================================================================
[3] Study of different feature sets to learn perceptrons useful
    for extracting the ACE mentions of relations
================================================================

One of the ACE-2007 subtask related to Relation Extraction consists in
the recognition and classification of relation mentions (RCR) within
the ACE documents. A relation mention in ACE is a sentence or phrase
that expresses the relation. Both entities involved in the relation
must be mentioned in the sentence. For instance, the sentence 'Peter
and his daughter, Jane, will go to the cinema' is a relation mention
of the type 'family' involving two entity mentions 'Peter' and
'Jane'. In fact, a relation mention can be represented as the pair of
mentions of entities (<Peter,Jane> in the example).

(Read the ACE07-evaluation-information available in
http://www.nist.gov/speech/tests/ace/ace07/doc/ace07-evalplan.v1.3a.pdf
for detailed information on the types and subtypes of relation
mentions and on the evaluation metrics).

The goal of this work consists in comparing different feature sets and
possibly different variations of the perceptron learning algorithm, if
necessary, to learn a multiclassifier of relation mentions based on
SVM. We will assume that the NEs are already anotated (strong and weak
NEs). So, the ACE gold set of Named Entities will be used, and the
task is reduced to classify pairs of NE mentions occurring in the same
sentence into an ACE relation type:sybtype.


Available resources and NLP tools for feature generation:

- Preprocessed documents (just some linguistic attributes -word, POS,
  syntactic chunck-)
- A perceptron learning algorithm
- A procedure to generate the positive and negative examples for the
  learning algorithm
- SwiRL: A Semantic Role Labeling system (available at:
  http://www.lsi.upc.edu/~surdeanu)
- MALT: A dependency parser (available at:
  http://w3.msi.vxu.se/~nivre/research/MaltParser.html)

Please contact to Jordi Turmo before starting the work and to get all
the above resources.
================================================================