ASK-RoBERTa: A pretraining model for aspect-based sentiment classification via sentiment knowledge mining
Introduction
Sentiment analysis and opinion mining offer valuable opportunities for the extraction and analysis of emerging patterns that appear as a result of the rapid development of social media communities [1]. The opportunity to automatically capture the general public’s sentiments about social events, political movements, marketing campaigns, and product preferences has attracted interest of both the scientific community and the business world [2]. When interacting with social media, people generally express their thoughts and opinions on a wide range of aspects. The goal of aspect-based sentiment classification (ABSC) is to discover sentiment polarities (e.g. positive, negative, neutral) for specified aspects within sentences and documents in contrast to conventional sentiment analysis that predicts the overall sentiment value of a given comment [3]. One of the primary benefits of ABSC approaches is their ability to extract the sentiment and polarity of a specific aspect from its context sentence. Different aspects of a given document can be studied, potentially revealing highlights of the underlying aspects, polarities, and meanings. ABSC datasets generally encompass a wide range of contexts, components, and sentiments. Consider the sentence “Boot time is superfast, but the battery life is poor”. This sentence combines contrasting polarities as the processing time will be satisfactory but under the implicit condition that battery life will be improved, which is not the case. This is a straightforward example of how ABSC grew in popularity due to its ability to extract useful information from textual comments.
Most early ABSC models employ a recursive neural network to improve sentiment classification accuracy by incorporating syntactic structural data [4]. ABSC models capture an effective representation of syntactic information in the hidden state layer [5], [6]. They are particularly effective at filtering out irrelevant words and aspect information [7], [8]. Although ABSC models have achieved significant performance, they generally lack the ability to mine sentiment dependencies between the aspect terms and contextual words. The significance of sentiment words in context is ignored, and this negatively affects ABSC performance.
Sentiment analysis based on machine and deep learning faces many challenges such as insufficient labelled data and poor generalization ability. Researchers have combined sentiment knowledge with supervised data to improve the model performance. For example, when extracting and analysing sentiments from texts, sentiment knowledge based on quality background dictionaries can capture fine-grained supervision information [9], [10]. Since sentiment dictionary knowledge is integrated into the modelling language, a word vector representation improves the performance of sentiment analysis tasks [11].
Aspect terms and sentiment words have recently been introduced in mask language model pretraining tasks that improves BERT performance (i.e., models based on multiple sentiment classification tasks) [12]. Moreover, graph convolutional networks can be derived using dependency trees and sentiment common sense and dependencies associated with specific aspects and terms [13]. Sentiment knowledge can be used as an effective auxiliary for identifying and explaining inherent dependencies between the aspect terms and sentiment words. In fact, in a given sentence, the sentiment polarity of a given aspect is determined by its own meaning and that of its related words. While recent research has achieved valuable performance in extracting sentiment knowledge, it still lacks fine-grained mining of aspect terms. This leads us to introduce a twofold approach, whose first objective is to introduce a set of general rules for extracting aspect term polarities. The second part incorporates a pretrained RoBERTa [14] model with optimized mask rules to better capture the dependency between the aspect terms and sentiment words.
The main contributions of this paper are as follows:
- •
A set of aspect mining rules based on part-of-speech tagging and sentence dependency grammar that are applied to aspects containing sentiment words. These rules consider word dependencies, compounds, and conjunctions to improve the overall accuracy.
- •
An aspect sentiment knowledge-adaptive pretraining model. The mask rules of the pretraining model are optimized to better capture the dependency between the aspects and sentiments.
- •
Extensive experiments were applied to four SemEval datasets to evaluate ASK-RoBERTa performance and the superiority of the proposed model against the baselines was demonstrated.
The rest of the paper is organized as follows. Section 2 presents ABSC-related works and sentiment knowledge, while Section 3 develops the main principles of our modelling approach. Section 4 presents the experimental setup and evaluation results. Finally, Section 5 concludes the paper and outlines future directions.
