Research on Demand Mining Method for Short Life Cycle
Experiential Products Based on Structural Topic Model and
Experience Value
Zhongjun Tang and Xinhao Zhou
School of Economics and Management, Beijing University of Technology, Beijing 100124, China
Keywords: Short Life Cycle Experience Products, Demand Mining, Structural Topic Model, Experience Value.
Abstract: With the development of social economy, short life cycle experiential products occupy an increasingly
important position in the market. The demand for short life cycle experiential products through big data
methods is of great significance to product improvement and innovation. This paper proposes a short-life
cycle experience product demand mining method based on structural topic model and experience value. In
order to meet the short-term characteristics of short life cycle experiential products, collect user comment data
on the website, and adopt the structural topic model (STM) method, the user comment rating is used as the
covariate in STM model, extract the customer demand topic words and their corresponding emotional
tendencies and visualizing. The demand topics excavated using the STM method are divided into five
categories based on the experience value theory, so that the excavated short life cycle experiential product
demands are experiential. This paper takes a movie as an example to verify the effectiveness of the proposed
method. The method is effective and more accurate than traditional methods, which provides guidance for
enterprises to tap customer needs and product innovation.
1 INTRODUCTION
With the continuous change of consumer demand for
products and the rapid development of product
production technology, the update speed of the
products is accelerated, the innovation cycle is
shortened, and the short life cycle experience products
occupy an increasingly important position in the
market. Short life cycle experiential products have the
characteristics of short-term and experiential. The
short-term is mainly reflected in the rapid renewal of
products, the rapid decline of value, and the strong
substitution effect of competitive products. The
experiential is mainly reflected in that consumers can
only evaluate the product quality after consumption,
and the subjective participation is strong. Movies and
mobile games are typical short life cycle experience
products (TANG, CUI, TANG, ZHU 2019). It is of
great significance to study the method of demand
mining for short life cycle experiential products, and
to discover and analyze customer demand for real
enterprises, so as to improve and innovate products.
Traditional market research methods such as
structured questionnaire or semi-structured
questionnaire are mainly used in existing research on
customer demand mining. This kind of research
method requires high rationality in questionnaire
design and takes a long time to design questionnaire,
which is not suitable for short life cycle experience
products. With the development of the Internet and
online comment platform, customers will choose the
network platform to make subjective comments after
consumption. Online comments are directly from
customers, which have the characteristics of
spontaneity, authenticity, and have become a research
hotspot of researchers. Using text mining method to
mine customer demands from online reviews can help
manufacturers innovate products quickly. Combined
with customer experience value, this paper constructs
demand classification according to the experiential
characteristics of short life cycle experiential products.
In the same movie, when the same keyword appears
as a high-frequency word in both positive and negative
comments, it is difficult for researchers to judge its
true emotional tendency. Therefore, on the basis of the
existing online review demand mining methods, we
also demand to solve the problem that the frequency
of keywords in positive comments is equal to that in
negative comments, which makes it impossible to
judge the real tendency of a keyword in online
Tang, Z. and Zhou, X.
Research on Demand Mining Method for Short Life Cycle Experiential Products Based on Structural Topic Model and Experience Value.
DOI: 10.5220/0011754200003607
In Proceedings of the 1st International Conference on Public Management, Digital Economy and Internet Technology (ICPDI 2022), pages 665-673
ISBN: 978-989-758-620-0
Copyright
c
2023 by SCITEPRESS Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
665
comments.
In summary, this article proposes a short-life cycle
experience product demand mining method. This
method uses online reviews as the data source,
replaces traditional text mining methods with
structural topic models to improve the accuracy of
demand mining, and compares the results with
traditional text mining methods to illustrate the
importance of using of structural topic models. Then
combine the experiential value theory with the
unearthed demand topics. Take a movie as a case to
verify the effectiveness of the method proposed in this
article.
2 RELATED WORKS
2.1 Topic Model and Online Comment
Demand Mining Method
The existing online comment data demand mining
methods mainly focus on text mining and sentiment
analysis, and mainly use content analysis and machine
learning methods to conduct topic mining on online
comments (Sangwan, Malik, Sunesh, et al. 2017),
sentiment polarity judgment (Usama, Ahmad, Song, et
al. 2020), and sentiment intensity calculations. The
common method of topic mining is the Latent
Dirichlet Allocation (LDA)model. Researchers can
mine the focus of users through LDA, but the
emotions implied in the comment cannot be judged.
