[ANN][SoC] Ariel 0.1.0 released

= Ariel release 0.1.0

== About - Ariel: A Ruby Information Extraction Library
Ariel is a library that allows you to extract information from semi-structured
documents (such as websites). It is different to existing tools because rather
than expecting the developer to write rules to extract the desired
information, Ariel will use a small number of labeled examples to generate
and learn effective extraction rules. It is developed by Alex Bradbury and
released under the MIT license. Ariel was started as a Google Summer of Code
project mentored by Austin Ziegler in 2006.

== Install
gem install ariel

== Announcement

I'm happy to announce the release of Ariel 0.1.0, the result of my Summer of
Code work. This release should be easy to use, very functional, and hopefully
useful - so it's worth trying out. I've put a lot of effort in to writing
clear and straightforward documentation to get your started, so take a look
at the docs available at http://ariel.rubyforge.org. In particular, flick
through the tutorial and quick start guide. If you're interested, you may
also want to take a look at the theory page where I've made a good start on
describing the method Ariel uses to learn extraction rules. If you have any
problems or find any bugs, just send me an email or add it to the issue
tracker (see link below). Enjoy. See the FAQ for a vim snippet to make
labeling examples a little easier.

== Quickstart/Basic usage

* @require 'ariel'@
* Define a structure for the information you wish to extract:
    structure = Ariel::Node::Structure.new do |r|
      r.item :title
      r.item :body
      r.list :comments do |c|
        c.list_item :comment do |d|
          d.item :author
          d.item :body
        end
      end
     end
* Collect a few examples of the sort of document you wish to extract
information from (pages from the same website for instance).
* Label each example with tags such as <l:title>, <l:comment> and so on in the
  relevant places.
* Ariel.learn structure, labeled_file1, labeled_file2, labeled_file3
* Find the documents you want to extract information from.
* extractions = Ariel.extract structure, unlabeled_file1,
  unlabeled_file2
* extractions[0].search('comments/*/body').each {|e| puts e.extracted_text}
=> "Great stuff, loving it", "I love life", .....
* extractions[0].at('comments/34') => nil</tt> (there is no 34th comment, #at
  returns the first result rather than an array of matches).

== Credits
Ariel is developed by Alex Bradbury as a Google Summer of Code project under
the mentoring of Austin Ziegler.

== Links
SVN Repository: http://rubyforge.org/projects/ariel
Issue tracker: http://code.google.com/p/ariel/issues/
Documentation/homepage: http://ariel.rubyforge.org
RDoc: http://ariel.rubyforge.org/rdoc/

Very impressive library! I remember when you posted about this at the
beginning of the summer. I took the library and pointed it at the
USCCBs online version of the Bible and got some very impressive
results! It was able to identify book, chapter and verse with only a
few examples.

With only 3 sample pages, and the following structure I was able to
get very reliable results

structure = Ariel::Node::Structure.new do |r|
  r.item :book do |b|
    b.item :title
    b.item :chapter do |c|
      c.item :title
      c.list :verses do |v|
        v.list_item :verse
      end
    end
  end
end

I was particularly impressed that it understood how I re-used the
title tag in different contexts (i.e. for the book and chapter title).

If you'd like me to email you my structure files, my examples, and the
tests I use them in I'd be glad to. It's three small files but
probably too much for the list.

Comments and questions I jotted down while playing with this:

* Most of the chapter pages have footnotes interspersed throughout the
text. These are hyperlinks to anchors below the main body of the
chapter. Can Ariel correctly identify footnotes and pull in the text
for them?

* Ariel gets confused if you have tags in the example document that
are not in the structure, but look like Ariel tags. Example: I had a
<l:verses> tag which contained <l:verse> tags. I realized 'verses'
was not needed so removed the item definition from the structure but
not the example file. Ariel was not able to find the verse items until
I removed the <l:verses> tag.

* Typing "extracted_text" to get the text of each node is cumbersome.
If its not already, maybe overload to_s on nodes to display the text?

* Dealing with items is a little cumbersome. To get the number of
verses in a chapter, I have to type
e[:book][:chapter][:verses].children.length. Since I am already
treating the nodes like arrays, having a 'length' method would be
nice: e[:book][:chapter][:verses].length

* Better progress indication during learning phase. Hard to tell if
program is hung or if it is managing to do something. The CPU is
pegged but its hard to tell what progress is being made.

* More info about the search/at methods and expressions they can take.
RDOC and the tutorial only hint at what you can do.

* Falls apart if tags entered are not well formed and gives little
indication why. For example, I had missed an end tag on a list_item.
The program didnt use the examples provided (said "learning node X
with 2 examples" when I had 3) and then would quit with the error "No
examples are suitable for exhaustive rule learning"

Hi,

= Ariel release 0.1.0

== About - Ariel: A Ruby Information Extraction Library

_WAY_ cool!

