This commit is contained in:
syuilo 2017-09-06 21:29:56 +09:00
parent f33571f2f4
commit cf7b1c0c5d
5 changed files with 334 additions and 23 deletions

1
.gitignore vendored
View File

@ -3,6 +3,7 @@
/node_modules /node_modules
/built /built
/uploads /uploads
/data
npm-debug.log npm-debug.log
*.pem *.pem
run.bat run.bat

View File

@ -64,6 +64,7 @@
"@types/webpack": "3.0.10", "@types/webpack": "3.0.10",
"@types/webpack-stream": "3.2.7", "@types/webpack-stream": "3.2.7",
"@types/websocket": "0.0.34", "@types/websocket": "0.0.34",
"@types/msgpack-lite": "^0.1.5",
"chai": "4.1.2", "chai": "4.1.2",
"chai-http": "3.0.0", "chai-http": "3.0.0",
"css-loader": "0.28.7", "css-loader": "0.28.7",
@ -97,7 +98,6 @@
"accesses": "2.5.0", "accesses": "2.5.0",
"animejs": "2.0.2", "animejs": "2.0.2",
"autwh": "0.0.1", "autwh": "0.0.1",
"bayes": "0.0.7",
"bcryptjs": "2.4.3", "bcryptjs": "2.4.3",
"body-parser": "1.17.2", "body-parser": "1.17.2",
"cafy": "2.4.0", "cafy": "2.4.0",
@ -126,6 +126,7 @@
"monk": "6.0.3", "monk": "6.0.3",
"morgan": "1.8.2", "morgan": "1.8.2",
"ms": "2.0.0", "ms": "2.0.0",
"msgpack-lite": "^0.1.26",
"multer": "1.3.0", "multer": "1.3.0",
"nprogress": "0.2.0", "nprogress": "0.2.0",
"os-utils": "0.0.14", "os-utils": "0.0.14",

View File

@ -68,6 +68,9 @@ type Source = {
hook_secret: string; hook_secret: string;
username: string; username: string;
}; };
categorizer?: {
mecab_command?: string;
};
}; };
/** /**

View File

@ -1,36 +1,42 @@
import * as fs from 'fs'; import * as fs from 'fs';
const bayes = require('bayes');
const bayes = require('./naive-bayes.js');
const MeCab = require('mecab-async'); const MeCab = require('mecab-async');
import * as msgpack from 'msgpack-lite';
import Post from '../../api/models/post'; import Post from '../../api/models/post';
import config from '../../conf';
/**
* 稿稿
*/
export default class Categorizer { export default class Categorizer {
classifier: any; private classifier: any;
categorizerDbFilePath: string; private categorizerDbFilePath: string;
mecab: any; private mecab: any;
constructor(categorizerDbFilePath: string, mecabCommand: string = 'mecab -d /usr/share/mecab/dic/mecab-ipadic-neologd') { constructor() {
this.categorizerDbFilePath = categorizerDbFilePath; this.categorizerDbFilePath = `${__dirname}/../../../data/category`;
this.mecab = new MeCab(); this.mecab = new MeCab();
this.mecab.command = mecabCommand; if (config.categorizer.mecab_command) this.mecab.command = config.categorizer.mecab_command;
// BIND ----------------------------------- // BIND -----------------------------------
this.tokenizer = this.tokenizer.bind(this); this.tokenizer = this.tokenizer.bind(this);
} }
tokenizer(text: string) { private tokenizer(text: string) {
return this.mecab.wakachiSync(text); return this.mecab.wakachiSync(text);
} }
async init() { public async init() {
try { try {
const db = fs.readFileSync(this.categorizerDbFilePath, { const buffer = fs.readFileSync(this.categorizerDbFilePath);
encoding: 'utf8' const db = msgpack.decode(buffer);
});
this.classifier = bayes.fromJson(db); this.classifier = bayes.import(db);
this.classifier.tokenizer = this.tokenizer; this.classifier.tokenizer = this.tokenizer;
} catch(e) { } catch (e) {
this.classifier = bayes({ this.classifier = bayes({
tokenizer: this.tokenizer tokenizer: this.tokenizer
}); });
@ -49,7 +55,7 @@ export default class Categorizer {
} }
} }
async learn(id, category) { public async learn(id, category) {
const post = await Post.findOne({ _id: id }); const post = await Post.findOne({ _id: id });
Post.update({ _id: id }, { Post.update({ _id: id }, {
@ -64,7 +70,7 @@ export default class Categorizer {
this.save(); this.save();
} }
async categorize(id) { public async categorize(id) {
const post = await Post.findOne({ _id: id }); const post = await Post.findOne({ _id: id });
const category = this.classifier.categorize(post.text); const category = this.classifier.categorize(post.text);
@ -76,14 +82,12 @@ export default class Categorizer {
}); });
} }
async test(text) { public async test(text) {
return this.classifier.categorize(text); return this.classifier.categorize(text);
} }
save() { private save() {
fs.writeFileSync(this.categorizerDbFilePath, this.classifier.toJson(), { const buffer = msgpack.encode(this.classifier.export());
encoding: 'utf8' fs.writeFileSync(this.categorizerDbFilePath, buffer);
});
} }
} }