The sentiment polarity judgment and sentiment
intensity calculation are usually based on the
sentiment dictionary. But the method has high
requirements for the construction of sentiment
dictionary.
On the basis of the above, the researchers put
forward a method to combine the LDA model with
sentiment analysis to simultaneously obtain the topic
and the emotion corresponding to the topic from the
comments. Lin et al. (Lin, He 2009) added an
additional sentiment layer to the LDA framework and
established a joint sentiment/topic model (JST), which
links sentiment polarity, documents, topics, and
words. Jo et al. (Jo, Oh 2011) proposed the Aspect and
Sentiment Unification Model (ASUM) to extract
feature-sentiment pairs from documents. Dong et al.
(Dong, Ji, Zhang, et al. 2018) proposed an
unsupervised topic-sentiment joint probabilistic
model (UTSJ) on the basis of JST. After improving
the topic model, Rao et al. (Rao, Li, Mao, et al. 2014)
and Feng et al. (Feng, Rao, Xie, et al. 2019) proposed
the Sentiment Latent Topic Model (SLTM) and the
Multi-label Supervised Topic Model (MSTM), the
accuracy of emotion calculation in the topic-emotion
model is improved by adding different types of
supervision tags to different generation stages of the
model.
The above research mainly focuses on the
improvement of the accuracy and stability of the
emotion polarity extraction and calculation in the
topic-emotional model, and does not consider the
frequency difference of keywords in negative
comments and positive comments.
2.2 Structural Topic Model (STM)
Roberts et al. (Roberts, Stewart, Tingley, et al. 2014)
proposed structural topic model (STM) in 2014, which
effectively solved the problem of keyword frequency
of positive and negative reviews. Structural topic
model allows researchers to input some features of
reviews into the model as covariate parameters in
advance to explore the relationship between topics and
review features. Taking the positive and negative
tendency of reviews as covariates can overcome the
frequency problem of the above keywords, mine
customers' real satisfaction and dissatisfaction
demands, and improve the accuracy of demand
mining method. As shown in Figure 1, the topic
popularity covariate X affects the document-topic
probability generation, and the topic content covariate
Y affects the topic-word probability generation.
STM has been used in the text analysis of news
reports (Chandelier, Steuckardt, Mathevet, et al.,
Figure 1: Schematic diagram of structural topic model.
ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology
666
2018), political discourse (Curry, Fix 2019) and other
fields, using time as a covariate to analyze the trend of
thematic changes in news reports. Scholars such as
Tvinnereim (Tvinnereim, Fløttum 2015) used STM to
extract topics from semi-structured texts such as
question-and-answer texts (Lester, Kessler, Modisett,
et al. 2019) and aviation reports (Kuhn 2018), in order
to study the user characteristics obtained from semi-
structured texts and the relevance of texts. In view of
the above-mentioned differences between STM and
LDA, the use of STM to mine online user reviews of
experiential products has the following two
advantages: 1. STM enables researchers to analyze
document-level covariates (such as whether the
review is positive or negative, the time of the review,
The gender of the reviewer, etc.) introduce topic
popularity covariates to explore the impact of the
covariate on the document-topic probability, and at the
same time, explore the changes in the document-topic
probability with the change of the covariate, and
consider the heterogeneity of positive and negative
reviews. 2. STM enables researchers to introduce
document-level covariates into topic content
variables, thereby affecting topic-word probability,
and studying the distribution of topic words among
covariates.
3 MATERIAL AND METHODS
The research framework proposed in this paper is as
follows: First, use a web crawler program based on
Python language to crawl user online comments and
their corresponding product ratings, and complete data
preprocessing; Then, use STM to extract the subject
words in the user's online comments, filter the topics
related to the demands, and visualize the demands;
Finally, the product requirements obtained by
combining subject terms with experience value and
the correlation between topics are analyzed. In order
to illustrate the changes and advantages brought by the
introduction of covariates in STM, the Term
Frequency-inverse Document Frequency (TF-IDF)
algorithm was used to extract the keywords in the
user's positive and negative comments, and the subject
words obtained by STM. Make a comparison and
analyze the impact of the frequency difference
between keywords in positive reviews and negative
reviews. The research framework is shown in Fig. 2.
