Multilingual Part-of-Speech Tagging with Bidirectional Long Short-Term Memory Models and Auxiliary Loss

TL;DR

This paper evaluates bi-LSTM models for multilingual POS tagging, demonstrating their robustness to data size and label noise, and introduces a novel auxiliary loss to improve performance on morphologically complex languages.

Contribution

It introduces a new bi-LSTM model with an auxiliary loss for rare words, achieving state-of-the-art results across 22 languages and analyzing model robustness.

Findings

Bi-LSTMs outperform traditional taggers across languages.

The auxiliary loss improves accuracy on morphologically complex languages.

Bi-LSTMs are less sensitive to training data size and label noise.

Abstract

Bidirectional long short-term memory (bi-LSTM) networks have recently proven successful for various NLP sequence modeling tasks, but little is known about their reliance to input representations, target languages, data set size, and label noise. We address these issues and evaluate bi-LSTMs with word, character, and unicode byte embeddings for POS tagging. We compare bi-LSTMs to traditional POS taggers across languages and data sizes. We also present a novel bi-LSTM model, which combines the POS tagging loss function with an auxiliary loss function that accounts for rare words. The model obtains state-of-the-art performance across 22 languages, and works especially well for morphologically complex languages. Our analysis suggests that bi-LSTMs are less sensitive to training data size and label corruptions (at small noise levels) than previously assumed.