Title: Acronym Finder
^^^^^^^^^^^^^^^^^^^^^
Team Detail:
Abhishek Agarwal (05329022)
Saurabh P. Singh(05329037)
Ramanand J.(05329402)
***************************
Abstract:
^^^^^^^^
In this project, we designed an Acronym Finder using a freely available CRF package[1]. 
Details about the package can be found at Source Forge. In this features based approach, we have tried to capture various positional, regular expression based features. We have complete the design and implementation of the tool, but currently we are facing the problem of lack of datasets, because of which we cant comment on the accuracy.

CRF package: It tries to classify the data sequences based on various features. The training, as well as testing part comes with the package. We were left with two major task:
	1) To format the datasets, and make an in-memory representation of the data
sequences, and provide the interfaces which package expects.
	2) To code features specific to Acronym Finder. Features coded till
now are:
		a) Regular Expression features, like all letters capital,
having parenthesis, etc.
		b) Positional Features: Given an Acronym, Axxx Bxxx Cxxx Dxxx(ABCD) at i th place,
the i-j the token should begin with |i| - j + 1 letter. Similarly, it checks
on the other side also. Currently it does not takes care of the stop words
like and, of etc.
	
Formatting the datasets:
^^^^^^^^^^^^^^^^^^^^^^^
We have downloaded around 100 pages form wikipedia, from different field and
then converted manually into the XML files with following tags: <ACR> and
<OTHER>. After that we designed a XML parser for that which parsed the tagged
files and stored the data in various objects, in accordance with the interface
provided by the package.

Features:
^^^^^^^^^
Other than the common features provided by the author, there were some
features specific to Acronym Finder which were designed.

Accuracy:
^^^^^^^^
Currently, we are facing the problem of lack of tagged data sets, as it
requires a manual input, process is very slow. Therefore the results currently
observed are not the right reflections of the accuracies: Confusion matrix
once observed is as follows:

		ACTUAL OTHER	 ACTUAL ACRONYM
OBS OTHER         100			0
OBS ACR		  50		 	80
	
This clearly shows that we were having high recall value but low precision. We
were also able to find the reason for following observation. Lack of training
data resulted in higher weights for some negative features, like word with
initial letter Capital, but because of lack of negative example it gained high
weight. 
We have discussed with the author and we feel that if we increase the data set
to a reasonable amount, we can achieve pretty good accuracies.

******************************************************************************