KEY DATA EXTRACTION AND EMOTION ANALYSIS OF DIGITAL SHOPPING BASED ON BERT

Abstract

Purpose: The objective of this paper is to focus on extracting the key words about the product quality and the customer experience with the same in a more efficient and accurate way by pre-training the Bidirectional Encoders Representations from Transformers (BERT) model with the quality domain knowledge and classifying the result with deep learning technique.

Methodology: Dataset is considered to be amazon reviews which is a combination of single product-based customer reviews and several products and their reviews which is of medium – large size. This dataset is subjected to initial process of cleaning, data wrangling, Exploratory Data Analysis with pre-trained BERT along with a neural network classifier. The BERT classifier is loaded along with tokenizer in the input modules. The BERT model is configured and training for fine-tuning. The prediction is done based on the final fine tuning.

Findings: BERT model along with TF-IDF topic extraction model was implemented to analyse the trend and theme of the outbreak which eventually helped to analyse the public concerns and appropriate health support. Fine tuning of Chinese BERT model and softmax neural network layer was used to train the model to classify into three sentiments which resulted in 75.65% accuracy. Higher accuracy was expected to obtain but was in need to improve in the modelling and more datasets from different parts of the world will lead to much more accuracy in regards with public concerns. A function is generated to output a sample permutation and thus its replication which will be a single statistic. We consider a hypothesis that the words distribution is of with same identity and setting a value of probability as with minimum value of 5.9. When the p-value values to 0.0 which gives as the null hypothesis is invalid. The baseline will be TF-IDF model with logistic regression. Here a prediction function along with prediction matrix values is generated. The model weights and tuning are interpreted with the help of Eli5 library. The pre-train model is initialized and the configuration is used with layers of encoding and pooling with dimensionality of 768.With this initialization, logits for the input sequence is generated.

Originality/value: The BERT model which is pre-trained is enables with tokenizing the input dataset which is taken as amazon Alexa product review dataset. While the input is loaded, pre-cleaning process is done such as managing the equal negative and positive comments so that the predictions can be made easy. In order to identify how much is the difference between negative and positive comments we implement the testing of permutation and from that calculating the p-value.