The first step is data collection and preprocessing.
A web crawler program written in Python language
obtains user reviews and their ratings from experience
product review sites, and performs operations such as
filtering, cleaning, and de-duplication on the data. Use
the ‘prepDocuments’ function to normalize and index
the data, and delete comments with less than 50 words.
Use the ‘TextProcessor’ function of the STM package
in the R language to process the data, including the
extraction of stems, the removal of default stop words
and custom stop words (the, is ,at). And the deletion
of numbers and punctuation to ensure that the
covariate corresponding to each comment does not
contain missing values for subsequent use in the STM
model.
Figure 2: Research framework of demand mining method for short life cycle experience products.
Research on Demand Mining Method for Short Life Cycle Experiential Products Based on Structural Topic Model and Experience Value
667
The second step is demand mining and demand
analysis. Use the STM model to perform text mining
on the pre-processed review data, use the ratings
selected by the reviewer to divide the reviews into
positive reviews and negative reviews, and introduce
the positive and negative of the reviews as a topic
popular covariate into the STM and perform topic
extraction, According to the words in the topic to
manually classify and determine the topic label. Use
word cloud graphs to visualize words that appear in
key topics, and extract and analyze relevant comments
on certain key topics. Using the TF-IDF algorithm to
extract keywords for positive reviews and negative
reviews, compare and analyze the extracted keywords
with the subject terms extracted using the STM
method, explaining the difference between STM and
the conventional word frequency method for demand
mining. It reflects the necessity of using STM for
demand mining.
The third step is demand classification analysis.
Therefore, based on the use of STM for subject
extraction, experience value is added to classify user
demands, to obtain user demand classification for
short life experience products, and analyze each
experience value classification.
4 RESULTS & DISCUSSION
This article takes the typical product movie of short
life cycle experience product as an example, and
conducts demand mining and demand analysis from
movie user reviews.
4.1 Data Collection and Preprocessing
IMDB (Internet Movie Database, IMDB.com), as the
most detailed movie database in the world, provides a
platform for global movie critics to express their
personal opinions on movies that have been released.
Users demand to rate movies while commenting. The
number represents the overall degree of user
satisfaction with the movie, and reflects the user's
subjective judgment on the movie. This article uses
the score corresponding to the review as the standard
for dividing positive reviews and negative reviews.
Using the Python language web crawler to crawl
the comments from the IMDB website, the movie
ratings corresponding to the comments, the reviewer
ID, and the review time. This article selects the user
reviews of Captain Marvel released in 2019 as the data
source because the number of the reviews is
reasonable, and the difference in the number of
positive and negative comments is small. After
deleting invalid comments, 6144 comments were
finally collected, and the number of comments
corresponding to each comment star rating is shown in
Figure 3. Since the number of positive reviews with
more than 5 stars is much higher than the number of
negative reviews with less than 5 stars, in order to
balance the gap in the number of positive and negative
reviews, 1-6 star reviews are regarded as negative
reviews and 7-10 star reviews are regarded as positive
reviews. The positive and negative of is introduced as
a covariate into the STM model.
Figure 3: Star distribution of movie reviews.
4.2 Topic Extraction
Input the crawled movie reviews and their
corresponding pre-processed comment positive and
negative as topic popular covariates into STM, change
the number of topics several times for model training,
determine the optimal number of topics to be 20, and
use the STM model for topic extraction. Table 1 shows
the results of the extracted topics.
“Prob” represents the word with the highest
occurrence probability in the topic, but the word with
the highest occurrence probability in a certain topic
may also appear with a high probability in another
topic, so the degree of discrimination is insufficient.
STM introduces the FREX (Frequency-exclusivity)
statistic, which is defined as the ratio of topic-based
word frequency to word-topic exclusivity, which can
avoid quoting only the most commonly used words.
STM also introduces the metric “Lift”. This metric
refers to the probability of words appearing on the
topic divided by the probability of words appearing in
the entire corpus. This metric will highlight the more
common words in the topic than in the corpus, which
the frequency of occurrence in this topic is much
higher than in the entire corpus.
