This is the working area for the technical and legal teams. The focus of this effort is for recording ideas and considerations about improving machine detection and recognition of licenses and headers in code.
- Bugzilla bug 1302, which started this particular conversation: https://bugs.linuxfoundation.org/show_bug.cgi?id=1302
- Meeting minutes from Aug 6, 2015 Legal call: http://wiki.spdx.org/view/Legal_Team/Minutes/2015-08-06
- Matching Guidelines: http://spdx.org/spdx-license-list/matching-guidelines
- Pull Request Proposal for SPDX License List (Gary): https://docs.google.com/document/d/115Lis1SJV7Rp-XuNjIysU61urzFGjUOBPqEdVyGLsfI/edit
- Meeting minutes from Oct 29, 2015 Legal call: http://wiki.spdx.org/view/Legal_Team/Minutes/2015-10-29
- Initial proposal, full write-up (Kris): http://wiki.spdx.org/view/Legal_Team/Template_proposal
- Meeting minute from Feb 9 joint tech/legal call: http://wiki.spdx.org/view/Technical_Team/Minutes/2016-02-09
- Meeting minutes from Feb 18 legal call: http://wiki.spdx.org/view/Legal_Team/Minutes/2016-02-18
- Proposal for exact tag names and edits to matching guidelines: http://wiki.spdx.org/view/Legal_Team/Templatizing/tags-matching
- 'Plan for review of XML license files:'
(from JL) The current Matching Guidelines and license markup was developed over the past several years. While the guidelines were relatively easier to come up with, how much markup to use and how to implement it was an on-going discussion of much debate. As relevant to the proposals here, the discussion around how much markup to use recognized that more markup would increase consistent results by reducing the need for tool makers to interpret the guidelines. However, due to the amount of work of potentially having markup on every license file, a more conservative approach was taken with the recognition that we could always add more markup over time (but taking it away would be harder).
The template language was a result of a collaboration between members of the legal and technical teams. The goal of the template was developed with the following goals in mind:
- Enable both humans and, to a lesser extent, tools to be able to determine if two license texts were equivalent.
- Preserve the original language of the license text.
- Leverage existing annotations and tools (e.g. the use of regular expressions to denote matches to variable text).
The current templating language is documented in Appendix II of the specification (http://spdx.org/sites/spdx/files/SPDX-2.0.pdf)
- Please review the current Matching Guidelines here: http://spdx.org/spdx-license-list/matching-guidelines
- Also see Bugzilla bug 1302, which started this particular conversation: https://bugs.linuxfoundation.org/show_bug.cgi?id=1302
- Meeting minutes from Aug 6, 2015 Legal call: http://wiki.spdx.org/view/Legal_Team/Minutes/2015-08-06
(from Gary) As a result of the initial bug 1302 discussion, Gary suggested to enable pull requests to be made against the SPDX License List, which would encourage contributions to markup by others. Such pull requests/contributions would be moderated by the Legal team (especially regarding determinations of markup and potential impact on legal interpretation of the license) with help from the Tech team as needed. Everyone on the Aug 6 legal call agreed this would be a great approach. Gary, Jilayne, Kate have begun to discuss how to implement this. A process also needs to be outlined for how pull requests will be handled, etc. A rough draft process proposal can be found in the google document https://docs.google.com/document/d/115Lis1SJV7Rp-XuNjIysU61urzFGjUOBPqEdVyGLsfI/edit?usp=sharing
(from Kris) Things that would improve working with the data programmatically:
- Get away from spreadsheets as the index for the licenses. Use JSON or XML (leaning towards XML to keep the toolset the same as markup parsing). A tool to compile the spreadsheet (for humans to work with) down to an XML file would probably work.
- Put all information about a license in the same place for each license (headers, addons/exceptions/whatever). Likely within markup within the license files themselves.
- Use XML or some other widespread format for the markup. This ensures there are tools available to work with the data in any language without having to write custom parsing tools or use tools tied to a specific platform because they're the only ones that exist.
- As, and if, the data moves towards a more markup-heavy format, it might be wise to simply include the full text verbatim in its own section rather than "recreate" it from marked up data.
