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

export GO111MODULE=on
go mod download
go build -v gopkg.in/src-d/go-license-detector.v3/cmd/license-detector

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 (for a single license detection):

import (
    "gopkg.in/src-d/go-license-detector.v3/licensedb"
    "gopkg.in/src-d/go-license-detector.v3/licensedb/filer"
)

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

Library (for a convenient data structure that can be formatted as JSON):

import (
	"encoding/json"
	"fmt"

	"gopkg.in/src-d/go-license-detector.v3/licensedb"
)

func main() {
	results := licensedb.Analyse("/path/to/project1", "/path/to/project2")
	bytes, err := json.MarshalIndent(results, "", "\t")
	if err != nil {
		fmt.Printf("could not encode result to JSON: %v\n", err)
	}
	fmt.Println(string(bytes))
}

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

$ cd $(go env GOPATH)/src/gopkg.in/src-d/go-license-detector.v3/licensedb
$ mkdir dataset && cd dataset
$ unzip ../dataset.zip
$ # src-d/go-license-detector
$ time license-detector * \
  | grep -Pzo '\n[-0-9a-zA-Z]+\n\tno license' | grep -Pa '\tno ' | wc -l
$ # benbalter/licensee
$ time ls -1 | xargs -n1 -P4 licensee \
  | grep -E "^License: Other" | wc -l
$ # google/licenseclassifier
$ time find -type f -print | xargs -n1 -P4 identify_license \
  | cut -d/ -f2 | sort | uniq | wc -l
$ # boyter/lc
$ time lc . \
  | grep -vE 'NOASSERTION|----|Directory' | cut -d" " -f1 | sort | uniq | wc -l
$ # amzn/askalono
$ echo '#!/bin/sh
result=$(askalono id "$1")
echo "$1
$result"' > ../askalono.wrapper
$ time find -type f -print | xargs -n1 -P4 sh ../askalono.wrapper | grep -Pzo '.*\nLicense: .*\n' askalono.txt | grep -av "License: " | cut -d/ -f 2 | sort | uniq | wc -l
$ # LiD
$ time license-identifier -I dataset -F csv -O lid
$ cat lid_*.csv | cut -d, -f1 | cut -d"'" -f 2 | grep / | cut -d/ -f2 | sort | uniq | wc -l

Regenerate binary data

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

make bindata.go