A Contrastive Learning Framework with Tree-LSTMs for Aspect-Based Sentiment Analysis

Different from sentence-level sentiment analysis, the aspect-based sentiment analysis (ABSA) is a fine-grained task that aims to identify the sentiment polarity towards specific aspect terms in a sentence. However, the lack of fine-grained labeled data and the fact that a sentence may contain multiple aspects or complex implicit sentiment relations make ABSA still face challenges. Specifically, effectively exploiting syntactic dependencies to construct contextual information in a sentence to capture implicit sentiment polarities and constructing the data augmentation paradigm to obtain fine-grained aspect-specific information are the key concerns of this paper. To mitigate the above issues, we propose a Contrastive Learning Framework with Tree-Structured LSTM (CLF-TrLSTM), which applies a concatenated form of Tree-LSTMs and self-attention with window mechanism to utilize dependency tree to capture syntactic and contextual information of the sentence. Meanwhile, to alleviate the data scarcity problem, we use mask generation operation and contrastive learning to generate in-domain high-quality positive and negative samples, then encourage anchor sentences and positive samples to be more similar than negative example pairs, which can achieve alignment of different granularities. Finally, experimental results on three public datasets demonstrate that our proposed framework achieves the state-of-the-art performance and comprehensive analysis verifies the effectiveness of each component.