sentence-similarity

SentenceSimilarity

Javascript/Nodejs sentence similarity.

This algorithm computes 4 sentence similarity scores that can be used for choosing best matched scores based on multiple criteria. The 4 similarities are 'exact' which is the number of words that match exactly regardless of order. 'score' which is the score of the words that match either partially or exactly. Partial matches are determined by the a user selected word similarity measure, Levenshtein or Metaphone for example. 'order' is score provided based on the order of the words in the two sentences. 'size' is the score determined by comparing the number of words to match to the number of words that exist in the compared sentence. The matches between words are unique so no word in sentence 1 is matched with more than one word in sentence 2. If the same word exists twice in sentence 2, but not sentence 1 then preference is given to the left most word during matching.

By multiplying score*order*size or exact*order*size, one will get a score between 0 and 1. In addition, the user can supply wildcards that appear in parentheses 'Your name is (name)' where the wildcard '(name)' is ignored in the comparison. This allows the similarity score to be used in slot filling.

Example 1

"use strict"

let similarity = require('sentence-similarity')
let similarityScore = require('similarity-score')

let s1 = ['how','close','is','this','to','that']
let s2 = ['these','two','are','not','that','close']

let winkOpts = { f: similarityScore.winklerMetaphone, options : {threshold: 0} }

console.log(similarity(s1,s2,winkOpts))

gives

{ matched: [ 3, 5, -1, 0, 1, 4 ],
  matchScore: 
   [ 0.27777777777777773,
     1,
     0,
     0.9133333333333333,
     0.32499999999999996,
     1 ],
  exact: 2,
  score: 3.516111111111111,
  order: 0.06666666666666687,
  size: 0.16666666666666666 }

"matched" tells you which index of s2 best matches the corresponding word in s1. And matchScore give you the corresponding score for the matches determined in "matched". In this case, the algorithm suggests that

1. 'how=not' with score (0.277)
2. 'close=close' with score (1)
3. 'is=' doesn't have a match, score(0)
4. 'this=these' with score (0.91)
5. 'to=two' with score (0.325)
6. 'that=that' with score (1.0)

The third parameter (opts) to the similarity(s1,s2,opts) call is an object containing the similarity measure for comparing words in a sentence as well as a option object which can be used to define additional parameters. someFunc should return a value between 0 and 1 defining how similar the two "words" are, where 1 means they are exactly the same and 0 means the are not at all the same. I often use a threshold parameter that sets the similarity to 0 if the similarity is below that threshold.

let opts = { f: someFunc, options : {threshold: 0.1} }

where

let someFunc = function(a,b,options) {
  let a = levenshtein(a,b);//for example
  if(a<options.threshold) a = 0;
  return a
}

several example similarity functions are provided in SimilarityScore.js.

Example 2 (wildcard)

Wildcards are placed between parentheses and allow the location of slots and slot values to be identified. In the following example (name) is a wildcard and "ripley" is the slot value.

"use strict"

let ss = require('sentence-similarity')

let similarity = ss.sentenceSimilarity;
let similarityScore = ss.similarityScore;

let s1 = ['My','name','is','ripley']
let s2 = ['My','name','is','(name)']

let winkOpts = { f: similarityScore.winklerMetaphone, options : {threshold: 0} }

console.log(similarity(s1,s2,winkOpts))

outputs

{ matched: [ 0, 1, 2, -1 ],
  matchScore: [ 1, 1, 1, 0 ],
  exact: 3,
  score: 3,
  order: 1,
  size: 0.3333333333333333 }

where

matched: [ 0, 1, 2, -1 ]

indicates the index in array s2 that matches the given index in array s1. Index 4 in s1 does not match anything in s2, so it may be a slot value. We've developed a slot filler that uses this sentence similarity and it is in the npm module slot-filler.