Exploiting BERT with Global-Local Context and Label Dependency for Aspect Term Extraction
Qingxuan Zhang, Chongyang Shi · 2020
Aspect term extraction (ATE) is a subtask of aspect-based sentiment analysis (ABSA), which aims to extract all aspect-specific words in a sentence. Recent neural network methods ignore the problem that word may play different semantic roles in different sentences and have limitation in handling dependencies between labels. In this work, we first exploit BERT as embedding layer to obtain word-level representations and utilize BERT architecture to capture global sequence features. Then, a position-aware attention is proposed to extract local context information. Global-local context representations of words are built by merging the global sequence features and local context information, which can select related information from both sides: global sequence and local context. Finally, to model the label dependency, we construct a label dependency module based on RNN and CRF, where the previous label features are introduced as additional information for label relationship modeling. Experimental results on four benchmark datasets show that our proposed model obtains the state-of-the-art performance.