You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@opennlp.apache.org by Jason Baldridge <ja...@gmail.com> on 2011/07/01 03:30:00 UTC

a call to arms (of sorts) ;)

Seth Grimes wrote the following article on Stanbol, OpenNLP and such:

http://www.cmswire.com/cms/enterprise-cms/iks-means-semantic-intelligence-for-content-management-but-will-it-survive-011848.php

I would highlight this paragraph on page 2:

OpenNLP has Apache incubation status, which is a sort-of provisional
acceptance by the Apache Software Foundation. It provides only basic NLP
functions. It doesn't support extraction of facts, events, or relationships,
nor sentiment or pattern-based information such as telephone numbers and
e-mail addresses. It appears, however, that FISE should be able to
accommodate other annotators, whether installed or invoked via calls to
entity-resolution Web services. There are many NLP engines that are more
advanced and capable than OpenNLP.

I guess a question is whether OpenNLP should be doing any of those things,
or just stay focussed on core NLP tasks that other software builds on.

Jason

-- 
Jason Baldridge
Assistant Professor, Department of Linguistics
The University of Texas at Austin
http://www.jasonbaldridge.com
http://twitter.com/jasonbaldridge

Re: a call to arms (of sorts) ;)

Posted by Olivier Grisel <ol...@ensta.org>.
2011/7/1 Jörn Kottmann <ko...@gmail.com>:
>
>> OpenNLP has Apache incubation status, which is a sort-of provisional
>> acceptance by the Apache Software Foundation. It provides only basic NLP
>> functions. It doesn't support extraction of facts, events, or
>> relationships,
>> nor sentiment or pattern-based information such as telephone numbers and
>> e-mail addresses. It appears, however, that FISE should be able to
>> accommodate other annotators, whether installed or invoked via calls to
>> entity-resolution Web services. There are many NLP engines that are more
>> advanced and capable than OpenNLP.
>>
>> I guess a question is whether OpenNLP should be doing any of those things,
>> or just stay focussed on core NLP tasks that other software builds on.

Well this part does require large indices and we are still working on
in as part of Stanbol to get good precision and recall. Once we get to
the point where the linking / disambiguation part is good enough we
could summarize which approach worked best and try to come up with a
pure solr4j / lucene based implementation that does not require to
deploy a complete Stanbol server.

-- 
Olivier
http://twitter.com/ogrisel - http://github.com/ogrisel

Re: a call to arms (of sorts) ;)

Posted by Jörn Kottmann <ko...@gmail.com>.
He also writes:

"One IKS annotation feature, found in some but not all competing 
annotation engines,
is ability to resolve identified entities to Semantic Web uniform 
resource identifiers (URIs),
IDs that uniquely designate/things/. URIs enable the DBpedia enrichment 
links."

Having an entity identification system would also be awesome to have in 
OpenNLP.
And I guess taking a learn able approach could work nicely with our 
annotations project.

I believe that is very important, because people who do NER, often need 
to identify entities.

Jörn

On 7/1/11 3:30 AM, Jason Baldridge wrote:
> Seth Grimes wrote the following article on Stanbol, OpenNLP and such:
>
> http://www.cmswire.com/cms/enterprise-cms/iks-means-semantic-intelligence-for-content-management-but-will-it-survive-011848.php
>
> I would highlight this paragraph on page 2:
>
> OpenNLP has Apache incubation status, which is a sort-of provisional
> acceptance by the Apache Software Foundation. It provides only basic NLP
> functions. It doesn't support extraction of facts, events, or relationships,
> nor sentiment or pattern-based information such as telephone numbers and
> e-mail addresses. It appears, however, that FISE should be able to
> accommodate other annotators, whether installed or invoked via calls to
> entity-resolution Web services. There are many NLP engines that are more
> advanced and capable than OpenNLP.
>
> I guess a question is whether OpenNLP should be doing any of those things,
> or just stay focussed on core NLP tasks that other software builds on.
>
> Jason
>


Re: a call to arms (of sorts) ;)

Posted by James Kosin <ja...@gmail.com>.
Jason,

-1, I don't think this classified OpenNLP correctly.

The corref functions extract content and information... and the name 
finder could easily be trained to be able to pick out email and other 
content.
Phone numbers would actually be more difficult; because there are no 
real clues that it really is a phone number in most cases... and it 
depends on the format which is not standardized.  ie:  US could use:
1(757)555-1212, (757)555-1212, 757-555-1212, 1-757-555-1212, etc and 
that doesn't include the international numbers or formats seen there 
that also appear in the US.
E-mail addresses are a bit easier with the '@' character and other 
requirements; but, how many e-mail addresses are put into documents or 
articles on a regular bases.  This would be more specialized but very 
doable with the current architecture.

