New Generation Computing

https://doi.org/10.1007/s00354-024-00242-9

The Impact of Sentiment Scores Extracted from Product

Descriptions on Customer Purchase Intention

Yi Sun1 · Kaira Sekiguchi1 · Yukio Ohsawa1

Received: 20 August 2023 / Accepted: 22 November 2023

© The Author(s) 2024

Abstract
This study investigates whether and how the textual content of product descriptions,
especially the sentiment element, influences buyers’ purchase intentions. Using year-
round digital transaction data from Mercari, a leading e-Commerce platform in Japan,
we examine the interplay of hard and soft information signals exchanged between
sellers and buyers. The study addresses two crucial questions: (1) Do the descriptions
that sellers provide on product sales pages impact the buyer’s intent to purchase? and
(2) In what way does the description influence the buyer’s purchase intention? Quan-
titative analysis is used to understand the relationship between product descriptions,
sentiment elements, and purchase intentions. The results show that sentiment factors
in product descriptions can serve as high-quality “signals” that can help buyers make
informed purchasing decisions and reduce information asymmetry between buyers and
sellers. This research contributes to understanding decision-making in online markets,
particularly the role of soft information and sentiment analysis.

Keywords Purchase intentions · Sentiment analysis · Product descriptions ·

e-Commerce

B Yi Sun
[email protected]
Kaira Sekiguchi
[email protected]
Yukio Ohsawa
[email protected]

1 Department of Systems Innovation, School of Engineering, The University of Tokyo, 7-3-1 Hongo,
Bunkyo-ku, Tokyo 1138654, Japan

0123456789().: V,-vol 123

 New Generation Computing

1 Introduction

Currently, sustainable development has become a pressing need, reflecting not only
economic development but also the social and environmental dimensions of develop-
ment. The Sustainable Development Goals (SDG) of the United Nations emphasize
the need for a more balanced approach to address various social, economic, and envi-
ronmental issues [1].
A common thread that runs through these dimensions is the promotion of trust.
We believe that this is the key to sustainable development. Business transactions,
especially in the context of customer-to-customer (C2C) transactions, have been the
subject of various studies with a wide variety of perspectives [2]. However, the role
of trust in business has not been fully elucidated even in the C2C domain [3, 4].
In recent years, C2C platforms, especially e-Commerce marketplaces selling used
products, have received more attention. Under the influence of the COVID-19 pan-
demic, more and more people are choosing to buy products on used product trading
platforms [5]. Furthermore, e-Commerce platforms are now more than just a place to
sell second-hand products; more users are becoming accustomed to selling items on
this platform that are difficult to purchase offline, such as masks that have not been
seen in offline stores for some time [6].
Therefore, our research aims to mathematically reveal the role of trust in commerce
and explore how machine learning can help build trust, facilitating more sustainable
commerce in the process. In this study, we specifically explore a rarely discussed
aspect of C2C transactions, the sentiment elements conveyed by sellers in product
descriptions, and their impact on buyer intent [7, 8].
To facilitate informed purchasing decisions and successful transactions on e-
Commerce platforms, it is crucial to reduce the information asymmetry between buyers
and sellers [9, 10]. In the context of e-Commerce platforms, this usually means that the
seller has complete information on the product being sold, while the buyer has to decide
whether or not to buy the product, knowing nothing about the product except some
information provided by the seller and other buyers [11]. This information is called
"signaling" [12]. One way to reduce information asymmetry is to use high-quality
’signals’ to help buyers make a purchase decision. For example, online reviews pro-
vided by other buyers are an often studied signal that provides information about the
product and reflects the customer’s experience with the product [13, 14]. Research has
shown that the number and sentiment of online reviews can influence sales [15, 16]. In
the context of C2C marketplaces, product descriptions created by sellers are another
signal that has received less attention in previous research [17, 18]. Chang et al. [19]
examined the impact of the quality of the product posts on consumers on the second-
hand marketplace on Facebook through a questionnaire survey. Their study found, for
example, that the completeness of product information, the accuracy of information,
the aesthetics of the posts, and their popularity had significant effects on customers’
purchase intentions to varying degrees. However, their study did not analyze the text
of the posts themselves.

123
New Generation Computing

2 Literature Review and Conceptual Framework

This study draws on research in three areas: (1) signaling theory and its impact on
product sales; (2) the impact of text information on product sales; and (3) sentiment
analysis in text.

