Skip to content

Latest commit

 

History

History
43 lines (34 loc) · 2.25 KB

README.md

File metadata and controls

43 lines (34 loc) · 2.25 KB
Raw

language-detector

Introduction

This supervised learning model predicts the language of text samples using ngram frequency analysis. The objective of this project was to see if we could make a cheap and accurate prediction on the language using a limited range of morphological features. This includes:

  • n-gram combinations (i.e. "an", "ine", "hew")
  • suffixes (including last-letters)
    For example, if a name has the suffix "que", this usually points to French. The goal was to get the model to recognize this and make language predictions in a more efficient/innovative way than using an exhaustive search through a dictionary.

Program flow

Key words are first processed and mapped in dataParser.py. Features used are common ngrams and suffixes(including last letters), extracted using the nltk library in features/grams.py. These morphological features are assembled into a NumPy design matrix in features/extractFeatures.py. The data is then loaded into PyTorch tensors in launch.py to train a feedforward neural network. The model will make predictions on CV/Test sets and on any custom dataset in the test_cases directory (if configured).

Setup

Here is how to get started with running the project:

Installations

The main obstacle downloading torch on your machine. It is highly recommended you use Anaconda as your package manager as it will handle all dependencies (NumPy, SciPy, etc.). You will also need to install nltk:

conda install -c anaconda nltk

Configuration

Most important settings can be toggled in configuration.py. For example, we can alter the set sizes:

training_set_size = 10000
cv_set_size = 2000
test_set_size = 2000

Or how the model learns:

iterations = 20000
learning_rate = 0.001
minibatch = 100

And so on.

Running the Project

The program flow is dictated in launch.py. Run this file to train the model.

Acnowledgements

  • The pytorch model is loosely based off of this logistic regression classifier from Deep Learning Wizard
  • Training, CV, and test sets all extracted randomly using the 10000 most common words on Wikipedia per language