Jay Taylor's notes

back to listing index

src-d/go-license-detector

[web search]
Original source (github.com)
Tags: golang go license-identification github.com
Clipped on: 2018-05-11
Go Makefile

README.md

go-license-detector

Project license detector - a command line application and a library, written in Go. It scans the given directory for license files, normalizes and hashes them and outputs all the fuzzy matches with the list of reference texts. The returned names follow SPDX standard. Read the blog post.

Why? There are no similar projects which can be compiled into a native binary without dependencies and also support the whole SPDX license database (≈400 items). This implementation is also fast, requires little memory, and the API is easy to use.

The license texts are taken directly from license-list-data repository. The detection algorithm is not template matching; this directly implies that go-license-detector does not provide any legal guarantees. The intended area of it's usage is data mining.

Installation

go get -v gopkg.in/src-d/go-license-detector.v2/...

Contributions

...are welcome, see CONTRIBUTING.md and code of conduct.

License

Apache 2.0, see LICENSE.md.

Algorithm

  1. Find files in the root directory which may represent a license. E.g. LICENSE or license.md.
  2. If the file is Markdown or reStructuredText, render to HTML and then convert to plain text. Original HTML files are also converted.
  3. Normalize the text according to SPDX recommendations.
  4. Split the text into unigrams and build the weighted bag of words.
  5. Calculate Weighted MinHash.
  6. Apply Locality Sensitive Hashing and pick the reference licenses which are close.
  7. For each of the candidate, calculate the Levenshtein distance - D. the corresponding text is the single line with each unigram represented by a single rune (character).
  8. Set the similarity as 1 - D / L where L is the number of unigrams in the quieried license.

This pipeline guarantees constant time queries, though requires some initialization to preprocess the reference licenses.

If there are not license files found:

  1. Look for README files.
  2. If the file is Markdown or reStructuredText, render to HTML and then convert to plain text. Original HTML files are also converted.
  3. Scan for words like "copyright", "license" and "released under". Take the neighborhood.
  4. Run Named Entity Recognition (NER) over that surrounding context and extract the possible license name.
  5. Match it against the list of license names from SPDX.

Usage

Command line:

license-detector /path/to/project
license-detector https://github.com/src-d/go-git

Library:

import (
    "gopkg.in/src-d/go-license-detector.v2/licensedb"
    "gopkg.in/src-d/go-license-detector.v2/licensedb/filer"
)

func main() {
	licenses, err := licensedb.Detect(filer.FromDirectory("/path/to/project"))
}

Quality

On the dataset of ~1000 most starred repositories on GitHub as of early February 2018 (list), 99% of the licenses are detected. The analysis of detection failures is going in FAILURES.md.

Comparison to other projects on that dataset:

Detector Detection rate Time to scan, sec
go-license-detector 99% (897/902) 13.5
benbalter/licensee 75% (673/902) 111
google/licenseclassifier 76% (682/902) 907
boyter/lc 88% (797/902) 548
amzn/askalono 87% (785/902) 165
LiD 94% (847/902) 3660
How this was measured

Regenerate binary data

The SPDX licenses are included into the binary. To update them, run

make bindata.go
Press h to open a hovercard with more details.