2.1 Signaling Theory and Its Impact on Product Sales

In e-Commerce transactions, three parties are typically involved: the e-Commerce

platform, the seller, and the buyer. A prevalent feature of these interactions is that
these parties do not always share the same level of information about products and
the market. The variation in access and interpretation of product-related information
by these parties leads to an imbalance known as information asymmetry. One main
approach to mitigate this imbalance is the application of signaling theory [9, 12,
20]. This theory focuses on various types of signal that function as tools to mitigate
information asymmetry, enabling buyers to discern the quality of the product more
accurately. In particular, quality signals can significantly shape a buyer’s perception,
foster trust, and improve purchase intention, thus influencing product sales [9, 10].
Previous research indicates that a signal possesses two essential characteristics:
its observability and its cost [12]. In an e-Commerce environment, where buyers
cannot physically evaluate the product, they often depend on quality signals to assess
the quality of the product [20–22]. If sellers can provide reliable quality signals, it
helps buyers reduce information asymmetry, facilitates trust-building, and improves
purchase intention. For example, [20] illustrate that a higher quality website, which
acts as a signal, increases the probability of earning trust, stimulating purchase intent,
and securing a premium price from potential buyers.
Signals from e-Commerce can generally be classified into hard information and soft
information. Hard information is information that can be easily quantified and repre-
sented numerically, such as the seller’s reputation system, which includes reputation
scores, star ratings, or platform recommendations [11]. Metrics such as the number of
likes and comments a product receives also constitute hard information, with several
studies affirming the substantial influence of these factors on a product’s popularity
and subsequent sales [23]. For example, [13] found that the higher the number of
likes a post receives on platforms such as Facebook, the greater the impact it has on
shaping potential consumers’ attitudes towards the product positively. Additionally,
[19] propose that sellers aim to increase the number of likes, comments, and shares
on platforms like Facebook’s second-hand marketplace. The higher these figures, the
greater the popularity of the post, thereby convincing potential buyers that the product
offered is a worthwhile purchase."
On the other hand, soft information encompasses data that are not readily quan-
tifiable and often require contextual understanding [11]. Common examples of soft
information include online reviews and product descriptions. Previous research indi-
cates that positive reviews from sellers can promote increased product sales [24–26].
Essentially, the more accurate and detailed the signals a seller can convey, the stronger

123
 New Generation Computing

their influence on trust-building. This, in turn, amplifies the buyer’s willingness to

make a purchase [3, 4, 27].

2.2 The Impact of Soft Information Signals on Product Sales

In e-Commerce research, previous studies have shown mixed results on the impact
of online reviews on product sales. Kim et al. [6] suggest that the influence of online
reviews on customer purchase decisions is altered by other signals, indicating a need
to study additional soft information signals. Although research has focused on product
descriptions in crowdfunding, there is a lack of attention paid to product descriptions
in the context of second-hand e-Commerce. Chang et al. [19] found that more com-
plete and accurate information in product sales posts leads to a higher probability of
purchasing. However, they did not analyze the text of the posts themselves. To address
this gap, researchers are exploring the use of machine-learning approaches, such as
topic modeling and sentiment analysis.
Traditionally, most research on soft information signaling has focused on online
reviews. However, the results of these studies are inconsistent and even vary within
the same product category. For example, the impact of online reviews on movie ticket
box office sales varies from study to study. Some findings suggest a positive correla-
tion, while other studies conclude that there is no significant effect [14, 15]. Kim [6]
argues that while the content of online reviews does influence customers’ purchasing
decisions, the extent of the influence is moderated by other signals. Specifically, when
customers have access to other signals that reduce uncertainty about product quality,
they rely less on online reviews. This, in turn, decreases the influence of online reviews
on customers, helping to explain the conflicting findings in previous research on the
impact of online reviews [6].
In contrast, textual information about product descriptions, another form of soft
information signaling, has received relatively little attention, with most existing
research focusing on the crowdfunding domain. In the context of crowdfunding web-
sites, [17] explored the importance of project descriptions for project success. They
identified three new signals from the project descriptions: length, readability, and
tone. They found that these signals improved the accuracy of project success predic-
tion models. In addition, they found that this new information had a significant impact
on the success rate [17]. Jiang et al. [18] also noted that themes extracted from project
descriptions significantly affect performance.
In the field of second-hand e-Commerce, [19] investigated the correlation between
buyer purchasing decisions and the quality of product sales posts on Facebook. Their
findings suggest that buyers are more inclined to pay attention to articles with com-
prehensive and accurate information, and therefore regard these articles as a reliable
source for purchasing products [19]. However, their study did not include an analysis
of the text of the posts themselves.