An example of what a more normalized (to xml) data set might look like:
<licenses> <license identifier="Apache-2.0" file="Apache-2.0.xml">Apache License 2.0</license> ... </licenses>
<license> <original>...full original text content...</original> <template> <header>...Licensed under the Apache License, version 2.0, etc....</header> <body>Some text blah blah <match regex="foo|bar"/> more text</body> </template> </license>
<template> <header>....etc</header> <body>...THE SOFTWARE IS PROVIDED "AS IS" AND <alt regex="THE AUTHOR">ISC</match> DISCLAIMS ALL WARRANTIES...</body> </template>
The above carries the useful property that the text content of the body tag is the original source without any modification; processing can instead replace "alt" tag contents with their regex attribute to "convert" to a matchable format.
I don't actually like XML, but it seems to be a good fit for this purpose. It helps keep the intrusiveness of the markup into the license text focused and to a minimum, lets us wrap entire sections as optional, keep the canonical original text if desired, and specify alternate license formats such as the short-form license header and any desired metadata all in one place. And there are tools to read XML in every language.
I recommend that the template content be *without* the parts that the matching guidelines say to ignore, and preferably it should have normalized or marked up bullets, etc. Best of all would be, simply, normalized text content in a full regular expression that can be used to match directly, rather than a bunch of markup that must be turned into a regular expression -- but I think this would make maintenance much harder. Again, a compilation step could help. There's a lot of work to be done here that can be qualified as preprocessing, and there's no reason we can't make sure the preprocessing is "done right" and publish the results, instead of requiring programmers to reimplement their own version of the preprocessing with uncertain results.
A particular note about normalizing text: Bullet items are one of the problematic points for parsing because it can be troublesome to distinguish a bullet point like "ii." from the end of a sentence that got wrapped to the start of the next line. (I actually encountered this in my testing; a saving grace is that the bullets tend to be things that are hard to confuse for words, but it's really easy to write a regular expression that isn't that smart.)
For what it's worth, I'm quite happy to contribute when it comes to tooling; I already have code to do the most of this anyway. -Kris
Use and Feedback re: Current Markup
Feedback was solicited on the legal and tech mailing lists to this end:
We have a task to add markup to some of the standard headers and have also had input to add/edit markup on existing licenses. As a result of the latter, it has been raised that perhaps the markup could be improved. Before adding more markup (to standard headers, license text or both), it seemed prudent to start a discussion as to whether the existing markup is effective. Please ponder the following questions: a) have you used the existing markup for matching purposes? i) if no, why not? ii) if yes, has it been helpful/effective? Could it be improved, and if so, how? (this will likely involve putting forward a proposal for review)
Please record your input in a section here, by name or tool or both:
For the SourceAuditor commercial tools, the markup is used to validate that 2 licenses are equivalent per the matching guidelines. The open source SPDX tools uses the markup is used in a number of ways. The "compareSpdx" and "compareMultipleSpdx" commands use the markup to determine if the licenses are equivalent. There are library methods implemented to compare license text and report if the license text matches any of the SPDX LicenseList.
In all cases above, the markup is used to compare 2 existing known license text. It is NOT used to match license text against a library of possible license matches. In the commercial tool, a separate algorithm implements this functionality and the markup language turned out to be too inefficient for this purpose - at least for the performance requirements of our application.
Note: When we originally discussed the markup language, we debated whether to cover the use case of searching a library of possible license matches and the decision was taken not to support this.
In my opinion, the markup works fine for matching two license texts. If we wanted to support a searching use case, we would need to modify/extend the markup language to enable this to be efficient.
Performance would be greatly improved if we bounded the RE's. Something which is within the spec, but is not implemented in the templates. I would support implementing bounding all RE's in an upcoming release and changing the spec to encourage bounding. I think we could pick a pretty practical upper bound that would improve the performance.
Also note - this is more of an issue with the "matching against a library of licenses" issue than simply comparing two license texts since the former is more performance sensitive.
I will share a few points from my experience in templatization. I currently use a different templatization syntax that predates SPDX, but the principle of using regular expressions embedded within the license text is similar.
The major barrier to me adopting the SPDX templates is insufficient templatization within the existing licenses. The SPDX templates currently encode what I perceive to be the ‘official’ variations, i.e. organization name, person name, product name etc. However, real-world licenses contain may minor variations that may be inconsequential from a legal perspective, but nonetheless do not warrant separating out as separate licenses. Here is an example from the GPL-3.0 notice where it is common to see two variations in one of the sentences:
distributed in the hope that it will be useful distributed in the hope that they will be useful
The example above is fairly uncontroversial, I would hope. However, there are plenty of other examples that border on having a legal impact. For example, in these two BSD-2-Clause variations it is necessary to consider whether the additional word constitutes an acceptable minor variation or warrants a different classification altogether:
Redistributions of source code must retain the above copyright notice, this list of… Redistributions of source code must retain the above copyright notice unmodified, this list of…
It is the grey cases like these that make expanding the use of templating difficult. Inevitably it leads to having to make some judgements about the impact of a particular word or phrase on the legal interpretation, something that I am aware SPDX tries to avoid.
