Inducing Target-Specific Latent Structures for Aspect Sentiment Classification

Abstract

Aspect-level sentiment analysis aims to recognize the sentiment polarity of an aspect or a target in a comment. Recently, graph convolutional networks based on linguistic dependency trees have been studied for this task. However, the dependency parsing accuracy of commercial product comments or tweets might be unsatisfactory. To tackle this problem, we associate linguistic dependency trees with automatically induced aspectspecific graphs. We propose gating mechanisms to dynamically combine information from word dependency graphs and latent graphs which are learned by self-attention networks. Our model can complement supervised syntactic features with latent semantic dependencies. Experimental results on five benchmarks show the effectiveness of our proposed latent models, giving significantly better results than models without using latent graphs.

Eliminating Sentiment Bias for Aspect-Level Sentiment Classification with Unsupervised Opinion Extraction

“…For ALSC task, we evaluate our model on five datasets 1 , whose statistics are listed in Table 1, including (i) Laptop14 (SemEval-2014T4) Table 4: ALSC results on the five datasets. †Numbers are from Dai et al (2021), and others are from the original papers, i.e., ASGCN (Zhang et al, 2019a), CDT (Sun et al, 2019b), BiGCN (Zhang and Qian, 2020), DGEDT (Tang et al, 2020), RGAT , kumaGCN (Chen et al, 2020), PWCN-FT-RoBERTa . (Pontiki et al, 2014) with laptop reviews, (ii) Rest14 (SemEval-2014T4), Rest15 (SemEval-2015T12) and Rest16 (SemEval-2016T5) (Pontiki et al, 2014(Pontiki et al, , 2015(Pontiki et al, , 2016 with restaurant reviews, and (iii) Twitter (Mitchell et al, 2013) with tweets.…”

“…Despite their effectiveness, they lack interpretability to the sentiment prediction. For further boosted, many methods (Sun et al, 2019a;Chen et al, 2020;Mao et al, 2021;Dai et al, 2021) introduce the pre-trained Transformer, which also brings a little interpretability due to the highlight modification between aspect and opinion terms derived from potential syntax knowledge (Wu et al, 2020;Dai et al, 2021). However, the derived interpretability is far from human-level.…”

“…Most recent works capture the modification relation between aspect and opinion terms in an implicit manner, which is usually achieved by integrating graph neural networks (GNNs) over dependency parsing tree into text representation learning (Zhang et al, 2019a;Sun et al, 2019b;Tang et al, 2020;. Further, the performance can be significantly boosted when incorporating pre-trained Transformer encoder, e.g., BERT (Devlin et al, 2019) in a fine-tuning paradigm (Sun et al, 2019a;Tang et al, 2020;Chen et al, 2020). Although these works achieve excellent results even close to humans, they cannot derive the interpretability to explain why an aspect is associated with the polarity prediction.…”

Dual Graph Convolutional Networks for Aspect-based Sentiment Analysis

“…Most recently, several works explore the idea of combining different types of graph for ABSA task. For instance, (Chen et al, 2020) combined a dependency graph and a latent graph to generate the aspect representation. (Zhang and Qian, 2020) observed the characteristics of word co-occurrence in linguistics and designed hierarchical syntactic and lexical graphs.…”

“…More recent efforts (Zhang et al, 2019;Sun et al, 2019b;Huang and Carley, 2019;Zhang and Qian, 2020;Chen et al, 2020;Liang et al, 2020;Wang et al, 2020;Tang et al, 2020) have been de-voted to graph convolutional networks (GCNs) and graph attention networks (GATs) over dependency trees, which explicitly exploit the syntactic structure of a sentence. Consider the dependency tree in Figure 1; the syntactic dependency can establish connections between the words in a sentence.…”