According to the words extracted from the above
metrics, a total of 20 topic tags are designed. Among
them, 20 topics related to movie features or the
process of watching movies are selected: topic1
ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology
668
(Actor acting), Topic2 (Sexism), Topic3
(Comparative movies), topic4 (Leading actors), topic5
(Other actors), topic6 (IP animation) Topic7 (User
emotion), topic8 (Feminism), topic9 (Related
movies), topic10 (Movie plot), topic11 (Watching
experience), topic12 (Main IP movie), topic13
(Personal thoughts), topic14 (Marvel Universe),
topic15 (Personal expectations), topic16 (starring
posture), topic17 (User comments), topic18 (Plot
perception), topic19 (Plot trend) and topic20 (Deep
thinking). See Table 1.
4.3 Positive and Negative Topics
Discovery
Using the STM model to find the positive and negative
tendencies of all demand topics. The results are shown
in Figure 4. The demand topic with positive tendency
represents the positive feedback of commenting users
on the movie, while the demand topic with negative
tendency represents the negative feedback of
commenting users on the movie. The topics with
strong tendency can be identified and extracted
through STM, so as to more accurately identify the
demands of consumers.
Table 1: Movie demand topic extraction results.
Topic No. Topic Label Criteria Word1 Word2 Word3 Word4 Word5
1
Actor Acting
Prob charact feel marvel mcu like
FREX phase stake risk care style
2
Sexism
Prob femin white charact feminist men
FRE
X
garbage white femin insult sexism
3
Comparative
Movies
Prob like thing get make time
FREX obstacle shake scheme sound mood
4
Leading Actors
Prob larson brie charact jackon samuel
FRE
X
samuel jackson
p
erfor
m
larson acto
r
5
Other Actors
Prob man thor black better war
FREX iron thor panther man america
6
IP Animation
Prob furi great carol love nick
FRE
X
carol goos nic
k
fury danve
r
7
User Emotion
Prob good realli scene action great
FREX good funny pretty scene cool
8
Feminism
Prob just female woman power strong
FRE
X
strong supe
r
women girl female
9
Related Movies
Prob save money alita battle mcu
FREX save alita angel battle grace
10
Movie Plot
Prob marvel kree earth skrull captain
FRE
X
ve
r
kree earth ryan
oden
11
Watching
Experience
Prob watch see end will wait
FREX watch worth theater cinema must
12
Main IP Movie
Prob marvel captain avenge endgame universe
FRE
X
universe aven
g
e ca
p
tain thano marvel
13
Personal
Thoughts
Prob review see people enjoy hate
FREX review negat rate hate troll
14
Marvel
Universe
Prob marvel captain charact space point
FRE
X
space sovel nostalg limit relev
15
Personal
Expectations
Prob high recommend incred special effect
FREX recommend high incred balanc blow
16
Starring Posture
Prob marvel worst actress ever poor
FRE
X
worst eve
r
gadot
p
oo
r
shame
17
User Comments
Prob bore just bad charact time
FREX bore horrible asleep bland uninterest
18
Plot Perception
Prob polit emake comic express like
FRE
X
p
olite express view facial shove
19
Plot Trend
Prob stori origin mcu marvel great
FREX origin pack stori nice classic
20
Deep Thinking
Prob just felt realli way think
FRE
X
felt rush familia
r
q
uestion aroun
d
Research on Demand Mining Method for Short Life Cycle Experiential Products Based on Structural Topic Model and Experience Value
669
Figure 4: Movie demands thematic sentiment tendencies.
Figure 5: Cloud chart of demand subject words with positive.
Figure 6: Cloud chart of demand subject words with negative.
A total of 4 topics with strong positive and negative
tendencies are visualized by using word cloud
diagram. The results are shown in Figure 5 and Figure
6. The results of the cloud chart of the subject words
with positive tendency show that the comments in the
personal thought topic mainly focus on the personal
views of the comment users, mainly emotional words;
The topic words of plot trend focus on the originality
of the plot and the general story line of the movie
company where the movie is located. Combined with
the positive tendency of the topic, the comment users
have a positive evaluation of the plot trend of the
movie.
The cloud chart results of subject words with
negative tendency show that the subject words
evaluated by users are mainly the evaluation words
used by users for the movie, mainly adjectives. The
high-frequency words in the starring posture topic
ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology
670
Table 2: User comment keywords in positive and negative movie reviews.
