Introducing Syntactic Structures into Target Opinion Word Extraction with Deep Learning

Abstract

Targeted opinion word extraction (TOWE) is a sub-task of aspect based sentiment analysis (ABSA) which aims to find the opinion words for a given aspect-term in a sentence. Despite their success for TOWE, the current deep learning models fail to exploit the syntactic information of the sentences that have been proved to be useful for TOWE in the prior research. In this work, we propose to incorporate the syntactic structures of the sentences into the deep learning models for TOWE, leveraging the syntax-based opinion possibility scores and the syntactic connections between the words. We also introduce a novel regularization technique to improve the performance of the deep learning models based on the representation distinctions between the words in TOWE. The proposed model is extensively analyzed and achieves the state-of-the-art performance on four benchmark datasets.

Enhancement of Target‐Oriented Opinion Words Extraction with Multiview‐Trained Machine Reading Comprehension Model

“… TS-GCN [ 5 ] is a syntax-based method that leverages GCN to capture dependency knowledge and utilizes a memory-based strategy to learn multiscale information. ONG [ 4 ] is also a syntax-based method that incorporates the syntactic structures of the sentences through learning syntax-based opinion possibility scores and syntactic connections. …”

“…ONG [ 4 ] is also a syntax-based method that incorporates the syntactic structures of the sentences through learning syntax-based opinion possibility scores and syntactic connections.…”

“…Directional information is also implicitly utilized by IOG [ 1 ], where tokens in the different sides of the target are first encoded separately and then aggregated with a recurrent network. Wu et al [ 3 ] and Veyseh et al [ 4 ] project the relative distance to the given target into embeddings as part of the model inputs. By informing the model with target knowledge, these methods achieve better performances than directly extract all opinions.…”

“…For example, by learning the conjunct relation among two opinions, these two phrases could benefit from each other when identifying the associated targets. State-of-the-art models, Zhang et al [ 5 ] and Veyseh et al [ 4 ], utilize the graph neural network to learn syntactical knowledge and capture opinion correlations. However, they rely on the dependency parser to generate graph structure, leading to a complex pipeline to train and deploy.…”

“…End-to-end neural methods [1] that directly take target information as inputs outperform those rule-based methods with a large margin. State-of-the-art methods [3][4][5] typically leverage extra information such as document-level sentiment representation and syntactic knowledge to assist the learning process. Nevertheless, existing works utilize position embedding to mark the target, which does not contain semantic information.…”