Dirichlet-Smoothed Word Embeddings for Low-Resource Settings

TL;DR

This paper enhances count-based word embeddings with Dirichlet smoothing to improve their performance in low-resource language settings, outperforming recent methods like PU-Learning.

Contribution

It introduces Dirichlet smoothing to PPMI-based embeddings, making them more effective for low-resource languages and domains.

Findings

Outperforms PU-Learning in low-resource scenarios

Achieves competitive results for Maltese and Luxembourgish

Improves bias correction for rare words

Abstract

Nowadays, classical count-based word embeddings using positive pointwise mutual information (PPMI) weighted co-occurrence matrices have been widely superseded by machine-learning-based methods like word2vec and GloVe. But these methods are usually applied using very large amounts of text data. In many cases, however, there is not much text data available, for example for specific domains or low-resource languages. This paper revisits PPMI by adding Dirichlet smoothing to correct its bias towards rare words. We evaluate on standard word similarity data sets and compare to word2vec and the recent state of the art for low-resource settings: Positive and Unlabeled (PU) Learning for word embeddings. The proposed method outperforms PU-Learning for low-resource settings and obtains competitive results for Maltese and Luxembourgish.