feat(*): add test cases for vector cosine similarity

Author: sanjayaksaxena

test/as-specs.js

@@ -38,13 +38,15 @@ var winkNLP = require( '../src/wink-nlp.js' );
 var its = require( '../src/its.js' );
 var as = require( '../src/as.js' );
 var model = require( './test-model/model.js' );
+var wordVectors = require( './test-model/languages/cur/models/test-vectors.json' );
+var similarity = require( '../utilities/similarity.js' );
 
 var expect = chai.expect;
 var describe = mocha.describe;
 var it = mocha.it;
 
 describe( 'its functions for .out()', function () {
-  var nlp = winkNLP( model );
+  var nlp = winkNLP( model, null, wordVectors );
 
   it( 'as.bow', function () {
     expect( nlp.readDoc( 'to be or not to be' ).tokens().out( its.value, as.bow ) ).to.deep.equal( { to: 2, be: 2, or: 1, not: 1 } );
@@ -73,9 +75,35 @@ describe( 'its functions for .out()', function () {
   } );
 
   it( 'as.vector', function () {
-    var doc = nlp.readDoc( 'the dog ran away' );
+    const sent = 'the dog ran';
+    const doc = nlp.readDoc( sent );
 
+    // Check exceptions when as.vector is combined with incompitable its helpers.
     expect( doc.tokens().out.bind( null, its.negationFlag, as.vector ) ).to.throw( 'winkNLP: as.vector is allowed only with its value or normal or lemma.' );
     expect( doc.tokens().filter( ( t ) => t.out().length > 0 ).out.bind( null, its.negationFlag, as.vector ) ).to.throw( 'winkNLP: as.vector is allowed only with its value or normal or lemma.' );
+
+    // Test actual behaviour of as.helper by averaging the vectors of tokens
+    // and computing the l2Norm.
+    const vDog = nlp.vectorOf( 'dog' );
+    const vRan = nlp.vectorOf( 'ran' );
+    const sentVector = doc.tokens().filter( (t) => ( t.out( its.type ) === 'word' && !t.out( its.stopWordFlag ) ) ).out(its.value, as.vector );
+
+    // Note, last entry in the array is `l2Norm`, that is why subtract 1 from the length.
+    const rVector = new Array( vDog.length - 1 );
+    rVector.fill( 0 );
+    let ssr = 0;
+    for ( let k = 0; k < rVector.length; k += 1 ) {
+      rVector[ k ] = +( ( vDog[ k ] + vRan[ k ] ) / 2 ).toFixed( 8 );
+      ssr += rVector[ k ] * rVector[ k ];
+    }
+    rVector.push( +Math.sqrt( ssr ).toFixed( 8 ) );
+    expect( rVector ).to.deep.equal( sentVector );
+
+    // Also test the similarity here itself to check both as.vector & similarity.
+    const docWith2S = nlp.readDoc( 'The table was in the drawing room. The desk was in the study room.' );
+    const s0Vector = docWith2S.sentences().itemAt( 0 ).tokens().filter( (t) => ( t.out( its.type ) === 'word' && !t.out( its.stopWordFlag ) ) ).out(its.value, as.vector );
+    const s1Vector = docWith2S.sentences().itemAt( 1 ).tokens().filter( (t) => ( t.out( its.type ) === 'word' && !t.out( its.stopWordFlag ) ) ).out(its.value, as.vector );
+
+    expect( similarity.vector.cosine( s0Vector, s1Vector ) ).to.equal( 0.824834 );
   } );
 } );

test/similarity-specs.js

@@ -137,3 +137,5 @@ describe( 'set-oo normal behaviour', function () {
     } );
   } );
 } );
+
+// vector-cosine behaviour is tested in the as-specs.js

utilities/similarity.js

@@ -199,7 +199,7 @@ similarity.vector.cosine = function ( vectorA, vectorB ) {
   }
 
   // Use `l2Norm` directly from each vector.
-  return +( sumOfProducts / ( vectorA[ length ] * vectorB[ length ] ) ).toFixed( 4 );
+  return +( sumOfProducts / ( vectorA[ length ] * vectorB[ length ] ) ).toFixed( 6 );
 }; // similarity.vector.cosine()
 
 // Export similarity