Sentiment Analysis of Product Reviews on E-Commerce Platforms

Based on NLP: Take HUAWEI P40 as an Example

Rui Yang, Yuzhu Xie, Jie Sun, Ling Ding, Zhemiao Zhang and Jianjun Wang*

School of Sciences, Sichuan Agricultural University, Ya’an, Sichuan,625014, China

Keywords: Natural Language Processing, HUAWEI, Sentiment Analysis, Online Shopping Review.

Abstract: As the boom and popularity of e-commerce greatly influenced people's consumption patterns and

socio-economic structures, the study of product reviews has great significance for guiding purchases. Most

of these product reviews are subjective and non-structural. Natural Language Processing (NLP) is an

effective method to address communication between humans and computers. It is commonly used to convert

unstructured language into structured language for computer recognition and processing to solve problems

that are difficult to solve manually. In this paper, we put out the previous HUAWEI Mall reviews about

HUAWEI P40 model mobile phone as the data source for natural language processing. And then, we use a

CNN model based on the language of Python to perform sentiment analysis of consumers' attitudes about

this phone. The results illustrate that consumers' concerns about the HUAWEI P40 phone are generally

divided into the following categories: photography, battery life, communication, running speed, display

effect and appearance, and their attitudes are generally positive and optimistic. In other word, the negative

sentiment towards this phone is not significant.

1 INTRODUCTION

1.1 Research Significance

In recent years, in the background of "big data" era,

with the popularity of online shopping, Taobao,

Jingdong and other major e-commerce platforms are

flourishing and competing fiercely (

TAO, ZHANG,

SHI, WEI 2020, ZENG, YU 2020, Feng 2019, Li

2019

). They are not only a platform for selling goods,

but also a platform for consumers to exchange

product use experience as well as quality evaluation.

Therefore, many online users often choose to check

the information and evaluation of a certain product

before buying or using it, and different consumers

have different performance tendencies for the same

type of products.

In order to improve the quality of customer

service, in addition to price wars, it is increasingly

important to understand the needs of customers and

listen to their voices, an important way is data

mining and sentiment analysis of consumers' text

reviews (

TAO, ZHANG, SHI, WEI 2020

). Getting

these information also can improve the

competitiveness of the corresponding products, and

can be used as a guide for consumers' purchases,

thus it is of great practical significance to help

consumers' choices become more rational.

1.2 Current Research Status at Home

and Abroad

Text sentiment analysis is a multidisciplinary

research field, which belongs to the scope of

computer linguistics. It has a wide range of

applications in computer science as well as

management, political science, economics,

sociology, and the booming of business in the

industry has further expanded its applications (

TAO,

ZHANG, SHI, WEI 2020

). Already in the 21st

century, as a new research direction, this field has

gradually become active and attracted increasingly

more researchers. In the application of sentiment

analysis, "user comment analysis and decision

making" is one of the most frequently used

applications of sentiment analysis technology (

Feng

2019

). Many domestic and foreign research

institutions have developed sentiment analysis

systems in various fields according to the specific

needs in real life to help users analyze and make

decisions on large amount of information. Many

802

Yang, R., Xie, Y., Sun, J., Ding, L., Zhang, Z. and Wang, J.

Sentiment Analysis of Product Reviews on E-Commerce Platforms Based on NLP: Take HUAWEI P40 as an Example.

DOI: 10.5220/0011768000003607

In Proceedings of the 1st International Conference on Public Management, Digital Economy and Internet Technology (ICPDI 2022), pages 802-806

ISBN: 978-989-758-620-0

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

commercial systems and applications based on

sentiment analysis have emerged. For example,

Google Shopping, a representative platform based

on product reviews, provides users with product

search and price comparison services for online

purchasing platforms. OpinionE Q allows

commercial organizations and individuals to

customize product analytics services on demand (

2019

). However, due to the increasingly swelling

problem of the diversity of sentiment research

objects and the complexity of sentiment tasks.

Sentiment objects range from positive and negative

tendency comments on products and services to

categorizing user and topic sentiments in social

media categories, with more diverse expressions and

a greater variety of emotions, and a corresponding

shift in research content, including more attention to

the user's information as well as the change of the

emotion for the user (

Li 2019, Ziang, Wang, Li,

Lévy, Nie, Jurafsky, Ng 2017, Lan, Chen,

Goodman, Gimpel, Sharma, ALBERT 2020,

Young, Hazarika, Poria, Cambria 2018, Rush

2018, Vaswani, Shazeer, Parmar, Uszkoreit, Jones,

Gomez, Kaiser, Polosukhin 2017

). On the other

hand, consumers' comments on short texts are

diverse in their expression of emotions, either in a

straightforward manner or through rhetorical devices

or even irony. The multiple forms of expression of

emotions such as irony make the analysis inaccurate.