Keywords
(Negative)
Frequency of
occurrence
Keywords
(Positive)
Frequency of
occurrence
marvel 0.0116 marvel 0.0122
captain 0.0053 captain 0.0062
like 0.0053 story 0.0042
character 0.0052 brie 0.0038
brie 0.0039 mcu 0.0033
larson 0.0036 watch 0.0024
story 0.0032 reviews 0.0019
mcu 0.0030 fury 0.0019
boring 0.0023 fun 0.0019
acting 0.0021 carol 0.0018
jackson 0.0015 avengers 0.0018
endgame 0.0015 origin 0.0015
samuel 0.0011 acting 0.0012
mainly describe the appearance of the heroine of the
movie.
4.4 Comparative Analysis of TF-IDF
Method and STM
TF-IDF algorithm is used to extract the keywords of
negative comments and positive comments
respectively, and the results are compared with the
topics extracted by STM and their comment tendency.
Since the TF-IDF value is the product of the two
statistics of word frequency and inverse document
frequency, it is difficult to understand its intuitive
meaning. Select the effective words related to the topic
from the words with high TF-IDF value to calculate
their occurrence frequency in the document. The
results are shown in Table 2.
Through comparative analysis, it is found that
most of the negative or positive keywords extracted by
TF-IDF method are consistent with the tendencies of
their topics, but there are also exceptions. The
keyword "endgame" extracted from the negative
comments appears in the topic12 "main IP movie"
with a positive tendency, and the keyword "story"
appears in the topic19 "plot trend" with a positive
tendency, which indicates that the reviewers'
comments on the main IP movie and movie plot
corresponding to the movie are mostly in the range of
negative comments. However, "main IP movie" and
"plot trend" are not the aspects that the commentators
are dissatisfied with the movie. On the contrary, the
commentators' emotional tendency towards these two
topics is positive.
Accordingly, the keyword "Brie" with high TF-
IDF value extracted from the positive comments, as a
part of the name of the leading actor Brie Larson,
appears in the topic topic4 "leading actor" which tends
to be negative, indicating that the commentators also
refer to the leading actor more in the positive
comments with high scores, but the commentators
tend to comment more negatively on the actor. This is
also verified by the fact that the leading actor's
surname "Larson" appears in negative comments with
a high frequency, while the frequency in positive
comments decreases significantly.
This shows that the frequency difference between
negative comments and positive comments of key
high-frequency words obtained by TF-IDF or word
frequency calculation method will affect the results of
demand mining. In the case of positive and negative
comments with the same keyword and unable to judge
the word tendency, or the original words belonging to
a certain tendency are attributed to the comment
keywords with the opposite tendency, lack of certain
accuracy.
4.5 Demand Classification Analysis
Combined with Experience Value
The topics are classified using five experience value
classification methods in the strategic experience
module (Bassi 2010). Classify the topic labels
according to the experience value classification, and
extract the positive and negative tendency of the topic
corresponding to each demand topic. Table 3 shows
the movie demand used in the demonstration. The
movie demand is mainly composed of the experience
value type, the topic label corresponding to the
experience value type and the positive and negative
tendency of the topic label.
Research on Demand Mining Method for Short Life Cycle Experiential Products Based on Structural Topic Model and Experience Value
671
Table 3: Combining movie demands with experience value types.
