Semantic Analysis of Online Initial and Additional Textual Reviews

of Pharmaceutical Products

Yumei Luo

, Jin Zhang

, Changlin Cao

, YaWen Zhong

and Qiongwei Ye

Department of Management Science, Yunnan University, Kunming, China

Guangzhou Institute for Modern Industrial Development in Greater Bay Area, Guangzhou, China

Business School, Yunnan University of Finance and Economics, Kunming, China

Keywords: Pharmaceutical Products, Initial Reviews, Additional Reviews, Semantic Analysis.

Abstract: In this paper, we selected the textual reviews of drugs on Ali Health Pharmacy as sample data. Firstly, the

text was pre-processed, including data cleaning, word separation, lexical annotation and removal of

deactivated words, in order to lay the data foundation for the subsequent analysis process. Secondly, we

conducted semantic analysis on the processed review texts, including word frequency analysis, LDA topic

clustering analysis and word frequency co-occurrence analysis. The results showed that the overall

distribution of semantic features in the text content of initial and additional comments was similar, but there

were differences in the attention of consumers to different topics in initial and additional comments.

1 INTRODUCTION

More and more consumers making their purchase

decisions based on online reviews. Additional

reviews as a new form of online reviews, contained

the product information that is somewhat different

from consumers' initial reviews in both richness and

objective veracity (Wang et al. 2015), it is necessary

to conduct a comparative analysis of initial and

additional reviews. Text mining of the review content

of this kind of online reviews with no fixed form can

help to better understand the large amount of

information contained in the review text. With the

development of the economy, the popularization of

medical knowledge and the impact of the new

pneumonia epidemic, consumers are increasingly

inclined to buy common and safe drugs on the

Internet. The pharmaceutical products involved in the

research sample of this paper are all over-the-counter

drugs.

The mining and research on online reviews of

pharmaceutical products, as a special product, need

yet to be strengthened (Terra and Clarke,2003). The

safety of pharmaceutical products is directly related

to the safety of consumers' lives, and consumers will

be more cautious in making purchase decisions, and

online reviews of pharmaceutical products will play a

more important role in consumers' purchase

decisions. Therefore, in this paper, we use text mining

analysis to study the pharmaceutical products’

semantic aspects of online initial and additional

reviews. On the basis of word frequency analysis, this

paper further conducts cluster analysis to analyse the

topics that consumers pay attention to medical

products. In order to better grasp the semantics of

short texts, in addition to using LDA theme analysis,

this paper also conducts word frequency co-

occurrence analysis to better highlight the relevance

between words.

2 LITERATURE REVIEW

2.1 Comparative Study of Initial and

Additional Reviews

Many researchers pay attention to text reviews, Zhou

and Li (2017) and Wang and Zhou (2016) used

questionnaires to study the impact of consistency and

ambiguity between initial and additional comments

on consumer information adoption. Sun and Li (2017)

and Li and Chen (2016) explored the effect of

different additional forms of reviews and word-of-

mouth emotional direction on purchase intention

through a questionnaire study of college students; Hu

and Ning (2017) investigated the usefulness of online

initial and follow-up reviews based on time interval

and product type by subjects completing

Luo, Y., Zhang, J., Cao, C., Zhong, Y. and Ye, Q.

Semantic Analysis of Online Initial and Additional Textual Reviews of Pharmaceutical Products.

DOI: 10.5220/0012044000003612

In Proceedings of the 3rd International Symposium on Automation, Information and Computing (ISAIC 2022), pages 715-722

ISBN: 978-989-758-622-4; ISSN: 2975-9463

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

715

questionnaire scales. Zhang et al. (2020) crawled the

cell phone reviews on JingDong Mall as a research

sample and studied the association characteristics of

the time series of online users' review chasing

behaviour. Shi et al. (2016) selected review data on

Mall for a comparative study of online initial reviews

and online additional reviews in terms of product

type, product price, number of reviews, review

length, review interval, and review sentiment

intensity.

Therefore, more and more scholars have begun to

pay attention to the comparative analysis of the initial

reviews and additional reviews, but most of them

have used questionnaires to investigate the impact of

questionnaires on consumer satisfaction, purchase

intention, and the usefulness of reviews, etc., and to a

large extent, a large amount of online data has been

ignored.

