Commentary Analysis: A Machine or a Customer?

Abstract

In the field of e-commerce, shopping reviews play a significant role in influencing consumers' purchase intention. This paper utilizes natural language processing techniques to analyze the information in e-commerce reviews in detail in order to facilitate digital insights.

For Task 1, this paper utilizes Python to extract the review text, split words, and remove deactivated words to preprocess the data. After that, this paper plots the word clouds respectively, and also draws bar charts for the top 20 words in terms of frequency of occurrence in the reviews as well as the items with the most 5-point reviews in OVERALL. Finally, this paper preliminarily analyzes the importance of keywords in review texts based on the bureau TF-IDF method.

For task 2, after the same preprocessing, the fastText tool is utilized to construct word vectors combined with the weights of words obtained from TextRank for keyword extraction. After the feature extraction, a machine learning-based semantic classification model was constructed with the ASIN field of Appendix III and IV as the classification variable. By comparing different machine learning classifiers, the SVM classifier with the highest accuracy is selected as the algorithm for the final semantic analysis model in this paper, and the accuracy rate reaches 69%.

For Task 3, feature extraction based on TF-IDF Weighted Word2vec was processed for the comment text after preprocessing, and for the SVM classifier selected for Task 2, the PSO algorithm was used for optimization, and learning factors and inertia weights were added to improve it, and a text sentiment analysis model based on AIPSO-SVM was constructed. In this paper, experiments were conducted on the data of Appendix V and VI. The experimental accuracy rate reaches more than 75%.

For Task 4, this paper constructs an evaluation system through seven evaluation criteria, such as different angles of comment text, comment time, OVERALL and HELPFUL fields. And based on this, we established a machine review research and judgment model based on T-S fuzzy neural network, and combined with 2D Fourier transform to convert it into a threshold, and experimented with 6 attachments, according to the results, we wrote a letter of advice to help customers identify the categories of reviews.

The highlights of this pape are the comparison of multiple classifiers, the incorporation of an improved PSO algorithm to optimize the classifiers, and the construction of a discriminative model via T-S fuzzy neural networks.