Does BERT Understand Sentiment? Leveraging Comparisons Between Contextual and Non-Contextual Embeddings to Improve Aspect-Based Sentiment Models

When performing Polarity Detection for different words in a sentence, we need\nto look at the words around to understand the sentiment. Massively pretrained\nlanguage models like BERT can encode not only just the words in a document but\nalso the context around the words along with them. This begs the questions,\n"Does a pretrain language model also automatically encode sentiment information\nabout each word?" and "Can it be used to infer polarity towards different\naspects?". In this work we try to answer this question by showing that training\na comparison of a contextual embedding from BERT and a generic word embedding\ncan be used to infer sentiment. We also show that if we finetune a subset of\nweights the model built on comparison of BERT and generic word embedding, it\ncan get state of the art results for Polarity Detection in Aspect Based\nSentiment Classification datasets.\n