2.2 Text Analytics Research

For the mining analysis of text content, Blei and

Jordan (2003) proposed the thematic mining model

LDA (latent Dirichlet allocation) model in 2003.

Jiang et al. (2013) proposed an associative LDA

model for opinion mining problem domains and

applied it to user online reviews. Zhang and Qiao

(2020) established a text mining model based on

Bayesian correlation principles, completed the

mining of user comments, and calculated the

corresponding sentiment values. Liu and Jian (2018)

completed the data mining of Pecan fruit user reviews

on the Tmall website by establishing the user's

emotional index. Wen and Ding (2010) completed the

judgment of the user's product review sentiment and

calculated the related emotional intensity. Chatterjee

et al. (2020) use text mining techniques, machine

learning, and econometric techniques to discover

those sentiments that are more important in reflecting

and predicting customer satisfaction. Zhang et al.

(2020) used analytic methods such as word frequency

co-occurrence analysis for text mining and analysis of

online reviews. Yang et al. (2020) combined a

sentiment dictionary in the field of cigarettes to

analyze the word frequency of their high-frequency

sentiment words in order to compare two different

brands of cigarettes. Qiu and Sun (2021) used word

frequency analysis methods to find popular majors in

the three sample colleges, using ERNIE Pre-trained

models make further sentiment classifications of

comment data.

At present, domestic and foreign online review

research involves many business fields such as

movies, finance, hotels, social networks, etc., and the

research content involved in network review mining

is also very extensive, scholars mainly have more

research on the number, length and polarity of online

reviews, while the comparative analysis of the initial

evaluation and the lack of online user reviews in the

field of medical e-commerce drugs are relatively

lacking. Therefore, this paper takes the comment

content of related drugs on Alibaba Health Pharmacy

as a sample to conduct semantic analysis of the initial

post-evaluation, which not only enriches the

theoretical research of pharmaceutical e-commerce

and the comparative analysis of the initial and

additional comments, but also enriches the

application scope of text mining and analysis

technology.

3 RESEARCH MODELS AND

DATA PROCESSING

3.1 Research Model

Figure 1: Research model.

According to the research model in Figure1, it can be

seen that first, how to pre-process the text, including

data cleaning, word segmentation, lexical annotation,

removal of stop words, etc., to lay a data foundation

for the relevant analysis process of the following text;

Secondly, the semantic analysis of the processed

comment content includes: word frequency analysis,

LDA topic cluster analysis, and word frequency co-

occurrence analysis.

3.2 Data Sources and Pre-Processing

According to the data, more than 40% of Chinese

consumers will choose to buy drugs on Alibaba

Health, and the top 5 usage rates of China's

pharmaceutical e-commerce platforms in the first half

of 2021 are shown in Figure2.

ISAIC 2022 - International Symposium on Automation, Information and Computing

716

Figure 2: Top 5 in the first half of 2021 for the use of

China's pharmaceutical e-commerce platforms.

Therefore, this article chooses Ali Health as a

research platform, this article mainly through Python

crawler technology to crawl the online user review

information data of 320 drugs in the pharmaceutical

e-commerce plat-form Ali Health Pharmacy, and

there are 242 kinds of products left after removing

duplicate products. After screening and eliminating

meaningless comments such as automatic replies, a

total of 1249 initial reviews and 3667 additional

reviews were left, of which the existence of

posthumous reviews was used as the filter, and a large

number of comment data that were only

posthumously commented on were deleted.

The data is then pre-processed, including data

cleaning, word segmentation, part-of-speech

annotation, and removal of manual annotations of

stop words, in order to lay the data foundation for

subsequent analysis.

3.2.1 Participle and Lexical Annotation

This article uses the jieba package in the Python

language to segment each comment data, and the

word segmentation supports three modes, namely

precision mode, full mode, and search engine mode.

Among them, the precision mode tries to make the

most accurate segmentation of the sentence, which is

suitable for text analysis. Because this article requires

text analysis of the review data of pharmaceutical

products, the precise mode in the stutter participle is

selected to classify the comment text with participles

and part-of-speech labels.

