Aspect-level sentiment analysis with aspect-specific context position information
Bo Huang,Ruyan Guo,Yimin Zhu,Zhijun Fang,Guohui Zeng,Jin Liu,Yini Wang,Hamido Fujita,Zhicai Shi
DOI: https://doi.org/10.1016/j.knosys.2022.108473
IF: 8.139
2022-01-01
Knowledge-Based Systems
Abstract:In recent years, an increasing number of researchers have focused on the aspect-level sentiment analysis in the field of natural language processing. A coarse-grained sentiment analysis at the document level and a sentiment analysis at the sentence level can only judge an entire text comprehensively, whereas a fine-grained sentiment analysis distinguishes each concrete aspect of the text and makes separate judgments on the sentiment polarity. The word vector representation obtained by a recurrent neural network lacks a description of the distance relationship between the context words and aspect, and traditional models rarely consider the influence of the association between contextual sentences. In this paper, we propose an aspect-level sentiment analysis model with aspect-specific contextual location information. By designing two asymmetrical contextual position weight functions respectively, the model adjusts the weight of contextual words according to the positions of the aspect words in the sentences, and alleviates the interference of the difference in the number of words on both sides of the aspect words on the judgment of sentimental polarity. By utilizing single-sentence-level and multiplesentence-level bidirectional GRU layers, model will extract the influence of the contextual association of each sentence in the document on the aspect sentiment polarity of individual sentences. In addition, we analyze the distribution properties of hard samples and design a novel loss function for the class imbalance problem in the field of sentiment analysis. For dataset 15Rest, the accuracy of our model is 4.27% higher than that of ASGCN, whereas the f1-score, which is more indicative of the classification performance on an imbalanced dataset, can be seen to be improved by 4.31% in comparison to the ASGCN. (C) 2022 Elsevier B.V. All rights reserved.