302
src/tools/ai/naive-bayes.js Normal file
View File

@ -0,0 +1,302 @@
// Original source code: https://github.com/ttezel/bayes/blob/master/lib/naive_bayes.js (commit: 2c20d3066e4fc786400aaedcf3e42987e52abe3c)
// CUSTOMIZED BY SYUILO
/*
Expose our naive-bayes generator function
*/
module.exports = function (options) {
return new Naivebayes(options)
}
// keys we use to serialize a classifier's state
var STATE_KEYS = module.exports.STATE_KEYS = [
'categories', 'docCount', 'totalDocuments', 'vocabulary', 'vocabularySize',
'wordCount', 'wordFrequencyCount', 'options'
];
/**
* Initializes a NaiveBayes instance from a JSON state representation.
* Use this with classifier.toJson().
*
* @param {String} jsonStr state representation obtained by classifier.toJson()
* @return {NaiveBayes} Classifier
*/
module.exports.fromJson = function (jsonStr) {
var parsed;
try {
parsed = JSON.parse(jsonStr)
} catch (e) {
throw new Error('Naivebayes.fromJson expects a valid JSON string.')
}
// init a new classifier
var classifier = new Naivebayes(parsed.options)
// override the classifier's state
STATE_KEYS.forEach(function (k) {
if (!parsed[k]) {
throw new Error('Naivebayes.fromJson: JSON string is missing an expected property: `'+k+'`.')
}
classifier[k] = parsed[k]
})
return classifier
}
/**
* Given an input string, tokenize it into an array of word tokens.
* This is the default tokenization function used if user does not provide one in `options`.
*
* @param {String} text
* @return {Array}
*/
var defaultTokenizer = function (text) {
//remove punctuation from text - remove anything that isn't a word char or a space
var rgxPunctuation = /[^(a-zA-ZA-Яa-я0-9_)+\s]/g
var sanitized = text.replace(rgxPunctuation, ' ')
return sanitized.split(/\s+/)
}
/**
* Naive-Bayes Classifier
*
* This is a naive-bayes classifier that uses Laplace Smoothing.
*
* Takes an (optional) options object containing:
* - `tokenizer` => custom tokenization function
*
*/
function Naivebayes (options) {
// set options object
this.options = {}
if (typeof options !== 'undefined') {
if (!options || typeof options !== 'object' || Array.isArray(options)) {
throw TypeError('NaiveBayes got invalid `options`: `' + options + '`. Pass in an object.')
}
this.options = options
}
this.tokenizer = this.options.tokenizer || defaultTokenizer
//initialize our vocabulary and its size
this.vocabulary = {}
this.vocabularySize = 0
//number of documents we have learned from
this.totalDocuments = 0
//document frequency table for each of our categories
//=> for each category, how often were documents mapped to it
this.docCount = {}
//for each category, how many words total were mapped to it
this.wordCount = {}
//word frequency table for each category
//=> for each category, how frequent was a given word mapped to it
this.wordFrequencyCount = {}
//hashmap of our category names
this.categories = {}
}
/**
* Initialize each of our data structure entries for this new category
*
* @param {String} categoryName
*/
Naivebayes.prototype.initializeCategory = function (categoryName) {
if (!this.categories[categoryName]) {
this.docCount[categoryName] = 0
this.wordCount[categoryName] = 0
this.wordFrequencyCount[categoryName] = {}
this.categories[categoryName] = true
}
return this
}
/**
* train our naive-bayes classifier by telling it what `category`
* the `text` corresponds to.
*
* @param {String} text
* @param {String} class
*/
Naivebayes.prototype.learn = function (text, category) {
var self = this
//initialize category data structures if we've never seen this category
self.initializeCategory(category)
//update our count of how many documents mapped to this category
self.docCount[category]++
//update the total number of documents we have learned from
self.totalDocuments++
//normalize the text into a word array
var tokens = self.tokenizer(text)
//get a frequency count for each token in the text