Only one reply post just wasn't enough I thought, very cool project indeed :slight_smile:

Kash

···

--
Using Opera's revolutionary e-mail client: http://www.opera.com/mail/

Very impressive library! I remember when you posted about this at the
beginning of the summer. I took the library and pointed it at the
USCCBs online version of the Bible and got some very impressive
results! It was able to identify book, chapter and verse with only a
few examples.

Thank you so much for taking the time for writing this detailed email.

With only 3 sample pages, and the following structure I was able to
get very reliable results

structure = Ariel::Node::Structure.new do |r|
  r.item :book do |b|
    b.item :title
    b.item :chapter do |c|
      c.item :title
      c.list :verses do |v|
        v.list_item :verse
      end
    end
  end
end

I was particularly impressed that it understood how I re-used the
title tag in different contexts (i.e. for the book and chapter title).

This is because of the way it checks for nesting when extracting label tags,
which is why it gets confused when you have an extra tag.
<l:item><l:title>....</l:title>...</l:item><l:title>...</l:title>
When it encounters the first <l:item>, it increments the nesting level by one,
and again when it encounters the first <l:title>. The two closing tags
decrement it, and then when <l:title> (which we're searching for in this
example) is encountered and the nesting level is 0, we know it's the right
one.

If you'd like me to email you my structure files, my examples, and the
tests I use them in I'd be glad to. It's three small files but
probably too much for the list.

Yes, please do email them to me.

Comments and questions I jotted down while playing with this:

* Most of the chapter pages have footnotes interspersed throughout the
text. These are hyperlinks to anchors below the main body of the
chapter. Can Ariel correctly identify footnotes and pull in the text
for them?

I'll have to take a look at your examples, but if I understand correctly not
really. Perhaps you can extract footnote references (as in #footnote34) from
the relevant page section, and separately extract all footnotes (with a
footnote.reference). Then you can match them up, is this what you're trying
to do? I haven't thought about having linked items like that
before...interesting.

* Ariel gets confused if you have tags in the example document that
are not in the structure, but look like Ariel tags. Example: I had a
<l:verses> tag which contained <l:verse> tags. I realized 'verses'
was not needed so removed the item definition from the structure but
not the example file. Ariel was not able to find the verse items until
I removed the <l:verses> tag.

The checking and error reporting when parsing labeled documents isn't that
great at the moment, I'll have to rework it a bit to make it easier to work
out where there are errors if they exist. I'm not sure I follow here, you
have :verses in your example above. Putting list items in a container is the
recommended way of doing things:

<ul>
<l:verses><li><l:verse>Verse 1<l:verse></li>
<li><l:verse>Verse 2</l:verse></li>
<li><l:verse>Verse 3<l:verse></li></l:verses>
</ul>

You could put the <l:verses> right next to the first <l:verse> and the same
with the </l:verse> if you wanted.

When defining structure, you should really only put a list_item as a single
child of a list (internally a list is just an item.....I mean if you think
about it extracting the whole list above is the same as extracting any other
piece of text that occurs once). If you have multiple list_item's at the same
level I think you'd get a lot of things going wrong. I'll add a check for
this - a list_item should have no siblings.

* Typing "extracted_text" to get the text of each node is cumbersome.
If its not already, maybe overload to_s on nodes to display the text?

Will do this.

* Dealing with items is a little cumbersome. To get the number of
verses in a chapter, I have to type
e[:book][:chapter][:verses].children.length. Since I am already
treating the nodes like arrays, having a 'length' method would be
nice: e[:book][:chapter][:verses].length

This is a case where I'd like you to use #search. It's easier for you too -
what if no value for chapter was extracted for whatever reason? You'd get an
error with the code above (because you'd be using [] on the nil value
returned by e[:book][:chapter]), but e.search('book/chapter/verses/*').length
would just return 0. (e/'book/chapter/verses/*').length is equivalent. I
guess I haven't defined #size/#length because it makes sense when you're
talking about a list, but means little when you're talking about
e.chapter.size I think. #size = number of children seems reasonable enough
though, I'll add that.

* Better progress indication during learning phase. Hard to tell if
program is hung or if it is managing to do something. The CPU is
pegged but its hard to tell what progress is being made.

You're seeing at least messages like this?:
info: Learning rules for node version_history with 2 examples
info: Learnt start rules [#<Ariel::Rule:0xb79d7c64 @exhaustive=false,
@direction=:forward, @landmarks=[["<td>"], ["Versions"], ["<td>"]]>]

You can fill your screen with status updates by using the -D switch if using
the command line script, by setting $DEBUG or by Ariel::Log.set_level :debug

It's hard to know what status information to output. Other than printing the
name of the item we're learning rules for and the rules as they're learnt,
I'm not sure what else would mean something to the user who isn't familiar
with Ariel internals and wouldn't be too excessively verbose. If you just
want to know something's going on behind the scenes, then try one of the
switches above.

* More info about the search/at methods and expressions they can take.
RDOC and the tutorial only hint at what you can do.

