Semantic-Syntax Cascade Injection Model for Aspect Sentiment Triple Extraction

Abstract

Aspect sentiment triple extraction aims to extract all aspects, opinions, and sentiments in a sentence and pair them into triples. The main challenge lies at mining the dependency between the aspect and corresponding opinion with the specific sentiment. Existing methods capture the dependency via either pipeline framework or collapsed sequence labeling model. However, the pipeline framework may suffer from error propagation, while collapsed tags cannot deal with complex pairing situations where the overlap or long dependency exists. In this paper, we propose a novel semantic-syntax cascade injection model (SSCIM) to address above issues. SSCIM adopts a cascade framework with joint training schema, where its lower layer extracts the aspects and injects those aspects into the upper layer to extract opinion and sentiment simultaneously. Such design is inspired by the fact that the sentiment is often conveyed in opinions, and the joint training schema alleviates error propagation effectively. Moreover, a novel semantic-syntax information injection gate (IIG) is designed to bridge the upper and lower layers of our model, enabling SSCIM to better capture the dependency between aspects and opinions. Experimental results on four benchmark datasets demonstrate the superior performance of the proposed model over state-of-the-art baselines.