Does BERT look at sentiment lexicon?

TL;DR

This study investigates whether BERT models, specifically RuBERT, attend more to sentiment lexicons than neutral words, revealing that a majority of attention heads focus more on sentiment-related tokens, aiding interpretability.

Contribution

The paper provides an analysis of attention weights in RuBERT, showing that most heads prioritize sentiment lexicons over neutral words, enhancing understanding of neural network interpretability in sentiment analysis.

Findings

Approximately 75% of attention heads focus more on sentiment lexicons.

Fine-tuning RuBERT reveals attention distribution differences for sentiment vs. neutral words.

Most attention heads statistically prioritize sentiment lexicons.

Abstract

The main approaches to sentiment analysis are rule-based methods and ma-chine learning, in particular, deep neural network models with the Trans-former architecture, including BERT. The performance of neural network models in the tasks of sentiment analysis is superior to the performance of rule-based methods. The reasons for this situation remain unclear due to the poor interpretability of deep neural network models. One of the main keys to understanding the fundamental differences between the two approaches is the analysis of how sentiment lexicon is taken into account in neural network models. To this end, we study the attention weights matrices of the Russian-language RuBERT model. We fine-tune RuBERT on sentiment text corpora and compare the distributions of attention weights for sentiment and neutral lexicons. It turns out that, on average, 3/4 of the heads of various model var-iants statistically pay more attention to the sentiment lexicon compared to the neutral one.