Does Chinese BERT Encode Word Structure?

TL;DR

This paper investigates how Chinese BERT encodes word structure, revealing that word information is captured mainly in middle layers and varies across different NLP tasks.

Contribution

It provides the first detailed analysis of word feature encoding in Chinese BERT using attention and probing methods.

Findings

Word information is captured by Chinese BERT.

Word features are mostly in middle layers.

Different NLP tasks utilize word features to varying extents.

Abstract

Contextualized representations give significantly improved results for a wide range of NLP tasks. Much work has been dedicated to analyzing the features captured by representative models such as BERT. Existing work finds that syntactic, semantic and word sense knowledge are encoded in BERT. However, little work has investigated word features for character-based languages such as Chinese. We investigate Chinese BERT using both attention weight distribution statistics and probing tasks, finding that (1) word information is captured by BERT; (2) word-level features are mostly in the middle representation layers; (3) downstream tasks make different use of word features in BERT, with POS tagging and chunking relying the most on word features, and natural language inference relying the least on such features.