          # must match the version in pyproject.toml
   git push origin v0.0.1
   

   Verify the result at https://test.pypi.org/project/emb-diversity/.

3. Create a GitHub Release → PyPI. When the TestPyPI run looks good, create a GitHub Release for the tag. That triggers publish-pypi.yml, which uploads to real PyPI (https://pypi.org/project/emb-diversity/). Create the release either:

   • on GitHub: go to the repository's Releases page (right-hand sidebar of the repo, or .../releases) → Draft a new release → under Choose a tag pick the existing tag (e.g. v0.0.1) → add a title and notes → Publish release; or
   • with the GitHub CLI:

     gh release create v0.0.1 --title "v0.0.1" --notes "First release"
     

   Publishing the release (not just drafting it) is what triggers the workflow.

To build and validate locally before tagging (optional):

rm -rf dist              # clear artifacts from previous versions first
uv build                 # -> dist/emb_diversity-<version>.{tar.gz,whl}
uvx twine check dist/*   # validate metadata + that the README renders on PyPI

uv build only adds to dist/, so clear it first when building a new version — otherwise old artifacts linger and an upload would try (and fail) to re-publish them. CI doesn't need this: each run starts from a clean checkout.

Funding

This work is supported by the ERC Starting Grant DataDivers (101162980).

Citation

There is no paper yet, so if you use emb-diversity in your work, please cite the software:

@misc{emb_diversity,
  author = {Su, Cantao and Velayuthan, Menan and Ploeger, Esther and Nguyen, Dong and Wegmann, Anna},
  title  = {emb-diversity},
  url    = {https://github.com/nlpsoc/Diversity-Measurement},
}