They're very limited at the moment, there's nothing more to them than listing
parameters between /, and * are supported much like directory globbing.
There's no way to specify certain parameters (like to select only verses
lists with more than 5 children). But then Ruby has powerful array operations
like #select and #reject for this sort of querying. I made this interface as
basic as possible, not being sure what people would need/use. What sort of
queries would you like to perform? I was planning on adding range selection,
so you could do e.search 'book/chapter/verses/[0..5/whatever'. Clearly in
your structure it would be as easy to just slice the result array.

This is where I could really use some practical examples to beef out the
documentation. Maybe some of the functionalities people might want are easily
provided using Ruby's standard library, but the documentation should give
pointers on where to look, and suggest useful techniques.

* Falls apart if tags entered are not well formed and gives little
indication why. For example, I had missed an end tag on a list_item.
The program didnt use the examples provided (said "learning node X
with 2 examples" when I had 3) and then would quit with the error "No
examples are suitable for exhaustive rule learning"

Mentioned this problem with error reporting above. I've added it to the issue
tracker, this is definitely something that makes Ariel less user friendly.

Can you recreate this with one of your labeled files? The message only
learning node x with 2 examples when there are 3 seems a little odd. The "No
examples are suitable for exhaustive rule learning" takes a little bit of
explaining, that I probably don't have time to do properly. But basically,
taking the example I used above. I could have labeled it like this:

<ul>
<li><l:verses><l:verse>Verse 1<l:verse></li>
<li><l:verse>Verse 2</l:verse></li>
<li><l:verse>Verse 3<l:verse></l:verses></li>
</ul>

Remember how Ariel learns rules - it finds a rule that consumes all the tokens
up to the one that is labeled. Assume we're finding start rules (end rules
have the same issue), there are no tokens between the beginning of the
extracted verses list and the label. So the only possible rule is an empty
rule with no landmarks, which of course can't be applied exhaustively to
iterate over the whole list. This is why this example must be ignored in the
current (somewhat naive) implementation. I find it works pretty well, it just
could be better. Looking in to this is one of my post-SoC aims. A problem
with lists is that you don't want to make users label every item, or even
count them. The good thing is that lists are generally very regular and have
simple rules to split them.

Returning to the example, if we can't make a start rule that locates the start
of the first verse, then how do we extract it? The answer is the end rule,
say we have an end rule that has </li> as a landmark, the lowest end location
will have a position less than the first start location, so Ariel assumes
that all tokens from the first to the lowest end location are a list item.
Hope that makes a little sense. This isn't something I've explained much/at
all in the documentation, because it requires quite a lot of understanding of
how Ariel works, and I'm hoping to look at ways to change the way this works.

Thanks so much again for taking the time to look through my project and share
your experiences, hope my response has been some help. Apologies if it's a
little long.

Regards,

Alex

···

On Wednesday 23 August 2006 00:25, Justin Bailey wrote:

A. S. Bradbury wrote:

They're very limited at the moment, there's nothing more to them than listing parameters between /, and * are supported much like directory globbing. There's no way to specify certain parameters (like to select only verses lists with more than 5 children). But then Ruby has powerful array operations like #select and #reject for this sort of querying. I made this interface as basic as possible, not being sure what people would need/use. What sort of queries would you like to perform? I was planning on adding range selection, so you could do e.search 'book/chapter/verses/[0..5/whatever'. Clearly in your structure it would be as easy to just slice the result array.

If I look at the queries that can be performed now they look somewhat similar to XPath expressions. The structure you define also looks a lot like an XML structure, so maybe the data extracted can be easily converted to XML? This would allow you to use REXML's powerful XPath support to query the results. Next to this the results can easily be exported as XML too.

Regards,

Peter

I hadn't got round to XML export yet, but this seems like an excellent reason
to bump it up higher on the todo list. I wanted to avoid reimplementing
XPath, which is why I chose a very simple query interface that is quite
similar to globbing directories (no ** at the moment). XML export of course
means I don't have to reimplement this myself, excellent. I'll add #to_xml to
the extracted data structures.

Alex

···

On Wednesday 23 August 2006 22:35, Peter C. Verhage wrote:

A. S. Bradbury wrote:
> They're very limited at the moment, there's nothing more to them than
> listing parameters between /, and * are supported much like directory
> globbing. There's no way to specify certain parameters (like to select
> only verses lists with more than 5 children). But then Ruby has powerful
> array operations like #select and #reject for this sort of querying. I
> made this interface as basic as possible, not being sure what people
> would need/use. What sort of queries would you like to perform? I was
> planning on adding range selection, so you could do e.search
> 'book/chapter/verses/[0..5/whatever'. Clearly in your structure it would
> be as easy to just slice the result array.

If I look at the queries that can be performed now they look somewhat
similar to XPath expressions. The structure you define also looks a lot
like an XML structure, so maybe the data extracted can be easily
converted to XML? This would allow you to use REXML's powerful XPath
support to query the results. Next to this the results can easily be
exported as XML too.