Reinforcement Learning-based Aspect Term Extraction using Dilated Convolutions and Differential Equation Initialization

Abstract

Aspect term extraction is a crucial subtask in aspect-based sentiment analysis that aims to discover the aspect terms presented in a text. In this paper, a method for ATE is proposed that employs dilated convolution layers to extract feature vectors in parallel, which are then concatenated for classification downstream. Reinforcement learning is used to save the ATE model from imbalance classification, in which the training procedure is posed as a sequential decision-making process. The samples are the states; the network, and the agent; and the agent gets a more significant reward/penalty for correct/incorrect classification of the minority class compared with the majority class. The training phase, which typically employs gradient-based approaches, including back-propagation for the learning process, is tackled. Thus, it suffers from some drawbacks, including sensitivity to initialization. A novel differential equation (DE) approach that uses a clustering-based mutation operator to initiate the BP process is presented. Here, a winning cluster is identified for the current DE population, and a new updating strategy is used to generate candidate solutions. The BERT model is employed as word embedding, which can be included in a downstream task and fine-tuned as a model, while BERT can capture various linguistic properties. The proposed method is evaluated on two English datasets (Restaurant and Laptop) and has achieved outstanding results, surpassing other deep models (Restaurant: Precision 85.44%, F1-score 87.35%; Laptop: Precision 80.88%, F1-score 80.78%).