BARLAT: A Nearly Unsupervised Approach for Aspect Category Detection

• Published in: Neural processing letters 2022-10, Vol.54 (5), p.4495-4519
• Main Authors: Kumar, Avinash, Gupta, Pranjal, Kotak, Nisarg, Balan, Raghunathan, Neti, Lalita Bhanu Murthy, Malapati, Aruna
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Clustering; Complex Systems; Computational Intelligence; Computer Science; Data mining; Datasets; Food; Neural networks; Regularization methods; Representations; Semantics; Sentences; Sentiment analysis; Supervised learning; Unsupervised learning
• ISSN: 1370-4621; 1573-773X
• DOI: 10.1007/s11063-022-10819-4

Aspect category detection is an essential task in aspect-based sentiment analysis. Most previous works use labeled data and apply a supervised learning approach to detect the aspect category. However, to avoid the dependency on labeled data, some researchers have also applied unsupervised learning approaches, wherein variants of topic models and neural network-based models have been built for this task. These unsupervised methods focus on co-occurrences of words and ignore the contextual meaning of the words in the given sentence. Thus, such models perform reasonably well in detecting the explicitly expressed aspect category but often fail in identifying the implicitly expressed aspect category in the sentence. This paper focuses on the contextual meaning of the word. It adopts a document clustering approach requiring minimal user guidance, i.e., only a small set of seed words for each aspect category to detect efficiently implicit and explicit aspect categories. A novel BERT-based Attentive Representation Learning with Adversarial Training (BARLAT) model is presented in this paper, which utilizes domain-based contextual word embedding (BERT) for generating the sentence representation and uses these representations for clustering the sentences through attentive representation learning. Further, the model parameters are generalized better by performing adversarial training, which adds perturbations to the cluster representations. BARLAT is the first nearly unsupervised method that uses the contextual meaning of the words for learning the aspect categories through an adversarial attentive learning approach. The performance of BARLAT is compared with various state-of-the-art models using F1-score on Laptop and Restaurant datasets. The experimental results show that BARLAT outperforms the best existing model by a margin of 1.1% and 2.3% on Restaurant and Laptop datasets, respectively.