123
New Generation Computing

2.3 Methods of Text Analysis

This study offers insights into the application of machine-learning techniques in mar-
keting research, focusing on the extraction and analysis of sentiment from textual data.
With the advancement of digital technologies, researchers are using large amounts of
unstructured data, primarily textual data, to gain actionable insights. In this context,
topic modeling, widely recognized in the machine learning community, has been
conceptualized as the grouping of large amounts of discrete textual data to reveal
underlying patterns by aggregating semantically related words [28].
Blei et al. [28] further refined the premise of topic modeling by proposing the Latent
Dreichler Assignment (LDA) model. The model posits that each topic is represented as
a distribution of words in a document, and, conversely, each document is represented
as a distribution of different topics [29]. However, as with most approaches, topic
modeling faces a number of challenges. A prominent issue is the number of topics
manually specified, which has a critical impact on the validity of the model [30].
Several studies confirm this view [18, 31, 32].
In 2018, [33] took an innovative perspective by comparing thematic modeling with
physical systems. Koltkov advocates the inclusion of physical metrics such as entropy
calculations, arguing that they can enhance topic models. By integrating benevolent
entropy, a method was developed to autonomously determine the optimal number
of topics. The underlying principle of this approach is that the maximum actionable
information is obtained when the entropy associated with the number of topics (similar
to the temperature in a physical system) is minimized [34].
Although topic modeling offers a significant number of possibilities for textual
data review, it remains critical to discriminate methods that are appropriate for spe-
cific research needs. While recognizing the potential of topic modeling, this study
relies heavily on sentiment analysis to achieve its unique goals. Through the use of a
sentiment dictionary, this study reveals the subtle sentiment tone that permeates prod-
uct descriptions and other textual content on e-Commerce platforms. This approach
provides insight into recognizable emotions, both positive and negative, that can sig-
nificantly impact purchase decisions [35].
Sentiment is a subjective emotion that can influence an individual’s thoughts and
judgments [7, 8]. In the previous studies, textual data have often been used to extract
sentiments such as praise or criticism; these insights can be used by others as a reference
for decision-making [16]. As a result, sentiment analysis, especially for product or
service reviews, has gained considerable popularity [8, 16, 36]. The main method of
calculating sentiment scores is to utilize the word frequency of words in a text that
have been assigned a sentiment value in a predefined dictionary to derive a sentiment
score for the text. Some studies have aggregated positive and negative sentiments to
arrive at a final score [8], while others have calculated positive and negative sentiments
separately and examined their impact on the audience [18]. Some researchers have
even hypothesized that customer attitudes towards sentiments in texts change with
product quality [6].
Despite these advances, there is a clear research gap in analyzing the sentiment
of descriptions made by sellers in business settings. Product descriptions are one of

123
 New Generation Computing

Fig. 1 Signal exchanges between sellers and buyers, adapted from Sun [5]

the critical components of online product information. These pieces of text, carefully
crafted by sellers, provide crucial insights into a product’s features, usability, and
other salient details that are often the deciding factor for customers contemplating a
purchase. This research embarks on a journey to uncover the hidden signals within
these descriptions and their influence on consumers’ purchase decisions. Based on the
results of previous studies, this paper also uses sentiment scores computed through
word frequency to observe whether this variable will actually affect buyers’ purchasing
decisions, in which aims to shed light on an integral part of online consumer behavior.

2.4 Research Framework

The primary objective of this study is to explore the influence of product descrip-
tions and sentiment components, calculated based on word frequency, on customers’
purchase intentions. To accomplish this, the research scrutinizes product description
texts drawn from sales data collated from Mercari, one of the largest online shopping
platforms in Japan. Figure 1 provides a visual representation of the prospective signals
exchanged between sellers and buyers on the Japanese second-hand e-Commerce site
Mercari, a conceptual framework adapted from previous research [5].

The selling process begins with sellers listing their products on the platform, pro-
viding crucial details such as the product’s price, condition, accompanying photos,
and a written description. Customers receive these signals and, in response, provide
feedback in the form of ’likes’ and comments. This customer feedback, directed to the
seller, can serve as a reference point for other potential buyers considering the same
product.
In this setup, certain elements, such as the price of the product, its condition, the
number of ’likes’, and the number of comments, can be classified as hard information,
as they are quantifiable in numerical terms. On the contrary, the product description
and user comments are considered soft information due to their textual nature. These
soft signals often harbor the unspoken sentiments of both sellers and buyers. Fur-
thermore, product photos can encapsulate both hard and soft information; they carry

123
New Generation Computing

explicit digital information, but can also showcase contextual information that requires
interpretation. Buyers, upon receiving these signals, make a purchase decision based
on their interpretations.
The focus of this study lies in uncovering the hidden signals in the product
description texts written by sellers and their subsequent influence on buyer purchase
intentions, a subject scarcely discussed in previous research. To illustrate the impact
of product descriptions on customer behavior, this study aims to answer the following
questions: (1) Do the descriptions that sellers provide on product sales pages impact
the buyer’s intent to purchase? (2) If so, in what way does the description influence the
buyer’s purchase intention? To address these research objectives, the study employs
a quantitative methodology to dissect the relationship between product descriptions,
sentiment elements derived from word frequency, and purchase intentions. The find-
ings of this study have the potential to provide invaluable insights for online sellers
and marketers, allowing them to optimize product descriptions and improve customer
satisfaction.