3.2.2 Remove Stop-Words

In natural language processing, in order to improve

the efficiency of storage utilization and the efficiency

of computational analysis, certain words or words are

filtered out as needed before or after text analysis, and

these filtered out are stop words. This article collates

the commonly used Chinese stop words and adds

some stop words according to the corpus of comment

data to be analysed, a total of 786, some of which are

stopped words as shown in Table 1 below: After

removing the stop words and word segmentation from

the comment data, provide corpus for the following

analysis.

Table 1: Stop Words (partial).

①

a few u

thin

②

some down

ottle

③

all unde

eat one time

4 ” One still drin

few da

、

an still afterwards an

4 SEMANTIC ANALYSIS

4.1 Word Frequency Analysis

In this paper, after using Python's Jieba package to

segment the comment data, the frequency of the

vocabulary is counted, the high-frequency knowledge

meta words in the initial comment and the additional

comment are extracted, the synonyms are merged,

and the Wordcloud2 package is used to map the

vocabulary cloud.

Figure 3: Word cloud map for initial reviews.

Semantic Analysis of Online Initial and Additional Textual Reviews of Pharmaceutical Products

717

Figure 4: Word cloud map for additional reviews.

From Figure 3 and Figure 4 above, it can be found

that the online initial reviews and additional reviews

of drugs on the Alibaba Health platform, the overall

concern of consumers includes product features such

as "effect", "price" and "quality" of drugs, and the

attention to service features such as "customer

service" and "logistics" of drugs is also high. At the

same time, it can also be seen that consumers will

compare and analyse similar products on "Ali Health"

and "Tmall International". In this paper, the top 10

high-frequency knowledge meta-vocabularies in the

initial reviews and the additional review are selected

to make word frequency statistics table, as shown in

Table 2 and Table 3 below:

Table 2: Word frequency statistics for initial comments.

Ranki

Entry

Freque

ncy

Rankin

Entry

Freque

ncy

1 Effect 247 6 Problem 68

Genuine

product

84 7 Price 45

3 Products 83 8

Promoti

ons

4 Logistics 82 9 Quality 43

Customer

Service

70 10 Taste 43

Table 3: Word frequency statistics for additional comments.

Rank

ing

Entry

Freque

ncy

Ranki

Entry

Freq

uenc

1 Effect 200 6

Customer

Service

2 Products 71 7 Quality 38

3 Problem 59 8 Logistics 31

Genuine

product

48 9

Take

Medicine

5 Taste 39 10 Reviews 25

Further divide "effect", "appearance", "price" and

"date" into product characteristics of pharmaceutical

products, divide "customer service" and "logistics"

into service characteristics of pharmaceutical

products, analyze the similarity and difference of

semantic features of online users' initial reviews and

additional comments, and select the top 100 high-

frequency knowledge meta-vocabularies

respectively, and the statistical results of the two are

shown in Table 4 below:

Table 4: Comparison of high-frequency vocabulary in the

initial and additional reviews.

Comment

type

Classify

Products

Logis

tics

Custo

mer

Service

Effect

Appea

rance

Price Date

Initial

comments

23.73

16.95

5.08%

13.56

23.73%

Additional

comments

37.10

14.52

12.90

3.23%

11.29

20.97%

4.2 LDA Topic Analysis

For unstructured text content like comments,

computers cannot directly understand, we need to

introduce text mining to help us extract high-value

information hidden in text descriptions. In natural

language processing, LDA is an important generative

model that can be used to identify hidden topic

information in large scale documents or corpora

(Suominen and Toivanen 2016; Shan and Li 2010).

The LDA uses the bag-of-words approach and a joint

distribution to calculate the conditional distribution of

hidden variables under a given observable variable.

Observable variables are a set of words, and latent

variables are subjects (Lyu et al. 2022). The graphical

representation of LDA model is shown in Figure5.

Figure 5: Graphical model representation of LDA.

α and β each control a Dirichlet distribution to

obtain the distribution of words under different

topics, so it is important to determine the number of

topics in the LDA model (Lamba and

ISAIC 2022 - International Symposium on Automation, Information and Computing

718

Madhusudhan,2019; Kushkowski et al. 2020). This

article selects the optimal number of topics by calling

the sklearn package in Python, performing

unsupervised topic clustering analysis for initial

comments and additional comments, respectively,

and through topic confusion, GridSearchCV grid

search and cross-validation, and visualization.

