Word Sense Disambiguity methods of Senseval-3 (2004) and a novel algorithm.

Introduction:
  Polysemy, lexical ambiguity is the property of some words to have multiple meanings or senses.
The definition and the number of senses is a function of the dictionary used.  This problem is 
commonly referred to as Word Sense Disambiguity (WSD).  This is a difficult problem that is receiving
a great deal of attention from the research community  because its solutions can help other NLP 
applications such as Machine Translation (MT), Information Retrieval (IT), Text Processing, 
Grammatical Analysis, Information Extraction (IE), hypertext navigation, and so on.
  Grading of the sytems is figured by comparing the system results against the ones found by the 
linguistic annotators abitrated answers.  All senses used were from the WordNet 1.7.1 (from 
Princeton) lexical dictionary database, except for verbs in the English Lexical Task, which uses 
the WordSmyth dictionary.
  The fourth Senseval conference help in Barcelona, Spain, brought together 27 groups of researchers
from around the world.  They used a wide variety of algorithms and techniques for disambiguation 
with a good set of results.  The accuracy of the systems are graded using precision, recall, and accuracy.  
The measures that we chose for the evaluation of our methods are those typically used in the language
engineering and machine learning literature: recall,precision and accuracy. The recall measure counts 
the number of words that are assigned the correct sense, out of the total number of words to be 
assigned a sense. This corresponds to the ratio of true positive examples to the total number of 
positives in the test data. On the other hand, precision counts the number of words assigned the
correct sense, out of the number of word-senses considered positive by the decision tree, i.e., 
the ratio of true positive to true and false positive examples. In addition to these two measures 
the percentage correct classification (accuracy), which is a standard measure for machine learning 
methods is used. In summary the three ratios:
  recall = TP/P,
  precision = TP/(TP+FP),
  accuracy = (TP+TN)/(P+N),
where TP/FP and TN/FN stand for True/False Positive and
True/False Negative and P/N for Positive/Negative
examples.
  Two specific tasks for the English language were attempted, as well as many other languages.
The English All-Word Task is a test data from text of two Wall Street Journal articles, and one 
excerpt form the Brown Corpus.  These cover three seperate domains, editorial, news story, and 
fiction.  The English All-Words task has a total of 2,041 words of 5,000 runnings words.  The second
is the English Lexical Sample Task (LS), with 3944 examples extracted from the British National 
Corpus (BNC) and training data tagged with the Open Mind Word Expert system.  All the words were 
tagged with part of speech (POS) tags by Penn Treebank.  
  The baseline results of the two systems are figured by simply choosing the most popular sense for 
the word.  A baseline of 60.9% (KOC Univ), and 55.2% respectively were found.  Many of the 
systems submitted received well above this achieving a 65.2% and 72.9 precision for the two systems.  
  Four of the systems were studied in detail and used a number of techniques for WSD.  Language 
Computer Corporation (LCC) of Richardson, TX, University of Alicante in Spain (R2D2), Koc University of 
Turkey, and University of Antwerp did well on their task using and combining various algorithms.  
 These systems will be compared with an in-house set of software created by Midwestern Alumni James 
Ratcliff.
  A large training data set was supplied for the second task, and many supervised and unsupervised 
systems were trained on this.  The methods include Heuristic rules, Machine Learning, Incremental 
Disambiguation, Maximum Entropy, Specialized Hidden Markov Model, Relevant Domains, Neural Networks, 
Vector Space Model, Naive Bayes, Bag of Words, Local Collocation, and Genetic Algorithms.  All of 
these algorithms take advantage of as much information as possible, usually relying on the training 
data provided by Senseval-4 and previous sessions, WordNet glossaries (glosses-definitions) Synsets 
(synonyms and similar words), and other corpora.  WordNet provides a number of relationships such as 
hypernyms (is a relations), hyponyms, coordinate terms, domain classification, antonyms, and 
frequencies.  Context, collocations, and bags of words (set of non-stop words located near the term) 
often give disambiguation evidence.  A glance of a number of the terms quickly determines that any 
one of these may be the deciding factor in determining the sense of the word.  Determining which of
these is the most important is the hard task that Senseval and natural language processing word sense 
disambiguity is concerned with.