3 Data and Methodology

3.1 Data

The main objective of this study is to explore the relationship between product descrip-
tions and customers’ purchase intentions. More specifically, our objective is to figure
out whether and to what extent the sentiments embedded in product descriptions influ-
ence customer behavior. For this purpose, we use product sales data from Mercari, one
of the largest online shopping platforms in Japan, known for its extensive marketplace
for second-hand goods.
We chose to focus on the sales of all figure-based products for several reasons. In
particular, the wide variety of figures available on Mercari includes limited-edition
items that are usually sold at high prices. For example, an item initially priced at
100,000 yen can subsequently sell for 300,000 yen or even 500,000 yen due to its rarity
and collector demand. In such a market environment, both buyers and sellers are likely
to invest a great deal of effort in shaping the transnational relationship between them,
making the figure category a rich field for studying the impact of product descriptions
and their sentiment elements.
Our data set contains more than 4.4 million transactions recorded over a 1-year
period, covering product sales from January through December of a given year over
the past 5 years. This robust data set encompasses all aspects of each sale, including
the product description, the price listed, the number of product reviews received, the
number of “likes” collected, and the condition of the product at the time of sale.
Using text-analytic techniques and econometric models, this study seeks to reveal
hidden patterns and insights that can help sellers maximize profits and build stronger
relationships with customers. Additionally, these findings are valuable to Mercari,
which is working to optimize its platform and improve the user experience for both
sellers and buyers.

123
 New Generation Computing

Fig. 2 Methodology of this research

3.2 Methodology

The research methodology used in this study is a combination of quantitative analysis

techniques and a comprehensive study of text data. Our approach can be depicted in
Fig. 2, which we will detail later.
1. Data Collection: This stage involves collecting sales data from the Mercari plat-
form, including product descriptions, product prices, number of ’likes’, and the
number of reviews. This data set serves as a basis for subsequent stages.
2. Data Pre-processing: We pre-processed the collected data to eliminate inconsis-
tencies and prepare them for further analysis. This includes segmenting product
descriptions into individual words and associating them with their respective sen-
timent scores, using the Japanese sentiment dictionary [35].
3. Text Analysis (Sentiment Score Calculation): Using the pre-processed data, we
conduct a textual analysis of the product descriptions. Each word in the product
description is assigned a sentiment score from 1 to 5, indicating its degree of
negative to positive. These scores are then used to calculate an overall sentiment
score for each product description.
4. Model Building: We build a logistic regression model using the calculated senti-
ment scores along with other hard information such as the number of ’likes’, the
price of the product, and the number of reviews. These variables serve to predict
the sale outcome (sold or canceled) for each product.
With our data set and the process ready, we proceed as follows.
Firstly, we employ a Japanese sentiment dictionary [35] produced by psychologists
to translate the sentiments encapsulated in the product descriptions. These descriptions,
crafted by the sellers of Mercari, offer valuable information on their attitudes, positive
or negative, towards the products. We use this dictionary to assign sentiment scores to
product descriptions, treating these scores as practical representations of the seller’s
attitude. This process can also be expressed as follows:

Sentiment Score = scorew , (1)
w

where scorew is the sentiment score for the wth word in the product.
In the next step, we use the sentiment scores obtained in Step 1 as variables in the
logistic regression, expressed in Equation 2
 
P(Sold = 1|x)
ln = α + β1 ∗ Countlike + β2 ∗ Count comment
1 − P(Sold = 1|x)
+ β3 ∗ Priceproduct + β4 ∗ Sentiment Score. (2)

123
New Generation Computing

In this model, other signals with hard information, such as the number of ’likes’,
the price of the product, and the number of reviews, have also been used. We list the
variables’ definitions used in the model in Table 1.
The dependent variable “sold” in our study is a binary variable that represents
the status of the sale of a product. In this representation, a value of 1 means that a
product is sold out, while a value of 0 means that the sale has been canceled. We
have intentionally excluded products that are still available for sale. The rationale
behind this choice is that the outcome of these sales has yet to be determined, creating
an element of uncertainty. As such, we cannot conclusively determine whether the
product description has any influence on these active sales or not.
In this investigation, we have chosen to use sentiment scores as an independent
variable within our model. This decision has been driven by our aim of investigating
whether hidden information within product descriptions, as revealed by sentiment
scores, has any bearing on product sales. Accordingly, to verify whether sentiment
scores have a more pronounced effect on sales, we also utilize as independent variables
certain signals that contain ’hard information’ and are displayed on the product sales
page. These signals include, for example, the total number of ’likes’ given by potential
buyers, the total number of comments left by these buyers, and the price listed of the
product.
To comprehensively examine the effect of varying price points on our findings, we
segment the data and apply a logistic regression model at multiple price criteria. These
criteria involve systematically removing data below a set price threshold, starting at
intervals of 1,000 yen between 1,000 yen and 100,000 yen, and then shifting to intervals
of 10,000 yen between 100,000 yen and 200,000 yen. We have chosen not to segment
further beyond the 200,000 yen mark. The motivation behind this methodology is
to ensure an adequately sized sample for each segment, thus ensuring the statistical
robustness of our findings.
To evaluate our hypothesis in a more rigorous manner, we employ the technique of
K-Fold cross-validation. This approach aids in mitigating the risk of over-fitting our
model to the data. For the purposes of this study, we set the value of K at 5. This choice
balances the need for a thorough exploration of the data, while avoiding over-fitting,
thereby ensuring the reliability of our conclusions.

