SAMBERT: Improve Aspect Sentiment Triplet Extraction by Segmenting the Attention Maps of BERT

Anonymous

Abstract: Aspect Sentiment Triplet Extraction (ASTE) performs fine-grained sentiment analysis in a unified way through extracting sentiment triplets comprised of aspect terms, opinion spans, and their sentiment relations in sentences. The previous works show the adoption of BERT, which simply leverages its last layer output as the word representation, is beneficial for recognizing triplet elements. However, their methods limit the potential of pretrained knowledge in BERT, since the different layers can capture multi-level linguistic information existing in sentences, which are useful for ASTE as well. In this work, we explore to access the rich pretrained knowledge by fully leveraging its attention maps of different layers. To this end, we propose to Segment the Attention Maps of BERT (SAMBERT) by taking the merits of semantic segmentation, which can effectively discriminate the desired objects from others in an image. In this procedure, we can further reason over the knowledge of different levels in these attention maps to distinguish aspect terms, opinion spans and their sentiment relations from other parts, which results in a same-shape tagging matrix of word pairs for deriving sentiment triplets. Through the extensive experiments on four benchmarks, we demonstrate our method can achieve a new state of the art.