Filipino Word Embeddings

This repository contains download links to pretrained static word embeddings in Filipino.

Corpus Information

The training corpus comprises texts from Tagalog Wikipedia, news articles, and tweets. The corpus contains texts in Filipino and Taglish (code-switching between Tagalog/Filipino and English).

Data Source

Combination of the following:

• WikiText-TL-39
• NewsPH-NLI
• Unpublished Twitter dataset

Preprocessing Rules

• Retain stop words
• Retain commas
• Lowercase text
• Removed quotes and its contents
• Removed brackets, parenthesis, braces, and its contents
• Removed punctuations
• Removed symbols
• Replace numbers with xx_digit
• Replace amounts with xx_amount
• Replace percentages with xx_percentage

Corpus Statistics

After preprocessing, the corpus have:

• 4.68 million sentences
• 14.28 average sentence length
• 66.9 million tokens
• 1.08 million unique tokens

Models

Training setup

The Python library to train word2vec (Skip-Gram with Negative Sampling) and fastText is gensim v4.1.2. Default hyperparameters of gensim v4.1.2 were used except for min_count which is the minimum number of frequency for the word or token to be retained in the vocabulary. The min_count is set to 10 instead of 5 which means that words or tokens that appear less than 10 times in the whole training corpus will be excluded from the vocabulary and will be treated as "unknown word". All models were trained for 20 epochs.

Download via link

Model	Dimensions	Vocab Size	Google Drive ID	Download URL (Size)
word2vec	300	126,687	11eeNuhgEIHCjRFMY9W68WykshTlHmMep	link (269.9 MB)
word2vec	200	126,687	1LcirDe1tbPu2mBhkGQjzTfnEYrhbZcwCY	link (180.2 MB)
word2vec	100	126,687	17XPgQZ7z-jwNgYKHiUvztyVABMLXPqjW	link (90.7 MB)
fastText	300	126,687	12APTxpR9GIbdnawxrn-UahwEZbqxVHFw	link (2.34 GB)
fastText	200	126,687	184LscA1LRu5L6SUEVvaAq1STpr7O7iuT	link (1.55 GB)
fastText	100	126,687	1L8k32TO75J6X0Rzu-SrXQAVz96LFVqvV	link (797.4 MB)

Download via terminal

# Example: download word2vec (300 dimensions)

# Install gdown (if not yet installed)
pip install gdown

# Download data to current directory (you can get the id from the table above)
gdown --id 11eeNuhgEIHCjRFMY9W68WykshTlHmMep

# Extract/Unzip to current directory
unzip word2vec_300dim_20epochs.zip

How to load and use the model (word2vec)

# Example: load word2vec (300 dimensions)

# Import dependencies
from gensim.models import Word2Vec

# Load model
model = Word2Vec.load("word2vec_300dim_20epochs.model")

# Get embeddings for the word 'manila'
model.wv['manila']

# Get top 10 nearest neighbors of the word 'manila'
model.wv.most_similar('manila', topn=10)

How to load and use the model (fastText)

# Example: load fastText (300 dimensions)

# Import dependencies
from gensim.models import FastText

# Load model
model = FastText.load("fasttext_300dim_20epochs.model")

# Get embeddings for the word 'manila'
model.wv['manila']

# Get top 10 nearest neighbors of the word 'manila'
model.wv.most_similar('manila', topn=10)

Example notebook (model usage example, evaluation, visualization)

You can explore the embeddings through this notebook from repo or view at Google Colab for better viewing experience 😎🍟

Contents:

• How to download the model
• How to load the model
• Evaluation scripts
  • Nearest neighbors + 2D plots
  • Word analogy
  • Categorization (concept clustering) + 2D plots
• 3D interactive plot of embedding space (Tensorboard)

Notes

These models/embeddings are not examined for biases. These models were trained for research and/or educational purposes. Please use with caution.

If you used this work in academic contexts, please cite as:

@misc{velasco2021filwembs,
  author = {Dan John Velasco},
  title  = {{Filipino Word Embeddings}},
  year   = {2021},
  howpublished = {\url{https://github.com/danjohnvelasco/Filipino-Word-Embeddings}},
}