4 Empirical Result

Table 2 presents the results of the descriptive statistical analysis of the independent
and dependent variables in this study. The median sentiment score derived from the
text analysis was 18.88 and the mean was 26.85. Higher scores indicate more positive
sentiments embedded in the text. Understanding the sentiment score is critical, because
it provides an empirical way to understand customer sentiment. Higher sentiment
scores indicate greater acceptance of a product or service, which can inform marketing
strategies and potentially increase customer engagement.
When examining data related to product pricing, the median price was determined
to be 2280 yen, indicating that half of the products fall into the low price range.
However, the average price was as high as 21,142 yen, suggesting that the data set

123
123
Table 1 Definition of variables
Type Variable Definition

Dependent variable Sold A dummy variable 1 if the product d is in the state of ’sold’ and 0 if the product d is in the state of ’cancel’
Independent Variable Countlike The total number of ’likes’ that potential buyers clicked
Countcomment The total number of comments that potential buyers made
Priceproduct Product price
Sentiment score The sentiment score extracted from the product description
New Generation Computing
New Generation Computing

Table 2 Descriptive statistics

Variables Min Median Mean Max Std

Sentiment score 0 18.88 26.85 385.42 28.81

Price 300 2280 21,142 9,999,999 380671.44
Num of likes 0 1 2.30 357 3.29
Num of comments 0 0 1.16 479 2.61
Sold 0 1 0.61 1 0.49

contained high-priced products that deviated from the average price. These findings
emphasize the heterogeneity of product price ranges and the existence of high-end
products. Sellers can use this information to cater to different customer segments and
refine pricing strategies to maximize revenue potential.
In terms of buyer-seller interaction metrics, the median and mean number of “likes”
were 1 and 2.3, respectively. On the contrary, the median and mean number of reviews
were 0 and 1.16, respectively. These data suggest that a large portion of product buyers
are not actively interacting with sellers. From a business perspective, this finding
suggests that there is an opportunity to encourage customers to be more actively
engaged, making it possible to develop strategies that promote positive interactions
between buyers and sellers, thereby increasing customer satisfaction and loyalty.
To better understand sentiment scores and their inherent attributes, we logarithmi-
cally transformed the data and integrated them into a logistic regression model, as
shown in Fig. 3.
The variable coefficients in Fig. 3 provide important information from both a sta-
tistical and a business perspective. The blue line represents the constant coefficient,
which indicates that the customer’s initial purchase intention is negatively related to
the product price. In other words, an increase in the price of the product reduces the
likelihood that customers will make a purchase. This finding is a key to the pricing
strategy and implies that sellers should be careful with their pricing policies, because
too high a price will discourage potential buyers, which may reduce sales.
The orange line indicates the review coefficient. Initially, this line shows an upward
trend, indicating that an increase in the number of reviews positively influences cus-
tomers’ purchasing decisions, especially for low-priced items. However, this effect
decreases as the price of the product increases and eventually levels off. Therefore, for
low-priced items, sellers should stimulate and promote customer reviews to increase
sales. For higher priced items, a large number of reviews may be less persuasive,
suggesting the need for a different marketing strategy.
In contrast, the red line indicates the “like” factor. The impact of this factor may
be negligible initially, but increases as the price of the product rises. The number
of comments and likes are both important indicators of a product’s popularity. The
unique impact of these variables can be attributed to the unique dynamics of Mercari,
where reviews often go beyond simple product reviews to include aspects such as
price negotiation discussions. This observation suggests that sellers of high-priced
items should focus on accumulating “likes,” as this appears to enhance the appeal and

123
 New Generation Computing

Fig. 3 Sentiment scores’ coefficient in different price range

potential sales of their products. In addition, sellers should understand and adapt to
the different characteristics of their respective platforms.
The green line corresponds to the sentiment score factor. The impact of the sentiment
score increases with the price of the product. Customers may pay more attention
to the sentiments conveyed in product descriptions when considering higher priced
items. However, it should be noted that this trend reverses for all variables above the
100,000 yen threshold. For sellers, this emphasizes the need to create emotionally rich
product descriptions, especially for higher priced items. For ultra-high-priced items
(over 100,000 yen), a different strategy is required as the influence of sentiment scores
and “likes” diminishes.
Figure 4 presents the p values associated with each coefficient. For most price
ranges, the p values hover around 0, indicating that the coefficients are statistically
significant. Thus, the correlation between customer behavior and the evaluation vari-
ables (product price, number of reviews, number of “likes”, and sentiment scores)
is not coincidental. From a business perspective, this highlights the importance of
these variables in influencing customer buying behavior, suggesting that they should
be central to the development of business strategies.
Beyond the 100,000 yen range, the p value associated with sentiment scores deviates
from the nonsignificant region, suggesting that there are limitations to sentiment scores
calculated using word frequency, especially for high-priced products. This insight is
crucial for sellers of high-priced products, suggesting that relying solely on sentiment
scores calculated using traditional methods may not provide accurate insight into
customer sentiment, which may lead to ineffective business strategies.