var frequencyTable = self.frequencyTable(tokens)
/*
Update our vocabulary and our word frequency count for this category
*/
Object
.keys(frequencyTable)
.forEach(function (token) {
//add this word to our vocabulary if not already existing
if (!self.vocabulary[token]) {
self.vocabulary[token] = true
self.vocabularySize++
}
var frequencyInText = frequencyTable[token]
//update the frequency information for this word in this category
if (!self.wordFrequencyCount[category][token])
self.wordFrequencyCount[category][token] = frequencyInText
else
self.wordFrequencyCount[category][token] += frequencyInText
//update the count of all words we have seen mapped to this category
self.wordCount[category] += frequencyInText
})
return self
}
/**
* Determine what category `text` belongs to.
*
* @param {String} text
* @return {String} category
*/
Naivebayes.prototype.categorize = function (text) {
var self = this
, maxProbability = -Infinity
, chosenCategory = null
var tokens = self.tokenizer(text)
var frequencyTable = self.frequencyTable(tokens)
//iterate thru our categories to find the one with max probability for this text
Object
.keys(self.categories)
.forEach(function (category) {
//start by calculating the overall probability of this category
//=> out of all documents we've ever looked at, how many were
// mapped to this category
var categoryProbability = self.docCount[category] / self.totalDocuments
//take the log to avoid underflow
var logProbability = Math.log(categoryProbability)
//now determine P( w | c ) for each word `w` in the text
Object
.keys(frequencyTable)
.forEach(function (token) {
var frequencyInText = frequencyTable[token]
var tokenProbability = self.tokenProbability(token, category)
// console.log('token: %s category: `%s` tokenProbability: %d', token, category, tokenProbability)
//determine the log of the P( w | c ) for this word
logProbability += frequencyInText * Math.log(tokenProbability)
})
if (logProbability > maxProbability) {
maxProbability = logProbability
chosenCategory = category
}
})
return chosenCategory
}
/**
* Calculate probability that a `token` belongs to a `category`
*
* @param {String} token
* @param {String} category
* @return {Number} probability
*/
Naivebayes.prototype.tokenProbability = function (token, category) {
//how many times this word has occurred in documents mapped to this category
var wordFrequencyCount = this.wordFrequencyCount[category][token] || 0
//what is the count of all words that have ever been mapped to this category
var wordCount = this.wordCount[category]
//use laplace Add-1 Smoothing equation
return ( wordFrequencyCount + 1 ) / ( wordCount + this.vocabularySize )
}
/**
* Build a frequency hashmap where
* - the keys are the entries in `tokens`
* - the values are the frequency of each entry in `tokens`
*
* @param {Array} tokens Normalized word array
* @return {Object}
*/
Naivebayes.prototype.frequencyTable = function (tokens) {
var frequencyTable = Object.create(null)
tokens.forEach(function (token) {
if (!frequencyTable[token])
frequencyTable[token] = 1
else
frequencyTable[token]++
})
return frequencyTable
}
/**
* Dump the classifier's state as a JSON string.
* @return {String} Representation of the classifier.
*/
Naivebayes.prototype.toJson = function () {
var state = {}
var self = this
STATE_KEYS.forEach(function (k) {
state[k] = self[k]
})
var jsonStr = JSON.stringify(state)
return jsonStr
}
// (original method)
Naivebayes.prototype.export = function () {
var state = {}
var self = this
STATE_KEYS.forEach(function (k) {
state[k] = self[k]
})
return state
}
module.exports.import = function (data) {
var parsed = data
// init a new classifier
var classifier = new Naivebayes()
// override the classifier's state
STATE_KEYS.forEach(function (k) {
if (!parsed[k]) {
throw new Error('Naivebayes.import: data is missing an expected property: `'+k+'`.')
}
classifier[k] = parsed[k]
})
return classifier
}