================================================================
Comparative Study of Learning Approaches for Sequential Labeling:
A Case Study on Syntactic Chunking
================================================================

GOAL:
The goal of this work is to compare several learning
architectures for sequential learning on the concrete NLP task of
syntactic base chunking. The problem datasets and feature types
will be fixed and the comparative study will be performed among
learning approaches at several levels: accuracy on the task,
efficiency (training and test), parameter tuning (stability),
sensitivity to the features, etc.

SETTING 
The task setting will be fixed to that of the CoNLL-2000 shared task
on chunking. Please, visit the website:

http://www.cnts.ua.ac.be/conll2000/chunking 

to obtain the datasets, the problem description, and a large
state-of-the-art bibliography from which you can borrow the set of
standard features commonly used in the task.

COMPARATIVE STUDY
Regarding the learning approaches to compare, you can consider,
in increasing level of complexity:

* A simple HMM-based tagger (e.g., use TnT with the lexical
  specialization described at http://www.dsic.upv.es/~fpla/demo.html)

* A machine learning sequential tagger based on local classifiers
  trained for each of the labels. Here, we can diferentiate between
  greedy local inference with/without left chaining and Viterbi style
  decoding for optimizing the probability of the whole sequence. You
  may use: SVMtool (for fast linear SVM training/decoding), YamCha (if
  you want to use kernels), Maximum Entropy models (MEMM), which can
  be found in the MALLET suite.

* A global learning algorithm for sequential labeling. For instance
  you can use Conditional Random Fields (find them at MALLET) or
  SVMstruct if you want to apply a large margin 

Other approaches suggested by students can also be considered, but
first you have to have your proposal accepted by the course advisor.

Links to all the software mentioned above are posted at the course
webpage. You have to try a minimum of three approaches (one per
student). You can work independently, but make sure to share exactly
the same setting and then to make an integrated comparison and
presentation. The level of difficulty may vary greatly depending on
your choices. Please, contact L. Màrquez before starting your work.