123
New Generation Computing

Fig. 4 Sentiment scores’ p value in different price range

The results of the study reveal key insights related to two questions posed in Chap-
ter 2.5. (1) The study found that product descriptions significantly influence buyers’
purchase intentions. Specifically, the sentiment score in product descriptions (shown
as the green line in Fig. 3) increases with product price, suggesting that customers
may place more importance on the sentiment expressed in product descriptions when
considering higher priced products. (2) We note that these descriptions moderated
purchase decisions in a way that was related to price. For example, sentiment scores
derived from product descriptions had a significant effect on purchase decisions for
medium-priced products, but the effect appeared to diminish for ultra-high-priced
products (100,000 yen or more). This emphasizes that the impact of product descrip-
tions on purchase intention varies by product price point and suggests that sellers
should adjust product descriptions according to product price point.
This finding highlights the need to investigate alternative methods for calculating
sentiment scores for high-priced products. This avenue for future research will not
only address existing limitations, but also provide organizations with a more accurate
tool for measuring customer sentiment, thus enabling them to adjust their market-
ing strategies more effectively. From a business perspective, this will require sellers
to continually refine their analytic processes and keep up with the development of
sentiment analysis technology, especially for high-end products.

123
 New Generation Computing

5 Conclusion

The main objective of this study is to experimentally investigate whether and how
the textual content of product descriptions, especially the sentiment element, influ-
ences customers’ purchase intentions. In an era where sustainability is becoming
increasingly important, understanding trust relationships in business, especially on
customer-to-customer (C2C) platforms, has become a vital, yet often overlooked
aspect of sustainable business practices.
Based on year-round digital transaction data from Mercari, Japan’s leading C2C
second-hand trading platform, we found that product descriptions written by sellers
do influence buyers’ purchasing decisions to some extent. To realize this analysis, we
informatively used a set of well-established Japanese sentiment dictionaries to extract
sentiment scores from textual descriptions. This method assigns scores ranging from
1 (indicating negative sentiments) to 5 (indicating positive sentiments) to individual
words in the descriptions.
A noteworthy trend emerged during the analysis: As the price of the product
increased, the description had an increasingly positive effect on the purchase deci-
sion. This observation suggests a compelling relationship between positive sentiment
elements in product descriptions and better sales performance. We measure this sales
performance using a logistic regression model that incorporates various factors such
as the number of product reviews and likes. However, the relationship becomes statis-
tically insignificant when the price of the product exceeds the 100,000 yen threshold.
Our study demonstrates that on C2C platforms such as Mercari, customers tend to
take the sentiment tone of product descriptions more seriously as the price of the prod-
uct increases. For lower priced items, sentiment content seems to have less impact. This
observation suggests that textual content does influence customers’ purchase propen-
sity to some extent, with far-reaching potential implications for C2C e-Commerce
practices.
Our study has practical implications for e-Commerce managers. First, our results
emphasize the impact of product descriptions on consumers’ purchasing decisions,
especially the fact that sentiment factors can serve as high-quality “signals” that can
help buyers make informed purchasing decisions and reduce information asymmetry
between buyers and sellers. To increase sales success, sellers should pay attention
to the sentiment color of the language when creating product descriptions and use
positive words as much as possible to enhance buyers’ willingness to purchase. Second,
our study also provides information on the design and management of e-Commerce
platforms. Platforms can help sellers create more attractive product descriptions by
providing better tools and guidance, such as setting up some writing guides or tools
that automatically optimize product descriptions.
However, our study has limitations. We focus mainly on how hidden sentiment
scores in product descriptions affect purchase intention, thus avoiding the inclusion
of too many variables in the logistic regression model. Future studies can further
enrich this approach by incorporating more variables (e.g., time from listing to selling
of the item) to gain a more comprehensive understanding of the purchase decision
mechanism. Furthermore, it is necessary to address anomalous data, such as products
with unreasonably inflated prices, in subsequent studies.

123
New Generation Computing

In addition, the results of the current study are limited to specific product categories
on the Mercari platform. To gain a more comprehensive understanding of the observed
phenomena, future research should explore differences between product categories.
The limited explanatory power of our analysis for high-priced products means that
simple word frequency calculations may not be sufficient. This limitation suggests
that more sophisticated methods are needed to explore the sentiment scores contained
in the texts.
In conclusion, this study provides valuable information on the role of sentiment in
product descriptions and its impact on consumer purchase decisions. Although there
is room for improvement and further research, these findings contribute to an initial
understanding of how consumers respond to sentiment signals in product descrip-
tions. More importantly, they emphasize the potential importance of creating more
enjoyable and trustworthy shopping environments, especially on C2C platforms, to
improve sustainable business practices by strengthening the critical role of trust in
such environments.
Acknowledgements This research is part of the results of Value Exchange Engineering, a joint research
project between R4D, Mercari, Inc. and the RIISE, University of Tokyo. This study is also partly supported
by JSPS KAKENHI 20K20482.

