eyecite is an open source tool for extracting legal citations from text. It is used, among other things, to process millions of legal documents in the collections of CourtListener and Harvard's Caselaw Access Project, and has been developed in collaboration with both projects.
eyecite recognizes a wide variety of citations commonly appearing in American legal decisions, including:
- full case:
Bush v. Gore, 531 U.S. 98, 99-100 (2000) - short case:
531 U.S., at 99 - statutory:
Mass. Gen. Laws ch. 1, § 2 - law journal:
1 Minn. L. Rev. 1 - supra:
Bush, supra, at 100 - id.:
Id., at 101
All contributors, corrections, and additions are welcome!
If you use eyecite for your research, please consider citing our paper:
@article{eyecite,
title = {eyecite: A Tool for Parsing Legal Citations},
author = {Cushman, Jack and Dahl, Matthew and Lissner, Michael},
year = {2021},
journal = {Journal of Open Source Software},
volume = {6},
number = {66},
pages = {3617},
url = {https://doi.org/10.21105/joss.03617},
}
eyecite offers four core functions:
- Extraction: Recognize and extract citations from text, using a database that has been trained on over 55 million existing citations (see all of the citation patterns eyecite looks for over in reporters_db).
- Aggregation: Aggregate citations with common references (e.g., supra and id. citations) based on their logical antecedents.
- Annotation: Annotate citation-laden text with custom markup surrounding each citation, using a fast diffing algorithm.
- Cleaning: Clean and pre-process text for easy use with eyecite.
Read on below for how to get started quickly or for a short tutorial in using eyecite.
Please see the issues list on GitHub for things we need, or start a conversation if you have questions or need support.
If you are fixing bugs or adding features, before you make your first contribution, we'll need a signed contributor license agreement. See the template in the root of the repo for how to get that taken care of.
The API documentation is located here:
https://freelawproject.github.io/eyecite/
It is autogenerated whenever we release a new version. Unfortunately, for now we do not support old versions of the API documentation, but it can be browsed in the gh-pages branch if needed.
Install eyecite:
pip install eyecite
Here's a short example of extracting citations and their metadata from text using eyecite's main get_citations() function:
from eyecite import get_citations
text = """
Mass. Gen. Laws ch. 1, § 2 (West 1999) (barring ...).
Foo v. Bar, 1 U.S. 2, 3-4 (1999) (overruling ...).
Id. at 3.
Foo, supra, at 5.
"""
get_citations(text)
# returns:
[
FullLawCitation(
'Mass. Gen. Laws ch. 1, § 2',
groups={'reporter': 'Mass. Gen. Laws', 'chapter': '1', 'section': '2'},
metadata=Metadata(parenthetical='barring ...', pin_cite=None, year='1999', publisher='West', ...)
),
FullCaseCitation(
'1 U.S. 2',
groups={'volume': '1', 'reporter': 'U.S.', 'page': '2'},
metadata=Metadata(parenthetical='overruling ...', pin_cite='3-4', year='1999', court='scotus', plaintiff='Foo', defendant='Bar,', ...)
),
IdCitation(
'Id.',
metadata=Metadata(pin_cite='at 3')
),
SupraCitation(
'supra,',
metadata=Metadata(antecedent_guess='Foo', pin_cite='at 5', ...)
)
]
For a more full-featured walkthrough of how to use all of eyecite's functionality, please see the tutorial.
eyecite's full API is documented here, but here are details regarding its four core functions, its tokenization logic, and its debugging tools.
get_citations(), the main executable function, takes three parameters.
plain_text==> str: The text to parse. Should be cleaned first.remove_ambiguous==> bool, defaultFalse: Whether to remove citations that might refer to more than one reporter and can't be narrowed down by date.tokenizer==> Tokenizer, defaulteyecite.tokenizers.default_tokenizer: An instance of a Tokenizer object (see "Tokenizers" below).
For a given citation text such as "... 1 Baldwin's Rep. 1 ...", eyecite expects that the text
will be "clean" before being passed to get_citation. This means:
- Spaces will be single space characters, not multiple spaces or other whitespace.
- Quotes and hyphens will be standard quote and hyphen characters.
- No junk such as HTML tags inside the citation.
You can use clean_text to help with this:
from eyecite import clean_text, get_citations source_text = '<p>foo 1 U.S. 1 </p>' plain_text = clean_text(text, ['html', 'inline_whitespace', my_func]) found_citations = get_citations(plain_text)
See the Annotating Citations section for how to insert links into the original text using citations extracted from the cleaned text.
clean_text currently accepts these values as cleaners:
inline_whitespace: replace all runs of tab and space characters with a single space characterall_whitespace: replace all runs of any whitespace character with a single space characterunderscores: remove two or more underscores, a common error in text extracted from PDFshtml: remove non-visible HTML content using the lxml library- Custom function: any function taking a string and returning a string.
