https://doi.org/10.31449/inf.v48i5.5337                                              Informatica 48 (2024) 41–54  41 
Web News Media Retrieval Analysis Integrating with Knowledge 
Recognition of Semantic Grouping Vector Space Model  
Wenting Xiong 
School of Journalism and communication, Hunan Mass Media V ocational and Technical College, 
Changsha, Hunan, 410100, China 
E-mail: 2016150180@jou.edu.cn 
Keywords: vector space model, semantic grouping, searching for information, web news media retrieval  
Received: October 26, 2023 
Traditional Web news media retrieval technology can only meet the specific requirements of customers. 
Because of its universal characteristics, it cannot meet the needs of different environments, different 
purposes, and different times simultaneously. Researchers have proposed a search method for online 
news media, which is used for computing the semantic grouping vector space model. The customer's 
interest model is analyzed through the characteristics of the user's different classification areas. In 
this paper, we propose a vector space model that performs semantic grouping based on feature words. 
The model divides four groups that are relatively independent in the meaning of feature words in a 
news report: time, place, person, and event, and then forms four vector spaces and calculates the 
weight value and similarity of each vector space. Theoretical analysis and experimental results show 
that the improved model is suitable for searching Web news information and improves the calibration 
rate, query speed, and calibration rate.  
Povzetek: V študiji je predstavljen izboljšan model vektorskega prostora za iskanje spletnih novic, ki 
vključuje semantično grupiranje po ključnih besedah v nekaj kategorij.
 
 
 
1  Introduction  
Early search engines generally used content-based search 
methods, which were developed on the basis of the theory 
and technology of traditional information retrieval. The 
primary consideration is the relationship between web 
pages and search terms and the frequency and location of 
users querying documents. This method improves the 
search quality and accuracy to a certain extent, but these 
methods are based on keyword queries. Synonyms and 
polysemous words in natural language cannot be 
retrieved, so the search rate is not high and has certain 
limitations[1-2]. In recent years, the rapid progress of 
information technology has ushered in the digital age, and 
the digital and electronic files of past paper files are 
gradually being replaced. How to quickly and accurately 
query the user's demand information in the Web news 
media database? The current significant problem is how 
to deal with the issue of archive information supply and 
demand. Web news media retrieval is a helpful solution. 
By using a personalized search system to refer to the 
same information among users, new calculation methods 
for the semantic grouping vector space model can be 
discovered.  
 
 
 
 
 
The most important thing is the user type, but the 
participation of users is required because of the  
calculation method and collaboration Personalized search 
methods of semantic grouping vector space model have 
their characteristics.  
With the continuous development of the multimedia 
entertainment field, Web news media has gradually 
integrated into people's lives. However, how 
large-capacity Web news media data can better adapt to 
the environment and different user characteristics and 
introduce Web news media suitable for users is a 
challenge facing people. This paper presents the 
knowledge recognition of the semantic grouping vector 
space model in the Web news media retrieval process. It 
accurately conducts the Web news media retrieval by 
using the deep belief network algorithm in the audio 
segmentation stage of the Web news media. Finally, the 
results of the experimental analysis show that the method 
proposed in this paper can quickly retrieve the 
corresponding type of Web news media according to the 
user's preferences. 
 
 

42   Informatica 48 (2024) 41–54                                                                     W. Xiong 
References  Key findings Methodology Limitation 
10 The study focused on the Hermes 
Framework for Personalized News 
Services, with an emphasis on 
implementing Semantic Web 
standards and utilizing topic data. 
The study included 
approaches such as 
Ontology-based 
Knowledge Representation, 
Natural Language 
Processing (NLP) for 
Semantic Text Analysis, 
and Semantic Query 
Languages for Information 
Specification. 
The challenges included 
complexity, difficulty in 
acquiring knowledge, 
reliance on Semantic Web 
standards, and scalability 
concerns. 
11 The study examined the effective 
extraction of events from a news 
corpus, emphasizing the strength of 
Latent Dirichlet Allocation (LDA) in 
collecting semantic connections. It 
also evaluated the efficacy of 
Continuous Bag of Words (CBOW) 
and Skip-gram models in obtaining 
semantic similarity. 
The process of preparing 
the Online news corpus 
involved the utilization of 
diverse research 
approaches. 
The study addressed 
limitations within the news 
corpus, generalization 
difficulties, and reliance on 
algorithmic parameters. 
12 The research focused on gathering 
data using web crawling and 
extracting features using NLP 
techniques. 
The study comprised 
primary stages: Online 
News Corpus Preparation, 
NLP for Feature Extraction, 
and the Selection of 
Machine Learning Models. 
The study examined many 
factors pertaining to 
generalization, the 
ever-changing nature of fake 
news, and ethical 
implications. 
13 The study discussed the challenges of 
collecting high-level information, 
proposed a compact event 
representation, and introduced a 
hypothesis for the facilitation of 
social and geopolitical analysis in the 
context of the rise of the Social Web. 
The study emphasized the 
use of data mining tools to 
examine event 
representations, integrating 
case studies and user 
evaluation procedures. 
Factors including data 
velocity and volume, the 
study's dependence on data 
quality, and the extent of user 
examination were among the 
constraints. 
14 The study focused on the deployment 
of the Hermes Framework for 
Personalized News in the Hermes 
News Portal. 
Three main approaches 
were examined in the 
study: the implementation 
of the Hermes News Portal, 
semantic query languages, 
and semantic text analysis. 
The complex nature of 
complexity and the learning 
curve, as well as its 
dependence on the quality of 
ontology. 
15 The study concentrated on the 
challenges associated with traditional 
search engines and the integration of 
NLP. 
The study involved using 
Web Crawling and NLP 
techniques to gather and 
preprocess data, followed 
by applying Sentiment 
Analysis algorithms. 
The study factored in 
reliance on data sources, 
sentiment analysis 
limitations, and resolving 
plurality. 
16 The paper focused on the Semantic 
Model that employed Term 
Frequency-Inverse Document 
Frequency (TF-IDF) and evaluated its 
performance using evaluation metrics. 
The research focuses on 
three key aspects: data 
preprocessing, the 
utilization of machine 
learning and deep learning 
algorithms, and the creation 
of a robust prediction 
model. 
The study presented 
challenges in generalizing 
findings, allocating 
resources, and accounting for 
temporal variables. 
17 The study focused on analyzing the 
level of financial area analysis and the 
importance of material on news 
websites. 
The study utilized various 
methods such as data 
collection, the 
implementation of a 
supervised machine 
The Study complexities of 
the Lithuanian language 
presented research 
restrictions due to its 
algorithmic sensitivity. 