Funding Open Access funding provided by The University of Tokyo.

Declarations

Conflict of Interest On behalf of all authors, the corresponding author declares that there is no conflict of
interest or personal relationships that could have appeared to influence the work reported in this article.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence,
and indicate if changes were made. The images or other third party material in this article are included
in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If
material is not included in the article’s Creative Commons licence and your intended use is not permitted
by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the
copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

References
1. Oláh, J., Kitukutha, N., Haddad, H., Pakurár, M., Máté, D., Popp, J.: Achieving sustainable e-commerce
in environmental, social and economic dimensions by taking possible trade-offs. Sustainability 11(1),
89 (2018)
2. Sun, K., Liu, Y., Chen, X., Szolnoki, A.: Evolution of trust in a hierarchical population with punishing
investors. Chaos Solit. Fract. 162, 112413 (2022). https://doi.org/10.1016/j.chaos.2022.112413
3. Chen, J., Zhang, C., Xu, Y.: The role of mutual trust in building members’ loyalty to a c2c plat-
form provider. Int. J. Electron. Commer. 14(1), 147–171 (2009). https://doi.org/10.2753/JEC1086-
4415140105
4. Chen, D., Lai, F., Lin, Z.: A trust model for online peer-to-peer lending: a lender’s perspective. Inform.
Technol. Manag. 15, 239–254 (2014). https://doi.org/10.1007/s10799-014-0187-z
5. Sun, Y., Ohsawa, Y.: The impact of product descriptions on customers’ trust in sellers and purchase
intention, pp. 1930–1935 (2022). https://doi.org/10.1109/BigData55660.2022.10020527

123
 New Generation Computing

6. Kim, R.Y.: When does online review matter to consumers? The effect of product quality information
cues. Electron. Commer. Res. 21(4), 1011–1030 (2021). https://doi.org/10.1007/s10660-020-09398-
0
7. Fang, X., Zhan, J.: Sentiment analysis using product review data. J. Big Data 2(1), 1–14 (2015). https://
doi.org/10.1186/s40537-015-0015-2
8. Situmeang, F., Boer, N., Zhang, A.: Looking beyond the stars: a description of text mining technique
to extract latent dimensions from online product reviews. Int. J. Market Res. 62(2), 195–215 (2020).
https://doi.org/10.1177/1470785319863619
9. Mavlanova, T., Benbunan-Fich, R., Koufaris, M.: Signaling theory and information asymmetry in
online commerce. Inform. Manag. 49, 240–247 (2012). https://doi.org/10.1016/j.im.2012.05.004
10. Spence, A.: Signaling in retrospect and the informational structure of markets. Am. Econ. Rev. 92,
434–459 (2002). https://doi.org/10.1257/00028280260136200
11. Liberti, J.M., Petersen, M.A.: Information: hard and soft. Rev. Corporate Finance Stud. 8(1), 1–41
(2018). https://doi.org/10.1093/rcfs/cfy009
12. Connelly, B.L., Certo, S.T., Ireland, R.D., Reutzel, C.R.: Signaling theory: a review and assessment.
J. Manag. 37(1), 39–67 (2011). https://doi.org/10.1177/0149206310388419
13. Lee, K., Lee, B., Oh, W.: Thumbs up, sales up? The contingent effect of facebook likes on sales
performance in social commerce. J. Manag. Inform. Syst. 32, 109–143 (2015). https://doi.org/10.
1080/07421222.2015.1138372
14. Liu, Y.: Word of mouth for movies: its dynamics and impact on box office revenue. J. Market. 70,
74–89 (2006). https://doi.org/10.1509/jmkg.70.3.074
15. Chintagunta, P., Gopinath, S., Venkataraman, S.: The effects of online user reviews on movie box office
performance: accounting for sequential rollout and aggregation across local markets. Market. Sci. 29,
944–957 (2010). https://doi.org/10.2139/ssrn.1331124
16. Hu, N., Koh, N., Reddy, S.: Ratings lead you to the product, reviews help you clinch it? The mediating
role of online review sentiments on product sales. Decis. Support Syst. 57, 42–53 (2014). https://doi.
org/10.1016/j.dss.2013.07.009
17. Zhou, M., Lu, B., Fan, W., Wang, G.A.: Project description and crowdfunding success: an exploratory
study. Inform. Syst. Front. 20, 259–274 (2018). https://doi.org/10.1007/s10796-016-9723-1
18. Jiang, C., Han, R., Xu, Q., Liu, Y.: The impact of soft information extracted from descriptive text on
crowdfunding performance. Electron. Commer, Res. Appl. 43, 101002 (2020) https://doi.org/10.1016/
j.elerap.2020.101002
19. Chang, H., Lu, Y.-Y., Lin, S.: An elaboration likelihood model of consumer respond action to facebook
second-hand marketplace: impulsiveness as a moderator. Inform. Manag. 57, 103171 (2019). https://
doi.org/10.1016/j.im.2019.103171
20. Gregg, D., Walczak, S.: The relationship between website quality, trust and price premiums at online
auctions. Electron. Commerce Res. 10, 1–25 (2010). https://doi.org/10.1007/s10660-010-9044-2
21. Kardes, F.R., Posavac, S.S., Cronley, M.L.: Consumer inference: a review of processes, bases,
and judgment contexts. J. Consum. Psychol. 14(3), 230–256 (2004). https://doi.org/10.1207/
s15327663jcp1403_6
22. Rindova, V.P., Williamson, I.O., Petkova, A.P., Sever, J.M.: Being good or being known: an empirical
examination of the dimensions, antecedents, and consequences of organizational reputation. Acad.
Manag. J. 48(6), 1033–1049 (2005). https://doi.org/10.5465/amj.2005.19573108
23. Ter Huurne, M., Ronteltap, A., Corten, R., Buskens, V.: Antecedents of trust in the sharing economy:
a systematic review. J. Consumer Behav. 16(6), 485–498 (2017). https://doi.org/10.1002/cb.1667
24. Bente, G., Baptist, O., Leuschner, H.: To buy or not to buy: influence of seller photos and reputation
on buyer trust and purchase behavior. Int. J. Hum. Comput. Stud. 70(1), 1–13 (2012). https://doi.org/
10.1016/j.ijhcs.2011.08.005
25. Li, X., Guo, X., Wang, C., Zhang, S.: Do buyers express their true assessment? Antecedents and
consequences of customer praise feedback behaviour on taobao. Internet Res. 26, 1112–1133 (2016).
https://doi.org/10.1108/IntR-03-2015-0063
26. Thierer, A., Koopman, C., Hobson, A., Kuiper, C.: How the internet, the sharing economy, and rep-
utational feedback mechanisms solve the lemons problem. U. Miami L. Rev. 70, 830 (2015). https://
doi.org/10.2139/ssrn.2610255
27. Chen, X., Huang, Q., Davison, R., Hua, Z.: What drives trust transfer? The moderating roles of seller-
specific and general institutional mechanisms. Int. J. Electron. Commer. 20, 261–289 (2015). https://
doi.org/10.1080/10864415.2016.1087828