For simple plain text, you can insert links to citations using the annotate_citations function:
from eyecite import get_citations, annotate_citations plain_text = 'bob lissner v. test 1 U.S. 12, 347-348 (4th Cir. 1982)' citations = get_citations(plain_text) linked_text = annotate_citations(plain_text, [[c.span(), "<a>", "</a>"] for c in citations]) returns: 'bob lissner v. test <a>1 U.S. 12</a>, 347-348 (4th Cir. 1982)'
Each citation returned by get_citations keeps track of where it was found in the source text.
As a result, annotate_citations must be called with the same cleaned text used by get_citations
to extract citations. If you do not, the offsets returned by the citation's span method will
not align with the text, and your annotations will be in the wrong place.
If you want to clean text and then insert annotations into the original text, you can pass
the original text in as source_text:
from eyecite import get_citations, annotate_citations, clean_text source_text = '<p>bob lissner v. <i>test 1 U.S.</i> 12, 347-348 (4th Cir. 1982)</p>' plain_text = clean_text(source_text, ['html', 'inline_whitespace']) citations = get_citations(plain_text) linked_text = annotate_citations(plain_text, [[c.span(), "<a>", "</a>"] for c in citations], source_text=source_text) returns: '<p>bob lissner v. <i>test <a>1 U.S.</i> 12</a>, 347-348 (4th Cir. 1982)</p>'
The above example extracts citations from plain_text and applies them to
source_text, using a diffing algorithm to insert annotations in the correct locations
in the original text.
There is also a full_span attribute that can be used to get the indexes of the full citation, including the
pre- and post-citation attributes.
Note that the above example includes mismatched HTML tags: "<a>1 U.S.</i> 12</a>".
To specify handling for unbalanced tags, use the unbalanced_tags parameter:
unbalanced_tags="skip": annotations that would result in unbalanced tags will not be inserted.unbalanced_tags="wrap": unbalanced tags will be wrapped, resulting in<a>1 U.S.</a></i><a> 12</a>
Important: unbalanced_tags="wrap" uses a simple regular expression and will only work for HTML where
angle brackets are properly escaped, such as the HTML emitted by lxml.html.tostring. It is intended for
regularly formatted documents such as case text published by courts. It may have
unpredictable results for deliberately-constructed challenging inputs such as citations containing partial HTML
comments or <pre> tags.
If inserting text before and after isn't sufficient, supply a callable under the annotator parameter
that takes (before, span_text, after) and returns the annotated text:
def annotator(before, span_text, after):
return before + span_text.lower() + after
linked_text = annotate_citations(plain_text, [[c.span(), "<a>", "</a>"] for c in citations], annotator=annotator)
returns:
'bob lissner v. test <a>1 u.s. 12</a>, 347-348 (4th Cir. 1982)'
Once you have extracted citations from a document, you may wish to resolve them to their common references.
To do so, just pass the results of get_citations() into resolve_citations(). This function will
do its best to resolve each "full," "short form," "supra," and "id" citation to a common Resource object,
returning a dictionary that maps resources to lists of associated citations:
from eyecite import get_citations, resolve_citations
text = 'first citation: 1 U.S. 12. second citation: 2 F.3d 2. third citation: Id.'
found_citations = get_citations(text)
resolved_citations = resolve_citations(found_citations)
returns (pseudo):
{
<Resource object>: [FullCaseCitation('1 U.S. 12')],
<Resource object>: [FullCaseCitation('2 F.3d 2'), IdCitation('Id.')]
}
Importantly, eyecite performs these resolutions using only its immanent knowledge about each citation's
textual representation. If you want to perform more sophisticated resolution (e.g., by augmenting each
citation with information from a third-party API), simply pass custom resolve_id_citation(),
resolve_supra_citation(), resolve_shortcase_citation(), and resolve_full_citation()
functions to resolve_citations() as keyword arguments. You can also configure those functions to
return a more complex resource object (such as a Django model), so long as that object inherits the
eyecite.models.ResourceType type (which simply requires hashability). For example, you might implement
a custom full citation resolution function as follows, using the default resolution logic as a fallback:
def my_resolve(full_cite):
# special handling for resolution of known cases in our database
resource = MyOpinion.objects.get(full_cite)
if resource:
return resource
# allow normal clustering of other citations
return resolve_full_citation(full_cite)
resolve_citations(citations, resolve_full_citation=my_resolve)
returns (pseudo):
{
<MyOpinion object>: [<full_cite>, <short_cite>, <id_cite>],
<Resource object>: [<full cite>, <short cite>],
}
Internally, eyecite works by applying a list of regular expressions to the source text to convert it to a list of tokens:
In [1]: from eyecite.tokenizers import default_tokenizer
In [2]: list(default_tokenizer.tokenize("Foo v. Bar, 123 U.S. 456 (2016). Id. at 457."))
