Relation construction for aspect-level sentiment classification
Introduction
As an important Nature Language Processing (NLP) task, sentiment classification is widely applied in online review analysis [1], text mining [2] and emotion recognition [3]. Document-level [4], [5] and sentence-level sentiment analyses [6], [7] aim to predict the overall sentiment of the input text, which can be considered as coarse-grained sentiment. Given a set of aspects in one sentence, aspect-level sentiment analysis aims at identifying sentiment polarities towards these specific entities [8]. For example, in the sentence “The menu is limited but almost all of the dishes are excellent”, although the overall sentiment of the sentence is neutral, the polarity of “menu” is negative while the one for “dish” is positive.
For the aspect-level sentiment analysis, traditional methods mainly constructed sentiment lexicon dictionary [9] or adopted a feature-based SVM [10] for classification. However, such kind of lexicon dictionary and feature-based methods are labor-intensive. Recently, with the ability to handle complex structure sentences, neural network-based approaches have become the mainstream of sentiment classification. Besides, the attention mechanism has been proved effective for capturing potential semantic relations between the aspect and the context [8], [11], [12]. Apart from the attention mechanism, Zhang et al. [13] applied Graph Convolutional Networks (GCNs) to extract syntactical dependencies among the context, and presented its advantages when handling long sentences.
However, the above studies all associated a single aspect with its contextual words individually, and ignored the benefits brought by multiple relations among aspects. One aspect can be better classified when considering its aspect relations. For instance, as shown in Fig. 1, given the sentence “Nice food. The price is reasonable although the service is poor.”, we can obtain that “food” and “price” are in positive semantic while “service” is negative. Meanwhile, three relations are extracted based on three aspects, namely “food-price”, “food-service” and “price-service”. Suppose that the semantic word “reasonable” is masked, and the polarity of “price” is unknown. If taken relations “food-price” and “price-service” into consideration, we can easily deduce that the polarity of “price” is negative. Therefore, for the sake of better aspect-level sentiment analysis, two challenges should be addressed: 1) how to model relations among aspects; and 2) how to utilize these aspect relations for the aspect-level sentiment classification.
For the first challenge, several methods [14], [15] adopted the attention mechanism or GCNs over aspects to capture potential relations. However, they exaggerated aspect information since the aspect with neutral polarity may generate noises if considered. To address this issue, we only consider explicit aspect relations (“similar” and “opposite”) to avoid neutral aspect noises. Specifically, we model aspect relations by the subtraction between two aspect representations, then auto-annotate their labels based on aspect polarities. For the second challenge, we construct an auxiliary task named relation-level classification, and train the basic aspect classification task and the auxiliary task simultaneously in a multi-task framework to explore the benefits of the aspect relations. Our contributions are summarized as follows:
- •
To the best of our knowledge, this is the first attempt to extract aspect relations as an auxiliary task for aspect-level sentiment analysis, and we explore the benefits brought by aspect relations.
- •
We propose a novel relation construction multi-task learning network that utilizes the bi-attention mechanism to capture bidirectional semantic information between the context and the aspect, and we adopt aspect disagreement regularization to better identify aspect-specific features from overlapped representations.
- •
Extensive experiments on several benchmark datasets validate the effectiveness of the proposed model compared to several comparative baselines, and show the ability to handle relatively small datasets.
The rest of this paper is organized as follows: Section 2 reviews related works. The detailed modules of the proposed method are presented in Section 3. Experimental results and further analyses are shown in Section 4. Section 5 concludes the paper and discusses the future research.
Section snippets
Sentiment classification
Sentiment analysis has been served as an essential role in NLP tasks, and can be divided into three levels: document-level [4], [5], sentence-level [7], [16], and aspect-level [17], [18], [19]. In the document-level sentiment classification, Dou et al. [4] proposed a deep memory network to predict the sentiment polarity of a whole document. In the sentence-level classification, Liu et al. [20] investigated domain representations of multitask learning for the multi-domain sentiment analysis
Problem definition and methodology
In this section, we first give the problem definition and the associated notations. Afterwards, we introduce the detailed methodology and workflows. The overall architecture of our RMN is presented in Fig. 2, where aspect representations are extracted from L-layer GCNs, context representations are encoded by position encoder, and the final aspect-specific representations are generated by the bi-attention module.
