Consumer feedback is highly valuable in business to assess their performance and is also beneficial to customers as it gives them an idea of what to expect from new products. In this research, the aim is to evaluate different deep learning approaches to accurately predict the opinion of customers based on mobile phone reviews obtained from Amazon.com. The prediction is based on analysing these reviews and categorizing them as positive, negative, or neutral. Different deep learning algorithms have been implemented and evaluated such as simple RNN with its four variants, namely, Long Short-Term Memory Networks (LRNN), Group Long Short-Term Memory Networks (GLRNN), gated recurrent unit (GRNN), and update recurrent unit (UGRNN). All evaluated algorithms are combined with word embedding as feature extraction approach for sentiment analysis including Glove, word2vec, and FastText by Skip-grams. The five different algorithms with the three feature extraction methods are evaluated based on accuracy, recall, precision, and F1-score for both balanced and unbalanced datasets. For the unbalanced dataset, it was found that the GLRNN algorithms with FastText feature extraction scored the highest accuracy of 93.75%. This result achieved the highest accuracy on this dataset when compared with other methods mentioned in the literature. For the balanced dataset, the highest achieved accuracy was 88.39% by the LRNN algorithm.

中文翻译：

---

通过Word嵌入和RNN变体对Amazon Online Reviews的情感分析进行评估

消费者反馈对于评估他们的绩效在业务中具有非常重要的价值，同时也使客户对新产品有何期望，因此对客户也很有益。在这项研究中，目标是评估不同的深度学习方法，以根据从Amazon.com获得的手机评论准确地预测客户的意见。该预测是基于对这些评论的分析并将其归类为正面，负面或中立的。已经实现和评估了不同的深度学习算法，例如具有四个变体的简单RNN，即长短期记忆网络（LRNN），组长短期记忆网络（GLRNN），门控循环单元（GRNN）和更新递归单位（UGRNN）。所有评估的算法都与单词嵌入相结合，作为特征提取方法进行情感分析，包括“ Sklip-grams”的Glove，word2vec和FastText。根据准确性，查全率，准确性和F1分数对平衡和非平衡数据集评估具有三种特征提取方法的五种不同算法。对于不平衡数据集，发现具有FastText特征提取功能的GLRNN算法得分最高，为93.75％。与文献中提到的其他方法相比，该结果在该数据集上实现了最高的准确性。对于平衡数据集，通过LRNN算法获得的最高准确度为88.39％。精度和F1得分（针对平衡和非平衡数据集）。对于不平衡数据集，发现具有FastText特征提取功能的GLRNN算法得分最高，为93.75％。与文献中提到的其他方法相比，该结果在该数据集上实现了最高的准确性。对于平衡数据集，通过LRNN算法获得的最高准确度为88.39％。精度和F1得分（针对平衡和非平衡数据集）。对于不平衡数据集，发现具有FastText特征提取功能的GLRNN算法得分最高，为93.75％。与文献中提到的其他方法相比，该结果在该数据集上实现了最高的准确性。对于平衡数据集，通过LRNN算法获得的最高准确度为88.39％。