Whether it is worth templating all the cases like these primarily depends on the goals of the SPDX templates. If they are for human use to see what official variations are permitted, then they are not necessary. On the other hand, if they are to be used by automated license scanning tools, then covering these cases is essential in order to have a tool that works effectively on real-world code. So I think an important point is to gain clarity on the purpose of the templates.
In terms of the current application of the templates, I have a technical concern over the use of unbounded regular expressions, for example:
<<var;name=copyrightHolderAsIs;original=THE COPYRIGHT HOLDERS AND CONTRIBUTORS;match=.+>>
This is unbounded because it will match any number of characters for the copyrightHolderAsIs field. The practical consequence of this is that regular expression matching can explode in terms of time. I don’t have a concrete example to hand, but my own experience with using the same unbounded regular expressions on real-world licenses is that I have seen it take minutes just to process one regular expression on a single file, and this does not scale well when there are millions of files to process. Clearly, in terms of English language there is no maximum size on the length of a copyright statement. Using an unbounded regular expression is therefore correct in theory but difficult to use in practice. I have had to use size bounded regular expressions in order to have a scanning tool that will complete in a reasonable time. The problem in switching to bounded regular expressions is in deciding on what is an acceptable upper bound on the size, and this can really only be judged by experimentation against real-world licenses, and does then require on-going tweaking as new license variations are discovered.
Neither of these are problems with templatization per-se, and they are more to do with the extent and way in which they are currently applied.
a) have you used the existing markup for matching purposes?
Yes and No: ScanCode uses an SPDX-inspired/derived markup, but instead of reusing the markup directly from the main license texts, markup is transformed in a simpler Template:Mustache-like syntax added to copies of these texts used only for detection purpose.
Because: - adding more markup to a reference license text makes this eventually no longer usable as a reference text and harder to read by humans - the many variations found in the wild make it hard to put all in a single template. - the markup syntax implies eventually an implementation using regular expressions. ScanCode does not use regex, but inverted indexes and string alignments.
ii) if yes, has it been helpful/effective? Could it be improved, and if so, how? (this will likely involve putting forward a proposal for review)
I think a simple markup is a very effective way to detect licenses with minor text variations and still call this an exact match. It is also a very effective way to indicate variations for humans. I find it hard personally to mix the human readability and technical detection concerns in the same file without compromises.
As food for thought, here are some examples of markup as used in ScanCode:
The syntax is using double curly braces to enclose variable parts. There is no regex involved. Optionally a number can be used after the opening braces to indicate the number of variable words, defaulting to 5 words. For instance Template:Copyright (c) 2015 Myco would match up to 5 words and Template:10 Copyright (c) 2015 Myco inc. would match up to 10 words.
I hope this helps even though this is a slightly different take.
a) have you used the existing markup for matching purposes?
i) if no, why not?
In a nutshell, partly due to timing of our dev efforts ahead of SPDX templatization rollout, partly due to performance in context of our internal Use Case of scanning every single open source file (not just a handful of applications).
Details: Black Duck has a corpus of license variants extracted from our Knowledge Base of around half a billion unique open source source/text and binary files. Several years ago, prior to SPDX license templatization, we went through multiple iterations of grouping license text variants, as well as license name/nickname variants. Our groupings were based on applying similarity algorithms, followed by human review. Our methodology has been in place for some time prior to the templates / regular expressions subsequently rolled out by SPDX.
Variations we've encountered, such as the street address of an organization, or typos / word substitutions that SPDX license templates might not have covered are some of the reasons we haven't yet gone through the exercise to see which license variants the SPDX templates might not 'match' to the license id's we've grouped them under.
We use our license scanner to scan (and rescan) every single open source file to populate our Knowledgebase of file-level license data. Our scanner uses multiple techniques to discover license references, but does not use regular expressions because of performance concerns.
Because Black Duck does not provide legal advice, consumers of our tools are able review the license text which our tools highlight in order to make a final determination. This fits well with the SPDX concept of separating discovered and concluded information.