Out[2]:
['Foo',
StopWordToken(data='v.', ...),
'Bar,',
CitationToken(data='123 U.S. 456', volume='123', reporter='U.S.', page='456', ...),
'(2016).',
IdToken(data='Id.', ...),
'at',
'457.']
Tokens are then scanned to determine values like the citation year or case name for citation resolution.
Alternate tokenizers can be substituted by providing a tokenizer instance to get_citations():
from eyecite.tokenizers import HyperscanTokenizer hyperscan_tokenizer = HyperscanTokenizer(cache_dir='.hyperscan') cites = get_citations(text, tokenizer=hyperscan_tokenizer)
test_FindTest.py includes a simplified example of using a custom tokenizer that uses modified regular expressions to extract citations with OCR errors.
eyecite ships with two tokenizers:
The default tokenizer uses the pyahocorasick library to filter down eyecite's list of
extractor regexes. It then performs extraction using the builtin re library.
The alternate HyperscanTokenizer compiles all extraction regexes into a hyperscan database
so they can be extracted in a single pass. This is far faster than the default tokenizer
(exactly how much faster depends on how many citation formats are included in the target text),
but requires the optional hyperscan dependency that has limited platform support.
See the "Installation" section for hyperscan installation instructions and limitations.
Compiling the hyperscan database takes several seconds, so short-running scripts may want to provide a cache directory where the database can be stored. The directory should be writeable only by the user:
hyperscan_tokenizer = HyperscanTokenizer(cache_dir='.hyperscan')
If you want to see what metadata eyecite is able to extract for each citation, you can use dump_citations.
This is primarily useful for developing eyecite, but may also be useful for exploring what data is available to you:
In [1]: from eyecite import dump_citations, get_citations
In [2]: text="Mass. Gen. Laws ch. 1, § 2. Foo v. Bar, 1 U.S. 2, 3-4 (1999). Id. at 3. Foo, supra, at 5."
In [3]: cites=get_citations(text)
In [4]: print(dump_citations(get_citations(text), text))
FullLawCitation: Mass. Gen. Laws ch. 1, § 2. Foo v. Bar, 1 U.S. 2, 3-4 (1
* groups
* reporter='Mass. Gen. Laws'
* chapter='1'
* section='2'
FullCaseCitation: Laws ch. 1, § 2. Foo v. Bar, 1 U.S. 2, 3-4 (1999). Id. at 3. Foo, s
* groups
* volume='1'
* reporter='U.S.'
* page='2'
* metadata
* pin_cite='3-4'
* year='1999'
* court='scotus'
* plaintiff='Foo'
* defendant='Bar,'
* year=1999
IdCitation: v. Bar, 1 U.S. 2, 3-4 (1999). Id. at 3. Foo, supra, at 5.
* metadata
* pin_cite='at 3'
SupraCitation: 2, 3-4 (1999). Id. at 3. Foo, supra, at 5.
* metadata
* antecedent_guess='Foo'
* pin_cite='at 5'
In the real terminal, the span() of each extracted citation will be highlighted.
You can use the context_chars=30 parameter to control how much text is shown before and after.
Installing eyecite is easy.
poetry add eyecite
Or via pip:
pip install eyecite
Or install the latest dev version from github:
pip install https://github.com/freelawproject/eyecite/archive/main.zip#egg=eyecite
To use HyperscanTokenizer you must additionally install the python hyperscan
library and its dependencies. python-hyperscan officially supports only x86 linux, though other configurations may be
possible.
Hyperscan installation example on x86 Ubuntu 20.04:
apt install libhyperscan-dev pip install hyperscan
Hyperscan installation example on x86 Debian Buster:
echo 'deb http://deb.debian.org/debian buster-backports main' > /etc/apt/sources.list.d/backports.list apt install -t buster-backports libhyperscan-dev pip install hyperscan
Hyperscan installation example with homebrew on x86 MacOS:
brew install hyperscan pip install hyperscan
- Update CHANGES.md.
- Update version info in
pyproject.tomlby runningpoetry version [major, minor, patch].
For an automated deployment, tag the commit with vx.y.z, and push it to main. An automated deploy and documentation update will be completed for you.
For a manual deployment, run:
poetry publish --build
You will probably also need to push new documentation files to the gh-pages branch.
eyecite comes with a robust test suite of different citation strings that it is equipped to handle. Run these tests as follows:
python3 -m unittest discover -s tests -p 'test_*.py'
If you would like to create mock citation objects to assist you in writing your own local tests, import and use the following functions for convenience:
from eyecite.test_factories import (
case_citation,
id_citation,
supra_citation,
unknown_citation,
)
When a pull request is generated for changes from changes to eyecite, a github workflow will automatically trigger. The workflow, benchmark.yml will test improvements in accuracy and speed against the current main branch.
The results are committed to an artifacts branch, and an ever updating comment in the PR comments with the output.
This repository is available under the permissive BSD license, making it easy and safe to incorporate in your own libraries.
Pull and feature requests welcome. Online editing in GitHub is possible (and easy!).