Optimizing small BERTs trained for German NER

TL;DR

This paper explores methods to optimize small BERT models for German NER, combining techniques from various BERT variants, proposing new fine-tuning methods, and introducing Whole-Word Attention to improve efficiency and performance.

Contribution

It introduces novel fine-tuning modifications and the Whole-Word Attention mechanism, enhancing small BERT models for German NER tasks with improved efficiency and accuracy.

Findings

Whole-Word Attention reduces memory usage with slight performance gains

New fine-tuning methods improve NER accuracy

Combining techniques from multiple BERT variants enhances small model performance

Abstract

Currently, the most widespread neural network architecture for training language models is the so called BERT which led to improvements in various Natural Language Processing (NLP) tasks. In general, the larger the number of parameters in a BERT model, the better the results obtained in these NLP tasks. Unfortunately, the memory consumption and the training duration drastically increases with the size of these models. In this article, we investigate various training techniques of smaller BERT models: We combine different methods from other BERT variants like ALBERT, RoBERTa, and relative positional encoding. In addition, we propose two new fine-tuning modifications leading to better performance: Class-Start-End tagging and a modified form of Linear Chain Conditional Random Fields. Furthermore, we introduce Whole-Word Attention which reduces BERTs memory usage and leads to a small increase in performance compared to classical Multi-Head-Attention. We evaluate these techniques on five public German Named Entity Recognition (NER) tasks of which two are introduced by this article.