Experiments
We conduct extensive experiments with several settings of the proposed method in this section, showing the significant results and effectiveness of the relation-level classification task.
Conclusion
In this paper, we mainly explore aspect relations for aspect-level sentiment analysis. Our work is the first attempt to construct explicit aspect relations as an auxiliary task. We generate aspect representations with the dependency graph based GCNs, and utilize the bidirectional attention mechanism to capture semantic relevance between the aspect and the context. No extra corpus is needed in our framework, and RMN performs better on relatively small datasets with the assistance of information
CRediT authorship contribution statement
Jiandian Zeng: Methodology, Software, Writing-original-draft. Tianyi Liu: Methodology, Software, Validation. Weijia Jia: Validation, Writing – review & editing. Jiantao Zhou: Validation, Writing – review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
We greatly appreciate anonymous reviewers and the associate editor for their valuable and insightful comments that greatly helped to improve the quality of this article. This work was supported by Macau Science and Technology Development Fund under SKL-IOTSC-2021-2023, 0072/2020/AMJ, 077/2018/A2, and 0060/2019/A1, by Research Committee at University of Macau under MYRG2018-00029-FST and MYRG2019-00023-FST, and by Natural Science Foundation of China under 61971476.
References (48)
- et al.
Cross-domain ontology construction and alignment from online customer product reviews
Inf. Sci.
(2020) - et al.
A novel method for sentiment classification of drug reviews using fusion of deep and machine learning techniques
Knowl.-Based Syst.
(2020) - et al.
A method for multi-class sentiment classification based on an improved one-vs-one (ovo) strategy and the support vector machine (svm) algorithm
Inf. Sci.
(2017) - et al.
Abcdm: An attention-based bidirectional cnn-rnn deep model for sentiment analysis
Future Gener. Comput. Syst.
(2021) - et al.
Three-way enhanced convolutional neural networks for sentence-level sentiment classification
Inf. Sci.
(2019) - et al.
Attention-based word embeddings using artificial bee colony algorithm for aspect-level sentiment classification
Inf. Sci.
(2021) - et al.
Ncrf++: An open-source neural sequence labeling toolkit
- B. Liu, L. Zhang, A survey of opinion mining and sentiment analysis, in: Mining text data, Springer, 2012, pp....
- et al.
Time-frequency representation and convolutional neural network-based emotion recognition
IEEE Trans. Neural Networks Learn. Syst.
(2020) Capturing user and product information for document level sentiment analysis with deep memory network
Investigating dynamic routing in tree-structured lstm for sentiment analysis
Attention-based lstm for aspect-level sentiment classification
Multi-grained attention network for aspect-level sentiment classification
Interactive attention networks for aspect-level sentiment classification
Aspect-based sentiment classification with aspect-specific graph convolutional networks
Iarm: Inter-aspect relation modeling with memory networks in aspect-based sentiment analysis
Modeling sentiment dependencies with graph convolutional networks for aspect-level sentiment classification
Knowl.-Based Syst.
Effective lstms for target-dependent sentiment classification
Progressive self-supervised attention learning for aspect-level sentiment analysis
Modelling context and syntactical features for aspect-based sentiment analysis
Enhancing cross-target stance detection with transferable semantic-emotion knowledge
Cited by (26)
Aspect-based sentiment classification with aspect-specific hypergraph attention networks
2024, Expert Systems with ApplicationsDRGAT: Dual-relational graph attention networks for aspect-based sentiment classification
2024, Information SciencesCAT: Continual Adapter Tuning for aspect sentiment classification
2024, NeurocomputingSentence-level sentiment classification based on multi-attention bidirectional gated spiking neural P systems
2024, Applied Soft ComputingA dependency-based hybrid deep learning framework for target-dependent sentiment classification
2023, Pattern Recognition LettersAspect-Pair Supervised Contrastive Learning for aspect-based sentiment analysis
2023, Knowledge-Based Systems