Hi Anton,
Extending the information extraction by adding recognition for new
classes, predicates or other language constructs is a complex task. As
explained in the documentation, such modifications consist of four steps.
1. Gather text samples that mention the new predicate. In your case,
find documents that mention acquisitions. Try to answer the following
questions:
- What words are used to describe acquisitions? Do those words
describe the event unambiguously?
- Does reading a single sentence give you enough information for the
event?
- Is it clear from the text who acquired what?
2. Add the new predicate the ontology - see
http://ontotext.com/kim/doc/sys-doc/ConfigInformationExtraction.html
3. Extend the KIM information extraction.
- Our algorithms are built on GATE and thus you can use arguably the
best natural language development environment to edit them.
- Play with GATE 4 at http://www.gate.ac.uk/ . GATE has extensive
documentation along with movies that explain working with the graphical
interface.
- GATE allows you to define JAPEs. If you are familiar with regular
expression, you will be able to understand JAPEs quickly. They are
regular expression which work on whole words, sentences and parts of
speech, instead of characters. More importantly, you can defines serie
of JAPEs where the output of one is the input of the next. JAPEs are the
easiest way for IT or academic professionals to get involved on
practical tasks. However, there are much more powerful tools in the
arsenal of KIM and GATE.
- Once you have created component(s) that create the necessary
annotations for the new predicate, you can integrate them with KIM.
- See the Developer's Guide for information on how to add facts to the
knowledge base with Java code.
- Start bin/startKIMGate.sh in the KIM installation. It will load the
GATE development environment with the KIM components. By that time, you
should be pretty familiar with the functioning of GATE. Add the newly
created ones and save the pipeline.
4. Evaluate the results.
- Does the new pipeline finds the acquisition every time? 50% of the
cases? Does your code find fake cases of acquisition?
- Don't be disappointed if doesn't. Compare your effort to the value
the upgraded information extraction delivers to you or your clients.
Usually, it is possible for automation information extraction to deliver
70% accuracy. Is it enough?
Hopefully, this brief introduction to natural language processing will
set you on the right path.
Best regards,
Marin Nozhchev
Technical Lead of the KIM Team
On 04/09/2009 04:50 PM, Anton Andreev wrote:
Hi,
I want to be able query for a triple like: Company A acquires/buys
company B.
How should I create this relation so the graph is first created and
then later I be able to query for A or B companies.
Where do I start?
I have to somehow create a predicate "isAcquired"?
Cheers,
Anton
------------------------------------------------------------------------
_______________________________________________
Kim-discussion mailing list
[email protected]
http://ontotext.com/mailman/listinfo/kim-discussion
_______________________________________________
Kim-discussion mailing list
[email protected]
http://ontotext.com/mailman/listinfo/kim-discussion