NeuralPhraseX

Takes a javascript object that contains a set of possible pattern matches, phrases and wildcards. These phrases are converted to sentence embeddings in tensorflow. A search phrase is passed to Phrasex and an object is returned containing a list of close phrases as well is the slots that are filled in. The algorithm relies on npm modules, neural-sentence-search, slot-filler and sentence-similarity. Each "match" is scored based on how well it matched. This can be used directly in processing text in a chatbot or a one shot information extraction pipeline.

Install

npm install neural-phrasex

How to use

let {Phrasex, UserData, BasicPhrasexDatabase} = require("neural-phrasex");

let simpleDatabase = {
  data: [
    {
      //The simplest entry has just a phrase and a class name - phraseType
      phrase: ["It is what it is."],
      phraseType: "isWhatIs"
    },
    {
      //A phrase and a response to a phrase can be included.
      phrase: ["what is your name?"],
      response: ["My name is Bot"],
      phraseType: "whatIsName",
      implies: ["whatIsName"]
    }, {
      //You can include just simple phrases with no wildcards
      exampleWildcards: { value: ["pig"], ans: ["animal"] },
      phrase: ["What is a (value)?"],
      response: ["A (value) is an (ans)"],
      phraseType: "whatIsThing",
    }, {
      //The wildcards are just examples of what could be put in the slots
      //These are necessary so that they can be passed to the neural network
      //to construct a sentence vector.
      exampleWildcards: { value: ["Seattle"], ans: ["Washington"] },
      phrase: ["where is (value)"],
      response: ["(value) is in (ans)"],
      phraseType: "whereIsThing"
    }
  ]
}

let compute = async () => {
  let ans = BasicPhrasexDatabase.generatePhraseDatabase(simpleDatabase)
  let phrasex = new Phrasex(ans)
  let res = await phrasex.initialize()

  let userData = new UserData();
  userData.initialize();

  let res1 = await phrasex.getWildcardsAndMatch("Where is Boston", "", userData)
  console.log(res1[0])

  let res2 = await phrasex.getWildcardsAndMatch("What is a coconut", "", userData)
  console.log(res2[0])
})

compute()

with result

    {
      source: {
        exampleWildcards: { value: [Array], ans: [Array] },
        phrase: 'where is (value)',
        response: [ '(value) is in (ans)' ],
        phraseType: 'whereIsThing',
        implies: [ 'whereIsThing', 'whereIsThing' ],
        meta: { groupInex: 6 },
        example: 'where is Seattle',
        storage: null,
        words: 'where is'
      },
      wildcards: { matched: true, value: 'Boston' },
      confidence: 1,
      wcScore: { score: 1, count: 1 },
      score: {
        queryIndex: [ [Object], [Object], [Object] ],
        score: 2,
        order: 1,
        size: 0.5,
        semantic: 0.580953918415687
      }
    }


    {
      source: {
        exampleWildcards: { value: [Array], ans: [Array] },
        phrase: 'What is a (value)?',
        response: [ 'A (value) is an (ans).' ],
        phraseType: 'whatIsThing',
        implies: [ 'whatIsThing', 'whatIsThing' ],
        meta: { groupInex: 4 },
        example: 'What is a pig',
        storage: null,
        words: 'What is a ?'
      },
      wildcards: { matched: true, value: 'coconut' },
      confidence: 1,
      wcScore: { score: 1, count: 1 },
      score: {
        queryIndex: [ [Object], [Object], [Object], [Object] ],
        score: 3,
        order: 1,
        size: 0.3333333333333333,
        semantic: 0.5399621217119523
      }
    }

The result is a variation of the original database source plus the wildcards for filling in the data.