Testing Java HTML parsers

A few weeks ago I had to code some data export for which I had to test the speed of a bunch of Java HTML parser/cleaner libraries to have a valid XHTML output.
Jens proposed me to publish the results here and I thought that would be a really great idea.
At first, I'd like to present each one to you first and hopefully give some useful pieces of information on them. "Maven" means if the library can be found in the Maven repositories.

Jericho HTML-Parser
License: Eclipse Public License/LGPL
Maven: Yes
Has many features, like recognizing PHP tags and is easy to use.

JTidy
License: MIT License
Maven: Yes, but only the "old" builds
JTidy is tiny and pretty fast, can output wellformed XHTML.
Has bad internal exception handling(lots of empty catch blocks!)

HTMLCleaner
License: BSD License
Maven: No
DOM based, supports XPATH(really cool). Has a good bunch of confuguration options.

NekoHTML
License: Apache Software License
Maven: Yes
Good, fast, seems to be famous

TagSoup
License: Apache 2.0
Maven: Yes
Parses HTML and provides a SAX handler. Entry class is "Parser" to which a custom SAX handler can be given.

HTML Parser
License: GPL

HotSAX
License: LGPL
HotSAX looked pretty fast, but according to the homepage it is still in pre-alpha stadium, so it was not useful for my task.

Java Swing HTML parser
Comes with Sun Java.
XHTML is a more strict form of HTML 4.01, but this parser only supports HTML 3.2, so it was not in question for my purposes. Just wanted to mention it here.

Cobra: Java HTML Renderer and Parser
License: LGPL 2.1
Major plus of this one is that it is capable of parsing js and CSS, too. The browser is a good start(my admiration for that project!) although it fails all ACID tests. But nevertheless, this hasn't to say anything about the parser's quality.
One con is that this library is really slow.

Mozilla Java HTML-Parser
License: Mozilla Public License 1.1 (MPL 1.1)
The setup is not really suitable for a multi-developer setup so it fell out of the test selection.

Test results

Task was to load a predefined, really errorneous HTML document and select all <a> tags.
I used JUnit tests for each parser/cleaner and the measurement was taken ten times, while the first one was skipped due to the compilation time.

Rank	Name	Time/ms	Deviation/ms
1	HTMLCleaner	95	±18
2	HotSAX	124	±19
3	JTidy	158,3	±17
4	Jericho HTML	150	±59
5	NekoHtml	380	±44
6	TagSoup	439	±50,5
7	Cobra	675	±100

Jtidy is listed before Jericho HTML because it had the better deviation.I first used HTMLCleaner, because its advance in time was really big. The problem was that it couln't handle some of the real input data. HotSAX was pre-alpha(although the results are very good), so JTidy was my next choice as I needed reliability. I had not a single problem with it, it works really fine.
Last point to say is that the results of the Cobra parser are very bad...

Jens is working on a website so I can provide you the testing source code, I will put a link to there when it is online.
If anyone is interested in more detailed statistics, just contact me and I'll put them here.

As someone recently has begun to work on JTidy again, I'll try the SVN version soon and tell you the results in another post, promise! I hope they improved the exception handling.
Have a look at this piece of code:
public Node parse(InputStream in, OutputStream out)
{
Node document = null;

try
{
document = parse(in, null, out);
}
catch (FileNotFoundException fnfe) {}
catch (IOException e) {}

return document;
}
That's gruesome, isn't it?