GitHub - karmdesai/fastTextWeb: fastTextWeb is a custom version of Facebook's text classification library (fastText) that is intended for use on the web. (original) (raw)

fastTextWeb

fastText is a library provided by Facebook, Inc. for efficient text classification and representation learning. It's great for creating highly accurate text classification models in a short period. However, it can't easily be used to serve predictions on the web - fastTextWeb aims to solve this issue. The project was compiled into JavaScript using Emscripten and can be used in a browser or run using Node.

Installation

$ npm install fasttextweb

Usage

fastTextWeb only supports two commands, which helps keep the project files small. I chose to remove all of the other functions because they're not required in order to make predictions. This means that a potential user could train a model using vanilla fastText and then use fastTextWeb to deploy the model and form predictions on the web.

All arguments are expected to be strings, so ensure you use quotation marks when you input arguments.

loadModel

Use loadModel to load a fastText model. The command takes in one argument, model.

$ node

const fastText = require('fasttextweb'); fastText.loadModel('working/');

This function only takes in one argument, model, which is the relative file path of the model you want to load.

makePrediction

Use makePrediction to obtain the most likely labels for a piece of text. The command takes in four arguments, predictionType, textToClassify, k, and th.

$ node

const fastText = require('fasttextweb'); fastText.loadModel('working/'); fastText.makePrediction(, , , );

This function takes in multiple arguments, which are all required.

<predictionType>: You can use either "predict" or "predict-prob". Use "predict" to obtain the k most likely labels for a piece of text. Alternatively, use "predict-prob" to obtain the k most likely labels for a piece of text with probabilities.

<textToClassify>: Simply enter the piece of text you want to classify, as a string.

<k>: The top k labels you want to return through the prediction.

<th>: The probability threshold ("0.0" was the default used during production).

Tests

The folder tests contains several usage examples of fastTextWeb. Below is an example use case of fastTextWeb with the balesModel dataset.

$ node

const fastText = require('fasttextweb'); fastText.loadModel("working/tests/balesModel/compressedModel.ftz"); The model has successfully been loaded!

fastText.makePrediction("predict", "today was a really amazing day!", "1", "0.0"); __label__B

fastText.makePrediction("predict-prob", "how about we put the spoons in the dishwasher?", "2", "0.0"); __label__C 0.959634 __label__B 0.0348174

What's Next

• Publish this package to the npm registry.
• Remove more unrequired functions and parts of the source code.
• Change some of the source code so that the 'index.js' file can be included in a website.

Acknowledgements

1. Stuart Watt (Chief Technology Officer of Turalt) - I really appreciate the resources, help, and support that Stuart has provided. This project really wouldn't have been possible without him!
2. fastText - A library provided by Facebook, Inc. for efficient text classification and representation learning.
3. Emscripten - A toolchain for compiling to WebAssembly, built using LLVM, that lets you run C and C++ on the web at near-native speed without plugins.