This project applies a CNN model with fewer

parameters but better performance to text data for

sentiment analysis, and uses the analysis results to

evaluate the attributes of products. This will help

merchants to improve their products and consumers

to choose more suitable products for themselves.

2 METHODS

2.1 Data Collection and Pre-processing

At present, online shopping had become one of the

important ways for people to trade and consume. By

the end of 2020, Taobao and Jingdong two major

e-commerce platforms had reached 1.2 billion active

purchasing users in the past year. Due to consumer

preference, digital enthusiasts tended to purchase

electronic products from official channels. As of

July 2021, as the P40 model phone's main and

official sales channel, more than 30,000 after-sales

reviews had been accumulated in the HUAWEI

Mall. If the entire model phone's review data were

crawled and studied, the process would be very

tedious, resulting in unnecessary human, financial,

and material waste. Therefore, this paper randomly

selected 500 real reviews of HUAWEI P40 model

mobile phones as the basis of the research. The

study first selected 10,000 reviews data of other

mobile phone products with known labels to train

the CNN model. The consumers' attitudes of

"positive", "negative" or "other" are determined

based on the content of the reviews of different

labels.

For text data, text data usually contained a lot of

noise that is not relevant to the task, so it should be

Pre-processed such as removing URL links,

removing emoji emoji, converting English upper

case to lower case and converting Chinese

traditional to simplified, etc.

Simultaneously, 492 valid questionnaires were

collected from online questionnaire survey of

HUAWEI mobile phone users. The results showed

that the users' concern about HUAWEI p40 mobile

phone are mainly focused on six aspects, such as

running speed, photography, battery life,

communication, appearance and display effect.

According to the above six tags, the dataset is

classified by Python in six labels to get six review

sets.

2.2 CNN Model Building

In the field of computer deep learning, CNN

convolutional neural network is one of the most

successful application of deep learning algorithm.

Consisting of neurons with learnable weights and

bias constants, each neuron receives some input and

does some dot product computation to output a score

for each classification.

Input OutputConv

Pool

Conv Pool

Figure 1: CNN convolutional neural network model.

Sentiment Analysis of Product Reviews on E-Commerce Platforms Based on NLP: Take HUAWEI P40 as an Example

803

Score

Display Effect

Appearance

Other

Running Speed

Communication

Battery life

Photograph

Figure 2: HUAWEI Mall HUAWEI P40 review sentiment analysis values (mean).

2.3 Model Training and Data

Processing

Due to the non-full connectivity and weight sharing

characteristics of CNN neural network, which makes

its complexity lower and the number of weights less,

so CNN convolutional neural network is chosen for

the review sentiment analysis of HUAWEI P40

mobile phone. The text in the review set was

Tokenized, and a vocabulary table was built. Then

the Token sequence was converted into a word

embedding matrix. The word embedding was then

fed into the CNN neural network for training. When

training the model, the number of traversals could

control the training time of the model. The smaller

the number of traversals, the shorter the training

time, but the training effect would also decrease.

The learning rate could be adjusted between 0.001

and 0.0001, and different learning rate models

would also have different training effects. After

multiple adjustments, we found that setting the

number of traversals to 30 and a learning rate of

0.001 has the best training effect. At the end of the

final training, the accuracy of the CNN model was

94%, and the model was considered good and the

next step of data processing could be performed.

The 500 data exported from HUAWEI Mall were

imported into the trained model for sentiment

analysis, and the following results were obtained

3 ANALYSIS OF THE RESULTS

3.1 The Overall Attitude of Consumers

Towards Each Label Shows a

Positive Trend

The sentiment analysis of HUAWEI P40 reviews in

HUAWEI Mall showed that the results were divided

into positive and negative scores, representing

whether consumers had a positive or negative

attitude towards the phone. Both positive and

negative scores ranged from 0 to 1, when the score

was greater than 0.5, it was considered that

consumers have an optimistic and positive attitude

towards this feature of the phone. when the score

was less than 0.5, consumers had an unfavorable

attitude. Meanwhile, the closer the value was to 0 or

1, the more absolute consumer attitude was.

The results of the sentiment analysis showed that

the sentiment scores about the HUAWEI P40 phone

concerned by online consumers were greater than

0.5 in all 6 aspects, which means that consumers

were more positive about this phone and generally

ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology

804

Table 1: Evaluation indicators Number of choices Percentage (492 surveys).

Evaluation index Select numbe

Proportion

Apearance 457 0.9289

Runnin

442 0.8984

Photograph 376 0.7642

Batter

life 257 0.5224

HongMeng system 141 0.2866

Communication 137 0.2785

Cost performance 122 0.2480

Effect of dis

105 0.2134

Hand feeling 97 0.1972

Others 70 0.1423

satisfied with it.