Figure 6:

Perplexity trend of LDA for initial comments.

Figure 7:

Perplexity trend of LDA for additional comments.

As can be seen from the above Figure6 and

Figure7, in the case of not considering the number of

topics of 1, whether for the initial comment or

additional comment, the LDA model has the lowest

degree of topic confusion when the number of topics

is 3, that is, the optimal number of topics at this time

is 3. In order to further verify the optimal number of

topics, this paper also uses grid search and cross-

validation to tune the model. The number of topics is

set to 3-9 when tuning this model, and the optimal

number of topics is searched in turn, and the data is

cross-validated in five folds, and the parameters of the

optimal model obtained are shown below:

Figure 8: The best parameters of the LDA model for initial

comments.

Figure 9: The best parameters of the LDA model for

additional comments.

It can also be clearly seen from the above Figure8

and Figure9 that the parameters in the optimal model

n components = 3, which is a further proof that for

online reviews of pharmaceutical products, whether it

is a first review or an additional review, the focus of

consumer attention is on three major themes. We

evaluated models with different numbers of topics (1

to 9 topics) on preliminary and post-evaluation data,

respectively, with the help of pyLDAvis in Python

Tool (Sievert and Shirley 2014) to visualize each

model. When the number of topics is 3, there are

relatively few connections be-tween the topics, which

means that the models are relatively independent and

have little to do with each other, indicating that the

choice of the number of topics in this article is

appropriate. However, the words of the LDA topic

model are discrete, and taking the preliminary

evaluation data as an example, as shown in Figure10.

Figure 10: Example diagram of LDA model visualization.

It is relatively inaccurate to determine what each

type of topic represents and which ones are included

under each topic Characteristic words, therefore, it is

necessary to do further word frequency co-occurrence

analysis on the comment text data, and word

frequency co-occurrence analysis pays more attention

to the connection between words than LDA model

analysis, which can more clearly show the connection

between subject words.

Semantic Analysis of Online Initial and Additional Textual Reviews of Pharmaceutical Products

719

4.3 Word Frequency Co-Occurrence

Analysis

In this article, the initial and additional comments are

sorted by the number of word frequencies, and the top

1,000 are selected to fit the number of word

frequencies, and the results are shown in the

following figure:

As can be seen from Figure11 and Figure12, the

number of word frequencies for initial comments and

additional comments conforms to a stricter power

index distribution, and the goodness of fit of the two

is 0.92 and 0.97, respectively. In this paper, the words

in the top 100 of the word frequency quantity are

selected for word frequency co-occurrence analysis,

and the high-frequency knowledge meta-vocabulary

is selected It is to remove words that have only

appeared once or twice, to eliminate some of the noise

in the corpus, and to make high-frequency words

more representative and more effective in reflecting

the information contained in the comments.

Figure 11: Word frequency count fitting for initial reviews.

Figure 12: Word frequency count fitting for additional

reviews.

Word frequency co-occurrence analysis is an

unsupervised cluster analysis. It is generally believed

that the more words appear in the same document, the

more closely related the two topics are. According to

the literature, there are many ways to calculate the

similarity between words, such as Pointwise Mutual

Information, Likelihood ratio, Average Mutual

Information and so on (Terra and Clarke 2003). This

paper calculates the semantic similarity between

high-frequency words in word frequency analysis

based on the Pointwise Mutual Information, that is,

the PMI value, so as to achieve co-occurrence

between words. This article uses Gephi to visualize

the word frequency co-occurrence matrix, as shown

in the following Figure13 and Figure14.

Figure 13: Initial comments word frequency co-occurrence.

Figure 14: Additional comments word frequency co-

occurrence.

