LDA

Latent Dirichlet Allocation via collapsed Gibbs sampling.

Usage

var lda = require( '@stdlib/nlp/lda' );

lda( docs, K[, options] )

Latent Dirichlet Allocation via collapsed Gibbs sampling. To create a model, call the lda function by passing it an array of strings and the number of topics K that should be identified.

var model;
var docs;

docs = [
    'I loved you first',
    'For one is both and both are one in love',
    'You never see my pain',
    'My love is such that rivers cannot quench',
    'See a lot of pain, a lot of tears'
];

model = lda( docs, 2 );
// returns {}

After initialization, model parameters are estimated by calling the .fit() method, which performs collapsed Gibbs sampling.

The model object contains the following methods:

model.fit( iter, burnin, thin )

model.fit( 1000, 100, 10 );

The iter parameter denotes the number of sampling iterations. While a common choice, one thousand iterations might not always be appropriate. Empirical diagnostics can be used to assess whether the constructed Markov Chain has converged. burnin denotes the number of estimates that are thrown away at the beginning, whereas thin controls the number of estimates discarded in-between iterations.

model.getTerms( k[, no = 10] )

Returns the no terms with the highest probabilities for chosen topic k.

var words = getTerms( 0, 3 );
/* returns
    [
        { word: 'both', prob: 0.06315008476532499 },
        { word: 'pain', prob: 0.05515729517235543 },
        { word: 'one', prob: 0.05486669737616135 }
    ]
*/

Examples

var getSpeeches = require( '@stdlib/datasets/sotu' );
var roundn = require( '@stdlib/math/base/special/roundn' );
var STOPWORDS = require( '@stdlib/datasets/stopwords-en' );
var lda = require( '@stdlib/nlp/lda' );

var speeches;
var terms;
var model;
var str;
var i;

function getText( e ) {
    str = e.text.toLowerCase();
    STOPWORDS.forEach( remove );
    return str;

    function remove( word ) {
        var RE = new RegExp( '\\b' + word + '\\b', 'gi' );
        str = str.replace( RE, '' );
    }
}

speeches = getSpeeches({
    'range': [ 1930, 2010 ]
}).map( getText );


model = lda( speeches, 3 );

model.fit( 1000, 100, 10 );

for ( i = 0; i <= 80; i++ ) {
    str = 'Year: ' + (1930+i) + '\t';
    str += 'Topic 1: ' + roundn( model.avgTheta.get( i, 0 ), -3 ) + '\t';
    str += 'Topic 2: ' + roundn( model.avgTheta.get( i, 1 ), -3 ) + '\t';
    str += 'Topic 3: ' + roundn( model.avgTheta.get( i, 2 ), -3 );
    console.log( str );
}

terms = model.getTerms( 0, 20 );
for ( i = 0; i < terms.length; i++ ) {
    terms[ i ] = terms[ i ].word;
}
console.log( 'Words most associated with first topic:\n ' + terms.join( ', ' ) );

terms = model.getTerms( 1, 20 );
for ( i = 0; i < terms.length; i++ ) {
    terms[ i ] = terms[ i ].word;
}
console.log( 'Words most associated with second topic:\n ' + terms.join( ', ' ) );

terms = model.getTerms( 2, 20 );
for ( i = 0; i < terms.length; i++ ) {
    terms[ i ] = terms[ i ].word;
}
console.log( 'Words most associated with third topic:\n ' + terms.join( ', ' ) );