The real problems in all these is finding and getting the training data 
for these types of information.

James

On 6/30/2011 9:30 PM, Jason Baldridge wrote:
> Seth Grimes wrote the following article on Stanbol, OpenNLP and such:
>
> http://www.cmswire.com/cms/enterprise-cms/iks-means-semantic-intelligence-for-content-management-but-will-it-survive-011848.php
>
> I would highlight this paragraph on page 2:
>
> OpenNLP has Apache incubation status, which is a sort-of provisional
> acceptance by the Apache Software Foundation. It provides only basic NLP
> functions. It doesn't support extraction of facts, events, or relationships,
> nor sentiment or pattern-based information such as telephone numbers and
> e-mail addresses. It appears, however, that FISE should be able to
> accommodate other annotators, whether installed or invoked via calls to
> entity-resolution Web services. There are many NLP engines that are more
> advanced and capable than OpenNLP.
>
> I guess a question is whether OpenNLP should be doing any of those things,
> or just stay focussed on core NLP tasks that other software builds on.
>
> Jason
>


Re: a call to arms (of sorts) ;)

Posted by Jörn Kottmann <ko...@gmail.com>.
On 7/1/11 8:53 AM, Olivier Grisel wrote:
> I think it would be great if OpenNLP could provide feature extractors
> and pre-trained models to perform Semantic Role Labeling for relation
> extraction (maybe for a restricted list of predicates). But this
> probably a non trivial task and it is only interesting if we can get a
> good F1 score (or at least precision) for this task.

+1 to work on this. We can simply do the work on the sandbox, and one day
we might feel it is now mature enough to be part of the OpenNLP Tools 
project
itself.

Jörn

Re: a call to arms (of sorts) ;)

Posted by Olivier Grisel <ol...@ensta.org>.
2011/7/1 Jason Baldridge <ja...@gmail.com>:
> Seth Grimes wrote the following article on Stanbol, OpenNLP and such:
>
> http://www.cmswire.com/cms/enterprise-cms/iks-means-semantic-intelligence-for-content-management-but-will-it-survive-011848.php
>
> I would highlight this paragraph on page 2:
>
> OpenNLP has Apache incubation status, which is a sort-of provisional
> acceptance by the Apache Software Foundation. It provides only basic NLP
> functions. It doesn't support extraction of facts, events, or relationships,
> nor sentiment or pattern-based information such as telephone numbers and
> e-mail addresses. It appears, however, that FISE should be able to
> accommodate other annotators, whether installed or invoked via calls to
> entity-resolution Web services. There are many NLP engines that are more
> advanced and capable than OpenNLP.
>
> I guess a question is whether OpenNLP should be doing any of those things,
> or just stay focussed on core NLP tasks that other software builds on.

I think it would be great if OpenNLP could provide feature extractors
and pre-trained models to perform Semantic Role Labeling for relation
extraction (maybe for a restricted list of predicates). But this
probably a non trivial task and it is only interesting if we can get a
good F1 score (or at least precision) for this task.

-- 
Olivier
http://twitter.com/ogrisel - http://github.com/ogrisel

Re: a call to arms (of sorts) ;)

Posted by Jörn Kottmann <ko...@gmail.com>.
Thanks for pointing us to this article Jason.

On 7/1/11 3:30 AM, Jason Baldridge wrote:
> Seth Grimes wrote the following article on Stanbol, OpenNLP and such:
>
> http://www.cmswire.com/cms/enterprise-cms/iks-means-semantic-intelligence-for-content-management-but-will-it-survive-011848.php
>
> I would highlight this paragraph on page 2:
>
> OpenNLP has Apache incubation status, which is a sort-of provisional
> acceptance by the Apache Software Foundation. It provides only basic NLP
> functions. It doesn't support extraction of facts, events, or relationships,
> nor sentiment or pattern-based information such as telephone numbers and
> e-mail addresses.

Yeah, he is right we have no engine to extract events, facts, or 
relationships. The
reason we do not have this is that never anyone contributed it. There is 
also no reason
why we should not have some more advanced rule based NER system.
> I guess a question is whether OpenNLP should be doing any of those things,
> or just stay focussed on core NLP tasks that other software builds on.

Most of the components we have are independent of each other, when we 
can get
support for the above mentioned task, why not?

I believe it would be great if we create jiras for new tasks we are 
interested in, find
people who would like to work on it and then support them with ideas and 
discussion
on how things could be done.
We could highlight these tasks on the website in a contributions wanted 
page.

Jörn