Home

Awesome

sentiment

AFINN-based sentiment analysis for Node.js

CircleCI Greenkeeper badge

Sentiment is a Node.js module that uses the AFINN-165 wordlist and Emoji Sentiment Ranking to perform sentiment analysis on arbitrary blocks of input text. Sentiment provides several things:

Table of contents

Installation

npm install sentiment

Usage example

var Sentiment = require('sentiment');
var sentiment = new Sentiment();
var result = sentiment.analyze('Cats are stupid.');
console.dir(result);    // Score: -2, Comparative: -0.666

Adding new languages

You can add support for a new language by registering it using the registerLanguage method:

var frLanguage = {
  labels: { 'stupide': -2 }
};
sentiment.registerLanguage('fr', frLanguage);

var result = sentiment.analyze('Le chat est stupide.', { language: 'fr' });
console.dir(result);    // Score: -2, Comparative: -0.5

You can also define custom scoring strategies to handle things like negation and emphasis on a per-language basis:

var frLanguage = {
  labels: { 'stupide': -2 },
  scoringStrategy: {
    apply: function(tokens, cursor, tokenScore) {
      if (cursor > 0) {
        var prevtoken = tokens[cursor - 1];
        if (prevtoken === 'pas') {
          tokenScore = -tokenScore;
        }
      }
      return tokenScore;
    }
  }
};
sentiment.registerLanguage('fr', frLanguage);

var result = sentiment.analyze('Le chat n\'est pas stupide', { language: 'fr' });
console.dir(result);    // Score: 2, Comparative: 0.4

Adding and overwriting words

You can append and/or overwrite values from AFINN by simply injecting key/value pairs into a sentiment method call:

var options = {
  extras: {
    'cats': 5,
    'amazing': 2
  }
};
var result = sentiment.analyze('Cats are totally amazing!', options);
console.dir(result);    // Score: 7, Comparative: 1.75

API Reference

var sentiment = new Sentiment([options])

ArgumentTypeRequiredDescription
optionsobjectfalseConfiguration options (no options supported currently)

sentiment.analyze(phrase, [options], [callback])

ArgumentTypeRequiredDescription
phrasestringtrueInput phrase to analyze
optionsobjectfalseOptions (see below)
callbackfunctionfalseIf specified, the result is returned using this callback function

options object properties:

PropertyTypeDefaultDescription
languagestring'en'Language to use for sentiment analysis
extrasobject{}Set of labels and their associated values to add or overwrite

sentiment.registerLanguage(languageCode, language)

ArgumentTypeRequiredDescription
languageCodestringtrueInternational two-digit code for the language to add
languageobjecttrueLanguage module (see Adding new languages)

How it works

AFINN

AFINN is a list of words rated for valence with an integer between minus five (negative) and plus five (positive). Sentiment analysis is performed by cross-checking the string tokens (words, emojis) with the AFINN list and getting their respective scores. The comparative score is simply: sum of each token / number of tokens. So for example let's take the following:

I love cats, but I am allergic to them.

That string results in the following:

{
    score: 1,
    comparative: 0.1111111111111111,
    calculation: [ { allergic: -2 }, { love: 3 } ],
    tokens: [
        'i',
        'love',
        'cats',
        'but',
        'i',
        'am',
        'allergic',
        'to',
        'them'
    ],
    words: [
        'allergic',
        'love'
    ],
    positive: [
        'love'
    ],
    negative: [
        'allergic'
    ]
}

In this case, love has a value of 3, allergic has a value of -2, and the remaining tokens are neutral with a value of 0. Because the string has 9 tokens the resulting comparative score looks like: (3 + -2) / 9 = 0.111111111

This approach leaves you with a mid-point of 0 and the upper and lower bounds are constrained to positive and negative 5 respectively (the same as each token! 😸). For example, let's imagine an incredibly "positive" string with 200 tokens and where each token has an AFINN score of 5. Our resulting comparative score would look like this:

(max positive score * number of tokens) / number of tokens
(5 * 200) / 200 = 5

Tokenization

Tokenization works by splitting the lines of input string, then removing the special characters, and finally splitting it using spaces. This is used to get list of words in the string.


Benchmarks

A primary motivation for designing sentiment was performance. As such, it includes a benchmark script within the test directory that compares it against the Sentimental module which provides a nearly equivalent interface and approach. Based on these benchmarks, running on a MacBook Pro with Node v6.9.1, sentiment is nearly twice as fast as alternative implementations:

sentiment (Latest) x 861,312 ops/sec ±0.87% (89 runs sampled)
Sentimental (1.0.1) x 451,066 ops/sec ±0.99% (92 runs sampled)

To run the benchmarks yourself:

npm run test:benchmark

Validation

While the accuracy provided by AFINN is quite good considering it's computational performance (see above) there is always room for improvement. Therefore the sentiment module is open to accepting PRs which modify or amend the AFINN / Emoji datasets or implementation given that they improve accuracy and maintain similar performance characteristics. In order to establish this, we test the sentiment module against three labelled datasets provided by UCI.

To run the validation tests yourself:

npm run test:validate

Rand Accuracy

Amazon:  0.726
IMDB:    0.765
Yelp:    0.696

Testing

npm test