A semantically enhanced dual encoder for aspect sentiment triplet extraction

Jiang, Baoxing; Liang, Shehui; Liu, Peiyu; Dong, Kaifang; Li, Hongye

doi:10.1016/j.neucom.2023.126917

Cited by 5 publications

(1 citation statement)

References 12 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To understand the relationship between aspects, Jiang et al [33] combined the pre-trained model BERT with Bi-directional Long Short-Term Memory (Bi-LSTM) and Graph Convolutional Networks (GCN) to capture the surface and deep semantics of sentences. According to Yang et al [34], ASTE can be enhanced through aspect-view pairing.…”

Section: End-to-end Methodsmentioning

confidence: 99%

Multi-task Dual-Graph Network framework for Aspect Sentiment Triplet Extraction

Wang,

Zheng,

Tang

et al. 2024

Preprint

View full text Add to dashboard Cite

Aspect Sentiment Triplet Extraction is a task that extracts aspects, opinion words corresponding to aspects and sentiment polarity from review sentences. Although aspect emotion triple extraction technology has achieved remarkable results in recent years, it still has a few drawbacks. Many existing research methods ignore important issues, such as the dependency relationships between emotional triad elements and the detection of aspect and viewpoint boundary terms at the phrase-level. Due to the above problems, this paper proposes a multi-task joint learning architecture based on dual graph networks (DGNMT). Firstly, a shared sentence encoder module is proposed in the model. The middle-layer and top-level feature representations of the BERT encoder are applied to two different subtasks in this module. Secondly, Dual-Graph Network Learning module has been designed. As part of the dual-graph network learning layer, this module introduces a graph attention network based on phrase structure trees in the joint extraction task branch. For the triple extraction task branch, a graph convolution network is used based on the modified syntactic dependency graph in order to determine the syntactic relationships between triple elements. Additionally , the Multi-Task Tagging module is proposed to extract labels for tasks as well as triplet extraction tasks. And then, predictions are made based on 1D sequence tags and 2D word pair matrices, and the tag information of the two tasks is transformed into each other through a conversion strategy. According to the final experimental results, the model outperforms a state-of-the-art baseline model on four commonly used evaluation indicators.

show abstract

Section: End-to-end Methodsmentioning

confidence: 99%