Home

Awesome

lapis-bayes

test

lapis-bayes is a Naive Bayes classifier for use in Lua. It can be used to classify text into any category that has been trained for ahead of time.

It's built on top of Lapis, but can be used as a standalone library as well. It requires PostgreSQL to store and parse training data.

Install

$ luarocks install lapis-bayes

Quick start

Create a new migration that look like this:

-- migrations.lua
{
  ...

  [1439944992]: require("lapis.bayes.schema").run_migrations
}

Run migrations:

$ lapis migrate

Train the classifier:

local bayes = require("lapis.bayes")

bayes.train_text("spam", "Cheap Prom Dresses 2014 - Buy discount Prom Dress")
bayes.train_text("spam", "Older Models Rolex Watches - View Large Selection of Rolex")
bayes.train_text("spam", "Hourglass Underwire - $125.00 : Professional swimwear")

bayes.train_text("ham", "Games I've downloaded so I remember them and stuff")
bayes.train_text("ham", "Secret Tunnel's Collection of Rad Games That I Dig")
bayes.train_text("ham", "Things I need to pay for when I get my credit card back")

Classify text:

assert("ham" == bayes.classify_text({"spam", "ham"}, "Games to download"))
assert("spam" == bayes.classify_text({"spam", "ham"}, "discount rolex watch"))

Reference

The lapis.bayes module includes a set of functions that operate on the default classifier and tokenizer:

If the classifier or tokenizer need to be customized, then the classes will need to be manually instantiated.

num_words = bayes.train_text(category, text)

local bayes = require("lapis.bayes")
bayes.train_text("spam", "Cheap Prom Dresses 2014 - Buy discount Prom Dress")

Inserts the tokenized words from text into the database associated with the category named category. Categories don't need to be created ahead of time, use any name you'd like. Later when classifying text you'll list all the eligible categories.

The tokenizer will normalize words and remove stop words before inserting into the database. The number of words kept from the original text is returned.

category, score = bayes.classify_text({category1, category2, ...}, text)

local bayes = require("lapis.bayes")
print bayes.classify_text({"spam", "ham"}, "Games to download")

Attempts to classify text. If none of the words in text are available in any of the listed categories then nil and an error message are returned.

Returns the name of the category that best matches, along with a probability score in natrual log (math.log). The closer to 0 this is, the better the match.

The input text is normalized using the same tokenizer as the trainer: stop words are removed and stems are used. Only words that are available in at least one category are used for the classification.

Tokenization

Whenever a string is passed to any train or classify functions, it's passed through the default tokenizer to turn the string into an array of words.

Tokenization is more complicated than just splitting the string by spaces, text can be normalized and extraneous data can be stripped.

Sometimes, you may want to explicitly provide the words for insertion and classification. You can bypass tokenization by passing an array of words in place of the string when calling any classify or train function.

You can customize the tokenizer by providing a tokenize_text option. This should be a function that takes a single arugment, the string of text, and the return value is the tokens. For example:

local bayes = require("lapis.bayes")
bayes.train_text("spam", "Cheap Prom Dresses 2014 - Buy discount Prom Dress", {
  tokenize_text = function(text)
    -- your custom tokenizer goes here
    return {tok1, tok2, ...}
  end
})

Built-in tokenizers

Postgres Text is the default tokenizer used when no tokenizer is provided. You can customize the tokenizer when instantiating the classifer:

BayesClassifier = require "lapis.bayes.classifiers.bayes"
classifier = BayesClassifier({
  tokenizer = <tokenizer instance>
})

local result = classifier:classify_text(...)

Postgres Text

Uses Postgres tsvector objects to normalize text. This will remove stop words, normalize capitalization and symbols, and convert words to lexemes. Duplicates are removed.

Note: The characteristics of this tokenizer may not be suitable for your objectives with spam detection. If you have very specific training data, preserving symbols, capitalization, and duplication could be beneficial. This tokenizer aims to generalize spam text to match a wider range of text that may not have specific training.

This tokenizer requires an active connection to a Postgres database (provided in the Lapis config). It will issue queries when tokenizing. The tokenizer is uses a query that is specific to English:

select unnest(tsvector_to_array(to_tsvector('english', 'my text here'))) as word

Example:

local Tokenizer = require "lapis.bayes.tokenizers.postgres_text"

local t = Tokenizer(opts)

local tokens = t:tokenize_text("Hello world This Is my tests example") --> {"exampl", "hello", "test", "world"}

local tokens2 = t:tokenize_text([[
  <div class='what is going on'>hello world<a href="http://leafo.net/hi.png">my image</a></div>
]]) --> {"hello", "imag", "world"}

Tokenizer options:

URL Domains

Extracts mentions of domains from the text text, all other text is ignored.

local Tokenizer = require "lapis.bayes.tokenizers.url_domains"

local t = Tokenizer(opts)
local tokens = t:tokenize_text([[
  Please go to my https://leafo.net website <a href='itch.io'>hmm</a>
]]) --> {"leafo.net", "itch.io"}

Schema

lapis-bayes creates two tables:

Running outside of Lapis

Creating a configuration

If you're not running lapis-bayes directly inside of Lapis you'll need to create a configuration file that instructs your script on how to connect to a database.

In your project root, create config.lua:

-- config.lua
local config = require("lapis.config")

config("development", {
  postgres = {
    database = "lapis_bayes"
  }
})

The example above provides the minimum required for lapis-bayes to connect to a PostgreSQL database. You're responsible for creating the actual database if it doesn't already exist.

For PostgreSQL you might run the command:

$ createdb -U postgres lapis_bayes

We're using the standard Lapis configuration format, you can read more about it here: http://leafo.net/lapis/reference/configuration.html

Creating the schema

After database connection has been established the schema (database tables) need to be created. This is done using Lapis' migrations.

Create a file, migrations.lua, and make it look like the following:

-- migrations.lua

return {
  require("lapis.bayes.schema").run_migrations
}

You can test your configuration now by running the migrations with the following command from your shell: (Note you must be in the same directory as your code, migrations.lua and config.lua)

$ lapis migrate

You're now ready to start training and classifying text! (Go back to the top of this document for the tutorial)

Contact

Author: Leaf Corcoran (leafo) (@moonscript)
Email: leafot@gmail.com
Homepage: http://leafo.net
License: MIT