vectorspacemodel

• Install the npm package

A strongly-typed, light-weight library created in TypeScript that allows for an object-orientated, high-level, efficient approach to Vector Space Model computations.

Installation

npm install vectorspacemodel --save

Features

• VectorOperations, which contains useful vector calculation functions such as dotProduct and euclideanLength.
• SimilaritySchemas, which contains useful functions to compute relevance between two vectorized documents, such as cosineSimilarity.
• WordMappings, which contains commonly used word-equivalence functions such as caseInsensitive and noPunctuation.
• Every mathematical function is completely customizable, and every function type is strongly typed, well documented, and has existing examples in the aforementioned libraries. This allows for a general-use, object-orientated approach to Vector Space Modeling.

Usage

import { VectorSpaceModel, SimilaritySchemas, WeighingSchemas, WordMappings } from 'vectorspacemodel';
const vsm = new VectorSpaceModel(SimilaritySchemas.cosineSimilarity, WeighingSchemas.tfidf, WordMappings.caseInsensitive);
const docs = [
    {
        content: "This is test numero uno.",
        meta: /* Some map representing metadeta */,
    },
    {
        content: "This document will be the most relevant.",
        meta: /* Metadata is optional! */,
    },
    {
        content: "Ordered last in the results.",
        meta: /* metadata can be useful for important document-tracking purposes. */,
    },
    /** More documents... */
];
const queryResult = vsm.query(docs, "which document is the most relevant.", 3);

Example

const vsm = new VectorSpaceModel(SimilaritySchemas.cosineSimilarity, WeighingSchemas.tfidf, WordMappings.caseInsensitive);
const docs = [
    {
        content: "This is test numero uno.",
        meta: new Map([["a", 12]]),
    },
    {
        content: "This document will be the most relevant.",
        meta: new Map([["c", 12]]),
    },
    {
        content: "Ordered last in the results.",
        meta: new Map([["d", 12]]),
    },
    /** More documents... */
];
const queryResult = vsm.query("which document is the most relevant.", docs, 3);
console.log(queryResult);

Output

[
  VectorizedDocument {
    _vector: Map(16) {
      'this' => 0.6931471805599453,
      'document' => 0.6931471805599453,
      'will' => 1.3862943611198906,
      'be' => 1.3862943611198906,
      'the' => 0.28768207245178085,
      'most' => 0.6931471805599453,
      'relevant' => 0.6931471805599453,
      'is' => 0,
      'test' => 0,
      'numero' => 0,
      'uno' => 0,
      'ordered' => 0,
      'last' => 0,
      'in' => 0,
      'results' => 0,
      'which' => 0
    },
    _meta: Map(1) { 'c' => 12 },
    _content: 'This document will be the most relevant.',
    _score: 0.3180619512803787
  },
  VectorizedDocument {
    _vector: Map(16) {
      'this' => 0.6931471805599453,
      'is' => 0.6931471805599453,
      'test' => 1.3862943611198906,
      'numero' => 1.3862943611198906,
      'uno' => 1.3862943611198906,
      'document' => 0,
      'will' => 0,
      'be' => 0,
      'the' => 0,
      'most' => 0,
      'relevant' => 0,
      'ordered' => 0,
      'last' => 0,
      'in' => 0,
      'results' => 0,
      'which' => 0
    },
    _meta: Map(1) { 'a' => 12 },
    _content: 'This is test numero uno.',
    _score: 0.09348996506292016
  },
  VectorizedDocument {
    _vector: Map(16) {
      'ordered' => 1.3862943611198906,
      'last' => 1.3862943611198906,
      'in' => 1.3862943611198906,
      'the' => 0.28768207245178085,
      'results' => 1.3862943611198906,
      'this' => 0,
      'is' => 0,
      'test' => 0,
      'numero' => 0,
      'uno' => 0,
      'document' => 0,
      'will' => 0,
      'be' => 0,
      'most' => 0,
      'relevant' => 0,
      'which' => 0
    },
    _meta: Map(1) { 'd' => 12 },
    _content: 'Ordered last in the results.',
    _score: 0.014983676405859226
  }
]

Learn the Theory

Want to learn the theory behind vector space modeling, or information retrieval in general? Sources for further reading:

• https://nlp.stanford.edu/IR-book/information-retrieval-book.html
• https://www.sciencedirect.com/topics/computer-science/vector-space-models
• https://blog.christianperone.com/2013/09/machine-learning-cosine-similarity-for-vector-space-models-part-iii/

Contributions

Want to contribute to this project? Feel free to fork! I'm happy to make further additions.

An Apology

Yes, the names of classes, interfaces, and functions are very long at times, but often times, abbreviations cause for confusions, ambiguity, and difficulty reading code. I am always open to new suggestions for new naming schema, so please open an issue if you have concerns.

TODO

• Add more general-functionality functions to the following libraries:
  • VectorOperations
  • SimilaritySchemas
  • WordMappings
• Find different representations of vectors than Map<string, number> in order to give the user more options.