English NLP for Node.js and the browser. Try it in your browser!
35k gzipped, Part-of-Speech tagging (92% on Penn treebank), entity recognition, sentiment analysis and more, MIT licensed.
Install it with bower:
bower install --save compendium
Or clone this repo and copy the dist/compendium.minimal.js file into your project.
<script type="text/javascript"
src="path/to/compendium/dist/compendium.minimal.js"></script>
In order to ensure that Compendium will work as intended, you must specify the encoding of the HTML page as UTF-8.
Call the compendium.analyse function with a string as parameter, and get a complete analysis of the text.
console.log( compendium.analyse('Hello world :)') );
npm install --save compendium-js
var compendium = require('compendium-js');
console.log(compendium.analyse('Hello world :)'));
The main function to call is analyse.
It takes a string as unique argument, and returns an array containing an analysis for each sentence. For example, calling:
compendium.analyse('My name is Dr. Jekyll.');
will return an array like this one:
[ { time: 9, // Time of processing, in ms
length: 6, // Count of tokens
raw: 'My name is Dr. Jekyll .', // Raw string
stats:
{ confidence: 0.4583, // PoS tagging confidence
p_foreign: 0, // Percentage of foreign PoS tags, e.g. `FW`
p_upper: 0, // Percentage of uppercased tokens, e.g. `HELLO`
p_cap: 50, // Percentage of capitalized tokens, e.g. `Hello`
avg_length: 3 }, // Average token length
profile:
{ label: 'neutral', // Sentiment: `negative`, `neutral`, `positive`, `mixed`
sentiment: 0, // Sentiment score
amplitude: 0, // Sentiment amplitude
types: [], // Types ('tags') of sentence
politeness: 0, // Politeness score
dirtiness: 0, // Dirtiness score
negated: false }, // Is sentence mainly negated
entities: [ { // List of entities
raw: 'Dr. Jekyll', // Raw reconstructed entity
norm: 'doctor jekyll', // Normalized entity
fromIndex: 3, // Start token index
toIndex: 4, // End token index
type: null } ], // Type of entity: null for unknown, `ip`, `email`...
tags: // Array of PoS tags
[ 'PRP$', 'NN', 'VBZ', 'NNP', 'NNP', '.' ],
tokens: // Tokens details
[ { raw: 'My', // Raw token
norm: 'my', // Normalized
pos: 'PRP$', // PoS tag
profile:
{ sentiment: 0, // Sentiment score
emphasis: 1, // Emphasis multiplier
negated: false, // Is negated
breakpoint: false }, // Is breakpoint
attr:
{ acronym: false, // Is acronym
plural: false, // Is plural
abbr: false, // Is an abbreviation
verb: false, // Is a verb
entity: -1 } }, // Entity index, `-1` if no entity
//
// ... Other tokens
//
] } ]
From version 0.0.26, in order to speed up the analyse, one can use the skipDetectors argument of the analyse function to skip some specific detectors.
Skippable detectors are the following:
sentiment: Sentiment analysisentities: Entity extractionnegation: Negation detectiontype: Type detection (declarative, interrogative...)numeric: Numeric values extraction
For example, the following call to analyse won't run the entity extraction detector, meaning that Dr. Jekyll won't appear in the entities section of the analysis result:
compendium.analyse('My name is Dr. Jekyll.', null, ['entities']);
See also:
Handles decoding of HTML entities (e.g. & to &), and normalization of some abbreviations that involve breakpoints chars (e.g. w/ to with).
No good part-of-speech tagging is possible without a good lexer. A lot of efforts has been put into the Compendium's lexer, so it provides the right tokens to be processed. Currently the lexer is a combination of four passes:
- A first pass splits the text into sentences
- A second one applies some regular expressions to extract specific parts of the sentences (URLs, emails, emoticons...)
- The third pass is a char by char parser that splits tokens in a sentence, relying on Punycode.js to properly handle emojis
- The final pass consolidates tokens such as acronyms, abbreviations, contractions..., and handles a few exceptions
This very little piece runs after the lexer, and is in charge to normalize a few other slangs (e.g. gr8 to great).