Section snippets
Related work
ABSC research falls under the umbrella of entity-level sentiment analysis. In the early years, sentiment analysis was primarily based on sentiment dictionaries and common machine learning methods. Kamps et al. used a WordNet English sentiment dictionary to determine the sentiment polarity of English texts [15] used a WordNet English sentiment dictionary to determine the sentiment polarity of English texts. Although the classification method based on a sentiment network dictionary is relatively
ASK-RoBERTa: Aspect sentiment knowledge-adaptive pretraining model
The overall architecture of the aspect sentiment knowledge-adaptive pretraining model is shown in Fig. 1. For existing knowledge-enhanced language representation models [12], our model mines aspect-sentiment knowledge that contains sentiment words, aspect terms, and the polarity of these sentiment words. Given an input sentence, the model first mines the sentiment words in the sentence using the sentiment word dictionary. Then, a series of aspect mining rules are based on part-of-speech tagging
Dataset and hyperparameters
The experiments are conducted on four public benchmark datasets on the restaurants and laptops domain of SemEval 2014 task 4 [41] (Restaurant14, Laptop14), restaurants domain of SemEval 2015 task 12 [3] (Restaurant 15), and restaurants domain of SemEval 2016 task 5 [42] (Restaurant16). Each sample consists of a review sentence, an aspect term that consists of one or multiple words, and sentiment polarity towards the aspect. The main statistics of the datasets are shown in Table 6.
The training
Conclusion
The research reported in this paper develops an aspect sentiment knowledge-adaptive pretraining ABSC model. Aspect- sentiment masking and two sentiment pretraining objectives incorporate aspect-sentiment knowledge into the pretraining model. To mine accurate aspect terms, a series of rules are proposed based on part of speech and sentence dependency grammar. ASK-RoBERTa outperforms current deep learning models and several BERT-based models. In the ablation experiment, it can be proven that each
CRediT authorship contribution statement
Lan You: Conceptualization, Methodology, Funding acquisition, Project administration, Supervision, Resources. Fanyu Han: Methodology, Software, Writing – original draft, Writing – review & editing. Jiaheng Peng: Data curation, Validation, Software. Hong Jin: Conceptualization, Methodology, Supervision, Investigation. Christophe Claramunt: Writing – reviewing & 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.
Acknowledgements
This work was partially supported by the Technology Innovation Special Program of Hubei Province (No. 2022BAA044, No. 2021BAA188), the Key Project of Science and Technology Research Program of Hubei Provincial Education Department (No. D20201006), and the National Natural Science Foundation of China (No. 61977021).
References (50)
- et al.
Aspect-based sentiment analysis via affective knowledge enhanced graph convolutional networks
Knowl.-Based Syst.
(2022) - et al.
Aspect-level sentiment classification based on attention-bilstm model and transfer learning
Knowl.-Based Syst.
(2022) - et al.
Consensus vote models for detecting and filtering neutrality in sentiment analysis
Inf. Fusion
(2018) - et al.
Knowledge-enabled BERT for aspect-based sentiment analysis
Knowl.-Based Syst.
(2021) - et al.
A survey of opinion mining and sentiment analysis
Affective computing and sentiment analysis
IEEE Intell. Syst.
(2016)- M. Pontiki, D. Galanis, H. Papageorgiou, S. Manandhar, I. Androutsopoulos, Semeval-2015 task 12: Aspect based sentiment...
- L. Dong, F. Wei, C. Tan, D. Tang, M. Zhou, K. Xu, Adaptive recursive neural network for target-dependent twitter...
- D. Tang, B. Qin, X. Feng, T. Liu, Effective LSTMs for target-dependent sentiment classification, in: Proceedings of...
- M. Zhang, Y. Zhang, D.-T. Vo, Gated neural networks for targeted sentiment analysis, in: Proceedings of the Thirtieth...
Learning sentiment-specific word embedding for twitter sentiment classification
Roberta: A robustly optimized bert pretraining approach
GSMNet: Global semantic memory network for aspect-level sentiment classification
IEEE Intell. Syst.
Aspect level sentiment classification with attention-over-attention neural networks
Cited by (18)
Contrastive variational information bottleneck for aspect-based sentiment analysis
2024, Knowledge-Based SystemsStore, share and transfer: Learning and updating sentiment knowledge for aspect-based sentiment analysis
2023, Information SciencesHierarchical Spatiotemporal Aspect-Based Sentiment Analysis for Chain Restaurants using Machine Learning
2024, International Journal of Advanced Computer Science and ApplicationsImproving Clothing Product Quality and Reducing Waste Based on Consumer Review Using RoBERTa and BERTopic Language Model
2023, Big Data and Cognitive Computing