Among them, consumers think the best was the

"display effect" of HUAWEI P40, with a value of

0.999 and an image display of 1. Secondly, people

were generally satisfied with the "appearance" and

"running speed" of this phone, with a value of

0.9.The lowest rated label was "battery life" for the

HUAWEI P40, which indicated that consumers have

higher expectations for the battery life of this phone.

3.2 Consumers are More Diversified in

Terms of what they Care about

HUAWEI Phones

As of July 1, 2021, Mac Form collected 492 reviews

of the HUAWEI P40 model phone.

With multiple choices available to respondents,

the relevant high-frequency words are shown in

Table 1.

As shown in the table, the most frequent words

are "appearance" and " running speed", indicating

that consumers were more concerned about the

appearance of the phone on the outside and the

speed of the phone on the inside. From the

comparison of the results of sentiment analysis,

consumers were more optimistic about these two

features.

Secondly, the emergence of the terms "Cost

performance" and "Hongmeng system" is also

noteworthy. Consumers who buy HUAWEI phones

were also more concerned about the

price/performance ratio of domestic phones and

whether to use the domestic operating system

Hongmeng, with a certain subjective atiitude and

personal preference.

4 CONCLUSION

This paper used natural language processing

technology to carry out machine learning through

CNN model, in order to evaluate the products on

e-commerce platform, which taked HUAWEI P40

model mobile phone in HUAWEI Mall as an

example. The relative reviews were classified and

scored by tags, finally review sentiment analysis is

conducted. The sentiment analysis initially reveal

consumers' sentiment attitude towards the phone.

There are two conclusions and recommendations.

(1) Consumers focus on 6 aspects of

HUAWEI phones, such as "appearance", "running

speed" and "photography", and the proportion of

these three items is greater than 50%, indicating that

consumer preferences are more convergent. Mobile

phone manufacturers can improve and stabilize the

performance of these three aspects to achieve a

better user experience.

(2) Consumers' preferences for HUAWEI P40

mobile phone's 6 performance items

all showed a positive trend, indicating that users'

trust and love of this phone were high. This

phenomenn indicates that users had a high degree of

trust and affection for this phone, and the user

stickiness was high at this time. Producers can

accelerate the development of new byproducts,

which is to further expand the market and enhance

consumers' purchasing power.

The limitation of this paper is mainly the

limitation of the analysis sample.

The majority of reviews were filled in before

consumers had fully experienced the product

features, and it is possible that with the passage of

time, the consumer attitudes might still fluctuate

over time. Secondly, the survey sample was limited

by the number of HUAWEI P40 users, which made

it difficult to accurately collect questions, the

questionnaire respondents were HUAWEI mobile

phone users. Although it can largely represent the

attitude of HUAWEI product consumers towards

HUAWEI mobile phones, the classification of this

label was still not precise enough and may not

reflect 100% of consumers' concerns about the

Sentiment Analysis of Product Reviews on E-Commerce Platforms Based on NLP: Take HUAWEI P40 as an Example

805

HUAWEI P40. With the further promotion of

HUAWEI P40 mobile phone in the market, we can

use more diversified e-commerce platforms to

collect different samples for analysis in a stratified

manner.

REFERENCES

A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A

N. Gomez, L Kaiser, I Polosukhin. Attention Is All

You Need[D].2017.

Feng XZ. Sentiment Analysis of Commodity Reviews

Based on Deep Neural Network[D]. Beijing

University Of Posts And Telecommunications,2019.

Li J Q. Sentiment Analysis Research and Application of

Online Shopping Review Data [D]. Chengdu

University of Technology,2019.

Rush, Alexander. The Annotated Transformer[A].

Association for Computational Linguistics.

Proceedings of Workshop for NLP Open Source

Software (NLP-OSS) [C]. Melbourne, Australia.

2018.52–60.

TAO Y C,ZHANG X Q, SHI L,WEI L.Ｒesearch on

Multi-feature Fusion Method for Short Text Sentiment

Analysis [J]. Journal of Chinese Computer

Systems,2020,41(06):1126-1132.

T Young, D Hazarika, S Poria, E Cambria. Recent Trends

in Deep Learning Based Natural Language

Processing[D].2018.

ZENG X Q, YU H. Sentiment Analysis of Merchandises’

Comment Based on Python [J]. Computer Knowledge

and Technology ,2020,16(08):181-183.

Ziang X, Sida I. Wang, J Li, D Lévy, A Nie, D Jurafsky, A

Y. Ng. Data Noising as Smoothing in Neural Network

Language Models[D].2017.

Z Z Lan, M D Chen, S Goodman, K Gimpel, P Sharma, R

Soricut. ALBERT: A Lite BERT for Self-supervised

Learning of Language Representations[D].2020.

ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology

806