Outlook: SPDX license templates / regexes aren't as useful or efficient as other matching techniques we have, when automatically bulk scanning large codebases to determine ‘LicenseFoundInFile'.
But when a human is in the loop producing a final SPDX Document with Concluded License, SPDX license templates / regexes could be useful to focus legal review on deviations which may or may not be significant.
a) have you used the existing markup for matching purposes? i) if no, why not?
Didn't use the markup. At the time haven't found documented the pattern keywords to find inside each license term. For example, one can find [xxxx]-[xxxx], [Owner Organization], <AUTHOR>. However, didn't seemed to be used consistently throughout the list of licenses. That was my opinion at that time, might certainly be proven erroneous.
If help is welcome, I can volunteer for normalizing the data for the complete list (top to bottom) and document the used keywords. However, for our own tooling usage as template I'd replace the text on some specific license terms so that we could later generate new licence docs where only the copyright holder changes.
Such example would be http://spdx.org/licenses/BSD-3-Clause-Attribution.html where: "Universidad de Palermo, Argentina" (http://www.palermo.edu/).'" becomes "<AUTHOR>".
I have been thinking about the proposed format in detail, and it has become clear to me that it will be undesirable to attempt to maintain a canonical single source of truth in the format we would want to expose to consumers. This is made clearest by the synonym matching rule: it is unfeasible to ask or expect a person editing a license file to find, identify, and mark up all candidates for synonym replacement. It is, however, easily possible to expose data that has been thus marked-up to consumers of the SPDX data. This implies that there must be a build step of some sort in the mix.
Given that there would be a build step, the case for altering the source markup is weakened somewhat; it's not strictly *necessary* to support (my) goal of providing data that can be consumed directly without any programmatic logic required to match a license, but still retains a number of benefits including: utility in editing/managing the license content itself, ease of integration into tooling required to produce the "built" product, ready extension, familiarity to newcomers/potential contributors, and even brevity.
Given the above, there are really two classes of markup to consider:
1) Markup used to clarify / identify information that is obvious to a human but difficult for a computer to interpret (for example: identifying bullet points)
2) Markup used to provide more robust matching that is easy for a computer to generate but difficult for a human to maintain (for example: identifying synonymous word spellings)
The most useful focus for markup in the first category is ease of use by a human editor: it should be unobtrusive and succinct, lending to ready interpretation by a reader/editor of the license text in a legal-meaningful way and imposing a minimum of technical knowledge for correct and confident interpretation.
The most useful focus for markup in the second category is ease of consumption by a program: it should be thorough and explicit, lending to immediately useful applications with little to no business logic required in the consuming code.
I'm working on the proposal discussed in the last legal call, but splitting it into two parts by these guidelines. This approach feels a bit messy to me, but I see no way around it other than to write custom tooling that can abstract the category-2 markup away from a human and let them work in terms of category-1 markup.
Compatibility with Current Markup
If we do change the format in a way which is incompatible with the existing markup, we may break existing applications. I originally did not think this would be much of an issue based on the responses above, but then I ran across the comment https://github.com/sindresorhus/spdx-license-list/issues/6#issuecomment-150606727 where the existing syntax is used to extract the license text. There may be other such applications that we are not aware of. -- Gary O'Neall
This is actually a point in favor of going ahead: if people are writing regular expressions to parse the current data, that is happening because the tools to do it properly aren't available to them. Markup change would obviously be a major version bump, so tools such as the above would still work and continue to function on the v2.0 data. New/updated tools, however, would become less fragile by virtue of using parsing tools that know how to deal with XML, and the markup itself can thus be extended to incorporate new needs without breaking existing fragile regular expressions. -- Kris
Markup to add/consider
[JL] This is simply a list of various specific licenses that may need additional markup or things to consider related to the matching guidelines, to ensure we don't forget to add/edit later!
- Apache-2.0 - instructions as to how to apply license, markup as optional? (from Kris)
- MIT - "THE AUTHORS OR COPYRIGHT HOLDERS" - markup to be variable (from DMG's student paper)
- add "noninfringment" and "non-infringement" and "non infringement" as equivalent words in Matching Guideline #8 list of words (from DMG's student paper)
- add "contributors" and "co-contributors" and "co contributors" as equivalent words in Matching Guideline #8 list of words, needs some discussion (from DMG's student paper)
- ISC license needs markup (and link fixed)