As can be clearly seen from Figures 13 and 14

above, for unsupervised clustering analysis, the

model automatically divides high-frequency words

ISAIC 2022 - International Symposium on Automation, Information and Computing

720

into 3 major themes according to the co-occurrence

relationship between words. Each color represents a

theme, and the co-occurrence relationship between

words of the same color is stronger, and they belong

to the same theme. The blue and orange themes in the

preliminary review are also the themes under the

theme of product characteristics, and the distribution

of topics in the post-evaluation is similar to the above,

so the description below this article focuses on the

analysis of the three themes of "effect", "customer

service" and "logistics". The larger the node, the

higher the centrality, the more frequently it appears in

the comment, and the more vocabulary it is connected

to. For the theme of "effect", the size of the node in

the post-evaluation is significantly larger than that of

the preliminary evaluation, and the words associated

with it are also greater than the preliminary evaluation

in the post-evaluation, and in the preliminary

evaluation, for the special commodity of the drug,

consumers are more likely to hold a try-and-see

attitude, and will only be in the post-evaluation after

a period of use Further evaluation of the efficacy of

the drug, so that the attention to the topic of "effect"

in the follow-up evaluation has increased

significantly; For the two major themes of "logistics"

and "customer service", the node size in the post-

evaluation is significantly smaller than that of the

preliminary evaluation, and the "logistics" theme

word is closely related to the words "delivery" and

"attitude", and the keywords such as "after-sales

service", "store" and "problem" are closely related to

the theme word "customer service". This shows that

consumers' focus on pharmaceutical products is

reflected in both online initial and additional reviews,

and there are both similarities and differences

between the two.

5 CONCLUSIONS

Combined with the above figure and the above table,

coupled with the above series of analysis, it can be

concluded that the first comment and the additional

comment have the similarity of the overall trend in

the distribution of semantic features of the text

content, that is, the description of the product

accounts for the highest proportion, the description of

logistics accounts for the lowest, the description of

customer service is centered. Consumers' attention to

pharmaceutical products is concentrated in the three

themes of "effect", "customer service" and

"logistics". From the perspective of difference, the

additional content focuses more on the effect

experience of drugs. Compared with the initial

comment, the proportion of additional comments in

other aspects is lower than the initial comment, and

the attention to other aspects has decreased to a

certain extent. This paper reveals the as-sociations

and differences in semantic features of the initial and

additional reviews, and performs a visual analysis of

word frequency co-occurrence of high-frequency

vocabulary, and points out feature words that are

closely related to different topics.

It should be pointed out that there are certain

limitations in this paper, first of all, the medical

product review data selected in this paper is relatively

small, and future studies can verify whether there is a

consistent conclusion in the context of big data.

Secondly, this article only conducts textual mining on

the semantic aspects of the initial review of online

reviews, and subsequent research can pay more

attention to the emotions contained in the comment

text. Finally, we hope that this research will help

pharmaceutical e-commerce platforms to discover

consumers' concerns about pharmaceutical products

and make corresponding adjustments. Make the most

of your website and improve the quality of its services

while helping them better manage customer reviews.

ACKNOWLEDGEMENTS

This research was also supported by Kunming E-

commerce and Internet Finance R&D Center

(KEIRDC[2020]), the Prominent Educator Program

(Yunnan [2018]11), Yunnan Province Young

Academic and Technical Leader candidate Program

(2018HB027) & Yunnan Provincial E-Business

Entrepreneur Innovation Interactive Space

(2017DS012), National Natural Science Foundation

of China [71402159], Yunnan University Eastland

Yong and Middle-aged Backbone Teachers Training

Program (2018), 2021 Yunnan Provincial Academic

Degrees Committee Program

[SJZYXWALK202113], 2021 Yunnan University

Program [XJKCSZ202117].

REFERENCES

Blei, D. M., & Jordan, M. I. (2003). Modeling annotated

data ACM SIGIR FORUM,

Chatterjee, S., Goyal, D., Prakash, A., & Sharma, J. (2020).

Exploring healthcare/health-product ecommerce

satisfaction: A text mining and machine learning

application. Journal of Business research, 131(7), 815-

825.

Semantic Analysis of Online Initial and Additional Textual Reviews of Pharmaceutical Products

721

Ghose, A., & Ipeirotis, P. G. (2007). Designing Novel

Review Ranking Systems: Predicting Usefulness and

Impact of Re-views. International Conference on

Electronic Commerce,

Hu, C., & Ning, C. (2017). When is online post-evaluation

more useful than preliminary evaluation? -Analysis of

regulatory effects based on time interval and product

type. Prediction, 36(04), 36-42.

Hu, N., Koh, N., & Reddy, S. K. (2014). Ratings lead you

to the product, reviews help you clinch it? The

mediating role of online review sentiments on product

sales[J]. , 2014. Decision Support Systems (57), 42-53.