123
New Generation Computing

28. Blei, D., Ng, A., Jordan, M.: Latent dirichlet allocation. J. Mach. Learn. Res. 3, 993–1022 (2003).
https://doi.org/10.1162/jmlr.2003.3.4-5.993
29. Blei, D.M.: Probabilistic topic models. Commun. ACM 55(4), 77–84 (2012). https://doi.org/10.1145/
2107736.2107741
30. Tang, J., Meng, Z., Nguyen, X.L., Mel, Q., Zhang, M.: Understanding the limiting factors of topic
modeling via posterior contraction analysis, vol. 1, pp. 337–345 (2014). https://api.semanticscholar.
org/CorpusID:6527691
31. Reisenbichler, M., Reutterer, T.: Topic modeling in marketing: recent advances and research opportu-
nities. J. Bus. Econ. 89, 1–30 (2019). https://doi.org/10.1007/s11573-018-0915-7
32. Yamada, A., Hashimoto, S., Nagata, N.: A Text mining approach for automatic modeling of kansei
evaluation from review texts, pp. 319–328 (2018). https://doi.org/10.1007/978-981-10-8612-0_34
33. Koltcov, S.: Application of rényi and tsallis entropies to topic modeling optimization. Physica A: Stat.
Mech. Appl. 512, 1192–1204 (2018). https://doi.org/10.1016/j.physa.2018.08.050
34. Koltcov, S., Ignatenko, V.: Renormalization analysis of topic models. Entropy 22(5), 556 (2020).
https://doi.org/10.3390/e22050556
35. Nakayama, M., Hatanakaka, C., Konakawa, H., Suzuki, Y., Koh, A., Sugihara, Y., Kawai, T.: Japanese
dictionary for sentiment analysis of counselling text. In: Proceedings of the 9th International Conference
on Human-Agent Interaction, pp. 311–315 (2021). https://doi.org/10.1145/3472307.3484663
36. Fan, Z.-P., Che, Y.-J., Chen, Z.-Y.: Product sales forecasting using online reviews and historical sales
data: a method combining the bass model and sentiment analysis. J. Bus. Res. 74, 90–100 (2017).
https://doi.org/10.1016/j.jbusres.2017.01.010

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

123