Tagging is performed using a Brill tagger (i.e. a base lexicon and a set of rules), with the addition of some inflection-based rules.
It's been inspired by the following projects that are worth being checked out:
- Eric Brill tagger: latest implementation published under MIT license is available for download on the Plymouth University website at this link (direct download).
- Mark Watson's FastTag Java library, a very simple implementation of the Brill's tagger.
- NLP Compromise, another great JS NLP toolkit, with an interesting inflection-based approach
PoS tagging is tested a set of unit tests generated with the Stanford PoS tagger, double checked with common sense and another machine-learning oriented tagger, and is then evaluated using the Penn Treebank dataset.
In September 2015, Compendium PoS tagging score on Penn Treebank was 92.76% tags recognized for the browser version, and 94.31% for the Node.js version.
Warning: the following process has been proved hardly extensible, and isn't powerful enough given the amount of code already. It's being replaced in v1.0 by another one currently in development [September 5th, 2015].
Dependency parsing module. Still experimental, and requires a lot of additional rules, but promising.
Inspired in some extent by Syntex from Didier Bourigault ref. (fr).
Constraint based. Constraints are:
- The governor is the head of the sentence (it doesnt have a master)
- When possible, the governor is the first conjugated verb of the sentence
- All other tokens must have a master
- A token can have one and only one master
- A master can have one or many dependencies
- If no master is found for a token, then its master is the governor
Parsing is done through several passes:
- First pass define direct dependencies from left to right
- Second pass define direct dependencies from right to left
- Third pass consolidate linked indirect dependencies using existing masters
- Final pass consolidate unlinked indirect dependencies
Starting from here, some detectors handle further analysis of the text. They're in charge to add some metadata to the analysis, such as the sentiment score and label.
These detectors can work at three different levels:
- the token level
- the sentence level
- the text (global) level
Token level detectors add attributes to each token (sentiment and emphasis scores, normalized token...).
Sentence level detectors work accross many tokens (negation detection, entity recognition, sentiment analysis...).
Global level detectors (there are none yet) are supposed to provide a global analysis of the whole text: topics, global sentiment labelling...
The full lexicon for Node.js is based on the lexicon from Mark Watson's FastTag (around 90 000 terms, itself being imported from the Penn Treebank).
The minimal lexicon for the browser contains only a few thousands terms extracted from the full lexicon, and filtered using:
- the list of the 10000 most common English words, an extract from the Google's Trillion Word Corpus
- the list of scored sentiments words
- Compendium suffixes detector
- Compendium embedded knowledge
Here is the list of Part-of-Speech tags used by Compendium. See at the bottom newly introduced tags.
, Comma ,
: Mid-sent punct. : ;
. Sent-final punct . ! ?
" quote "
( Left paren (
) Right paren )
# Pound sign #
CC Coord Conjuncn and,but,or
CD Cardinal number one,two,1,2
DT Determiner the,some
EX Existential there there
FW Foreign Word mon dieu
IN Preposition of,in,by
JJ Adjective big
JJR Adj., comparative bigger
JJS Adj., superlative biggest
LS List item marker 1,One
MD Modal can,should
NN Noun, sing. or mass dog
NNP Proper noun, sing. Edinburgh
NNPS Proper noun, plural Smiths
NNS Noun, plural dogs
PDT Predeterminer all, both
POS Possessive ending 's
PP Personal pronoun I,you,she
PRP$ Possessive pronoun my,one's
RB Adverb quickly, not
RBR Adverb, comparative faster
RBS Adverb, superlative fastest
RP Particle up,off
SYM Symbol +,%,&
TO 'to' to
UH Interjection oh, oops
VB verb, base form eat
VBD verb, past tense ate
VBG verb, gerund eating
VBN verb, past part eaten
VBP Verb, present eat
VBZ Verb, present eats
WDT Wh-determiner which,that
WP Wh pronoun who,what
WP$ Possessive-Wh whose
WRB Wh-adverb how,where
Compendium also includes the following new tag:
EM Emoticon :) :( :/
Go to the wiki to get more details about the project.
The MIT License (MIT)
Copyright (c) 2015 Ulflander
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.