Jiang, W., Zhang, L., DaiI, Y., Jiang, J., & Wang, G. (2013).

Analysis on the Usefulness of Online Reviews For User

Needs. Chinese Journal of Computing Science, 36(01),

119-131.

Kushkowski, J. D., Shrader, C. B., Anderson, M. H., &

White, R. E. (2020). Information flows and topic

modeling in corporate governance. Journal of

Documentation, 76(6), 1313–1339.

Lamba, M., & Madhusudhan, M. (2019). Mapping of topics

in DESIDOC Journal of Library and Information

Technology, India: A study. Scientometrics, 120(2),

477–505.

Li, X., & Chen, Y. (2016). The Influence of Word-of-

Mouth Supplement Form on Purchase Intention: The

Moderating Effect of Word-of-Mouth Direction.

Psychological Journal, 48(06), 722-732.

Lyu, Y., Yin, M., Xi, F., & Hu, X. (2022). Progress and

Knowledge Transfer from Science to Technology in the

Research Frontier of CRISPR Based on the LDA

Model. Journal of Data and Information Science, 7(1),

1-19.

Qiu, J., & Sun, Y. (2021). Online Review Research on

University Environment Based on Text Mining. Library

Theory and Practice(5), 16-22.

Shan, B., & Li, F. (2010). A survey of topic evolution based

on LDA (in Chinese). Journal of Chinese Information

Processing, 24(06), 43-49+68.

Shi, W., Lu, W., Na, S., & Cai, J. (2018). A Comparative

Study into the Impact of Initial and Follow-on Online

Comments on Sales. Management Review, 30(01),

144-153.

Shi, W., Xue, G., Zhang, Q., & Wang, L. (2016). A

Comparative Study of Online Initial Review and Online

Additional Review. Management Science, 29(04), 45-

58.

Sievert, C., & Shirley, K. (2014). LDAvis: A method for

visualizing and interpreting topics. In Proceedings of

the workshop on interactive language learning,

visualization, and interfaces,

Sun, R., & Li, X. (2017). Research on the Influence of

Contradictory Post-evaluation on Consumers' Purchase

Intention. Geomatics and Philosophy and Social

Sciences of Wuhan University, 70(01), 75-86.

Suominen, A., & Toivanen, H. (2016). Map of science with

topic modeling: Comparison of unsupervised learning

and human-assigned subject classification. Journal of

the Association for Information Science and

Technology, 67(10-), 2464–2476.

Terra, E. L., & Clarke, C. L. A. (2003). Frequency estimates

for statistical word similarity measures[C]. Proceedings

of the 2003 human language technology conference of

the North American Chapter of the Association for

Computational Linguistics,

Wang, C., He, S., & Wang, K. (2015). Research on How

Additional Review Affects Perceived Usefulness of

Review. Journal of Management Sciences, 28(3), 102-

114.

Wang, J., & Zhou, S. (2016). Research on the Influence of

Consistent and Contradictory Online Reviews on

Consumer Information Adoption: The Mediating Role

of Perception Usefulness and the Moderating Role of

Self-Efficacy Based on Perception. Library and

Information Services (22), 94-101.

Wen, N., & Ding, S. (2010). A Review of Tendency

Analysis of Subjective Product Review Information.

Journal of Intelligence, 29(12), 70-74+48.

Yang, C., Zhang, H., Huang, J., & Wan, J. (2020). Text

sentiment analysis of cigarette online reviews. Chinese

Journal of Tobacco, 26(02), 92-100.

Zhang, N., & Qiao, D. (2020). Research on Sentiment

Analysis of Online Learning Review Based on Deep

Learning. Journal of Henan University of Urban

Construction, 29(04), 63-71+92.

Zhang, Y., Wang, Y., Peng, L., & Liu, Y. (2020). Research

on The Content Intelligence of Online User Additional

Comments Based on Text Mining: A Case Study of

Jingdong Mall Mobile Phone Review Data. Modern

Intelligence, 40(09), 96-105.

Zhou, H., & Li, S. (2017). The Impact of Online Additional

Comments on Consumers’ Information Adoption.

Sociology Mind, 07(02), 60-71.

ISAIC 2022 - International Symposium on Automation, Information and Computing

722