modnlp.tc.classify

Class BVBayes

java.lang.Object

modnlp.tc.classify.BVBayes

public class BVBayes
extends java.lang.Object

Naive Bayes classifier for multivariate Bernoulli models (documents represented as Boolean vectors). This class uses the CSV described in (Sebastiani, 2002). CSV values are not actual probabilities but the function is monotonically increasing on the estimated probability function. Categorise each news item in corpus_list according to categ using a Multivariate Bernoulli model and Naive Bayes. USAGE:

   BVBayes corpus_list categ prob_model threshold [parser [smoothing]]

   SYNOPSIS:"
     Categorise each news item in corpus_list according to categ using
     Boolean Vector Naive Bayes (see lecture notes ctinduction.pdf, p 7)

   ARGUMENTS:
    corpus_list: list of files to be classified
    categ: target category (e.g. 'acq'.) The classifier will define CSV 
           as CSV_{categ}
    pmfile: file containing a  probability model generated via, say, 
            modnlp.tc.induction.MakeProbabilityModel.
    threshold: a real number (for RCut thresholding) or the name of a 
              thresholding strategy. Currently supported strategies:
              - 'proportional': choose threshold s.t. that g_Tr(ci) is
                 closest to g_Tv(ci). [DEFAULT]
             - ...
    parser: parser to be used [default: 'NewsParser']
           see modnlp.tc.parser.* for other options 

    smoothing: 0: MLE (no smoothing), 1: Laplace, ...  [default: 0]
  

See Also:

BVProbabilityModel, NewsParser

Constructor Summary

Constructors

Constructor and Description
BVBayes(java.lang.String clist, java.lang.String pmfile) Set up the main user interface items

Method Summary

Methods

Modifier and Type	Method and Description
double	computeCSV(java.lang.String cat, ParsedDocument pni) CSV_i(d_j) = \sum_0^T tkj log p(t|c) * (1 - p(t|?c) / p(t|?c) * (1 - p(t|c) (where tkj \in {0, 1} is the binary weight at position k in vector d_j; multiplying by it causes terms that do not occur in the document to be ignored) Disadvantage of this approach: larger documents receive disproportionally large CSV's simply by virtue of containing more terms.
static void	main(java.lang.String[] args)
ParsedCorpus	parse(java.lang.String filename, java.lang.String plugin) pars: Set up a new parser object of type plugin, perform parsing, and return a ParsedCorpus

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

BVBayes

public BVBayes(java.lang.String clist,
       java.lang.String pmfile)

Set up the main user interface items

Method Detail

parse

public ParsedCorpus parse(java.lang.String filename,
                 java.lang.String plugin)
                   throws java.lang.Exception

pars: Set up a new parser object of type plugin, perform parsing, and return a ParsedCorpus

Parameters:

filename - the file to be parsed

plugin - the parser class name

Returns:

a ParsedCorpus

Throws:

java.lang.Exception - if an error occurs

computeCSV

public double computeCSV(java.lang.String cat,
                ParsedDocument pni)

CSV_i(d_j) = \sum_0^T tkj log p(t|c) * (1 - p(t|?c) / p(t|?c) * (1 - p(t|c) (where tkj \in {0, 1} is the binary weight at position k in vector d_j; multiplying by it causes terms that do not occur in the document to be ignored) Disadvantage of this approach: larger documents receive disproportionally large CSV's simply by virtue of containing more terms. Document lenght normalisation could be employed here to correct this distortion. A simpler approach would be to base the probability model on term events rather than document events, as described in (Mitchell 97, ch 6).

main

public static void main(java.lang.String[] args)