Type of
Experience value
Value Description
Topic serial
numbe
r
Topic label
Topic
tendenc
y
Sense
Affect consumers' external
perception
1 Actor Acting Negative
4 Leading Actors Negative
5 Other Actors Positive
11 Watching Experience Positive
16 Starring Posture Negative
Feel Affect consumer sentiment
7 User Emotion Positive
10 Movie Plot Positive
17 User Comments Negative
18 Plot Perception Negative
19 Plot Trend Positive
Think
Elicit consumers' ideas on
specific topics
3 Comparative Movies Negative
6 IP Animation Positive
9 Related Movies Negative
12 Main IP Movie Positive
Act
Stimulate consumers to
change their lifestyles and
expectations
13 Personal Thoughts Positive
15 Personal Expectations Positive
20 Deep Thinking Negative
Realte
Arouse consumers to think
about the connection with
society or culture
2 Sexism Negative
8 Feminism Negative
14 Marvel Universe Negative
5 CONCLUSIONS
This paper proposes and constructs a short life cycle
experience product consumer demand mining method
based on structural topic model. Using structural
topic model method, the demand topics of positive
and negative comments are extracted from online
comments. Combined with consumer experience
value, a structured experience product demand
classification is constructed. It provides new ideas for
improving the product quality and innovation
direction of experience products. Taking a movie as
an example, the effectiveness of this method is
verified. Compared with the traditional keyword
topic extraction method. The results show that the
demand mining method proposed in this paper
considers the impact of the inconsistent number of
positive and negative comments in online comments,
and effectively improves the accuracy of extracting
demand topics from online comments, helps to
provide guidance for enterprises and manufacturers
developing short life cycle experience products to
quickly obtain user demands, and promote the rapid
innovation and generation of short life cycle
experience products.
There are deficiencies in the selection of model
covariates. Only the scores of online comments are
considered. More online comment features or online
comment user features (such as comment time,
comment humanity) can be used as covariates for
demand mining.
ACKNOWLEDGMENT
This study is supported by the National Natural
Science Foundation of China (Grant No. 71672004).
REFERENCES
Bassi F. Experiential goods and customer satisfaction: An
application to films[J]. Quality Technology &
Quantitative Management, 2010, 7(1): 51-67.
Chandelier M, Steuckardt A, Mathevet R, et al. Content
analysis of newspaper coverage of wolf recolonization
in France using structural topic modeling[J]. Biological
conservation, 2018, 220: 254-261.
Curry T A, Fix M P. May it please the twitterverse: The use
of Twitter by state high court judges[J]. Journal of
Information Technology & Politics, 2019, 16(4): 379-
393.
Dong L, Ji S, Zhang C, et al. An unsupervised topic-
sentiment joint probabilistic model for detecting
ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology
672
deceptive reviews[J]. Expert Systems with
Applications, 2018, 114: 210-223.
Feng J, Rao Y, Xie H, et al. User group based emotion
detection and topic discovery over short text[J]. World
Wide Web, 2019: 1-35.
Jo Y, Oh A H. Aspect and sentiment unification model for
online review analysis[C]//Proceedings of the fourth
ACM international conference on Web search and data
mining. 2011: 815-824.
Kuhn K D. Using structural topic modeling to identify
latent topics and trends in aviation incident reports[J].
Transportation Research Part C: Emerging
Technologies, 2018, 87: 105-122.
Lin C, He Y. Joint sentiment/topic model for sentiment
analysis[C]//Proceedings of the 18th ACM conference
on Information and knowledge management. 2009:
375-384.
Lester C A, Kessler J M, Modisett T, et al. A text mining
analysis of medication quality related event reports
from community pharmacies[J]. Research in Social and
Administrative Pharmacy, 2019, 15(7): 845-851.
Rao Y, Li Q, Mao X, et al. Sentiment topic models for
social emotion mining[J]. Information Sciences, 2014,
266: 90-100.
Roberts M E, Stewart B M, Tingley D, et al. Structural topic
models for openended survey responses[J]. American
Journal of Political Science, 2014, 58(4): 1064-1082.
Sangwan N, Malik V, Sunesh, et al. A Survey for Text
Analytics Using R Language[J]. Advances in
computational ences and technology, 2017, 10(7):2173-
2182.
TANG Zhong-jun, CUI Jun-fu, TANG Xiao-wen, ZHU
Hui-ke. A Demand Feature Pattern Mining Method for
Short Life Cycle Experiential Product by Integrating
Content Analysis and Association Analysis[J]. Chinese
Journal of Management Science, 2019, 27(11): 166-175.
Tvinnereim E, Fløttum K. Explaining topic prevalence in
answers to open-ended survey questions about climate
change[J]. Nature Climate Change, 2015, 5(8): 744-
747.
Usama M, Ahmad B, Song E, et al. Attention-based
sentiment analysis using convolutional and recurrent
neural network[J]. Future Generation Computer
Systems, 2020, 113: 571-578.
Research on Demand Mining Method for Short Life Cycle Experiential Products Based on Structural Topic Model and Experience Value
673