Web News Media Retrieval Analysis Integrating with Knowledge…                      Informatica 48 (2024) 41–54   43                  
learning model, and 
hyper-parameter 
optimization using Grid 
Search. 
18 The study focused on the various 
challenges faced by conventional 
search engines, the integration of 
sentiment analysis technologies, and 
the subsequent deployment of 
intelligent search capabilities. 
The research involved 
various methodologies, 
such as data collection, web 
crawling, Natural Language 
Processing (NLP) for text 
preprocessing, and the 
implementation of 
sentiment analysis 
algorithms. 
The research involved 
depending on the data source 
and carefully considering 
temporal aspects. 
19 The study's primary findings included 
an analysis of annual publication 
trends and subject distribution, highly 
cited literature and research hotspots, 
and an evaluation of module 
functionality and integration. 
The research involved 
analyzing annual trends and 
subject distribution, as well 
as creating a deep-learning 
model. 
The study's findings are 
limited due to assumptions 
made in vote prediction, the 
complexity of the model, and 
resource constraints. 
 
To overcome these challenges, we proposed a vector 
space model that focuses on semantic grouping based on 
feature words. The paper aims to organize news 
information into different categories based on the 
meaning of specific words. 
 
2  Semantic grouping vector space 
model  
The knowledge recognition process based on the 
semantic grouping vector space model is mainly for the 
problems that occur during database retrieval. If this 
problem is completely solved, the semantic grouping 
vector space model knowledge algorithm needs to be 
used to optimize the parameters of the model. This paper 
combines the strategy of weight sharing. Weight sharing 
refers to making the semantic grouping vector space 
model different in connection mode and parameter 
sharing mode to the conventional vector model. The 
semantic grouping vector space model can be locally 
connected and data information can be shared. Weight 
sharing mainly refers to the collection of parameter data 
based on multiple nodes in the hierarchical process. The 
feasibility analysis of shared data parameters is primarily 
related to various goals in the calculation process.  
Different from the traditional method, the semantic 
grouping vector space model mainly uses the initial  
 
 
 
 
 
 
 
 
 
 
 
feature value of the input signal of the collected data and 
learns the input signal according to the hierarchical 
retrieval [3]. Usually, it includes the average time 
amplitude difference of the time domain features. This 
article uses the energy in a short time, etc., for the initial 
input characteristics of the model.  
The Web news media search signal can be represented by 
x(n), and the short-term energy balance can be expressed 
as follows: 
 
 
( ) ( ) ( ) ( ) ( ) ( )
2 22
n
mm
E x n w n m x m h n m x m h n

= −  = − 
= − = − =  


     
(1) 
 
However, the high dimensionality of the time-domain 
features under initialization will cause a lot of 
interference and noise. Therefore, the input search signal 
needs to be reduced in dimensionality.  
The primary cause analysis method is used to make 
statistics on the multiple variables of the investigation, 
and the internal structure among various variables can be 
analyzed by studying multiple main components. After 
the data of Web news media is processed by 
dimensionality reduction, the input data information can 
be retrieved. In deep learning, the semantic grouping 
vector space model can perform data processing on the 
output data. Among them, the Web news media retrieval 
structure at this stage is shown in Figure 1.  
 

44   Informatica 48 (2024) 41–54                                                                     W. Xiong                                                              
 
Figure 1: Semantic grouping vector space structural 
diagram. 
2.1  Deep belief network algorithm in the 
audio segmentation stage 
The Deep Belief Network (DBN) algorithm, utilized in 
audio segmentation, employs a hierarchical, generative 
model consisting of multiple layers of stochastic, latent 
variables. Initially trained layer by layer, the DBN 
captures intricate patterns in audio data through 
unsupervised learning. The top layer of the network 
functions as a discriminative model, facilitating the 
identification of relevant features for segmentation. 
Through iterative fine-tuning of weights during training, 
the DBN learns hierarchical representations, enabling the 
extraction of meaningful audio features. This hierarchical 
approach enhances the algorithm's efficiency in 
discriminating between various audio segments, thereby 
improving its ability to segment and classify different 
components within the audio signal accurately. 
 
2.2  Knowledge recognition algorithm of 
semantic grouping vector space model  
A news report contains four elements: time, place, person, 
and event. Therefore, for Web news information, at first, 
feature words are distinguished based on these four 
elements, respectively define the associated semantic 
groups to form 4 vectors, and then determine which 
vector space each feature word belongs to, and establish 
an inverse index corresponding to each vector space, 
calculate the weight value and similarity of feature words 
for each meaning group. Finally, the weighted sum of 
similarity is obtained, and the search results that are 
greater than a specific valve value are sorted using link 
analysis[4-5].  
 
2.3  Weight and similarity of feature word  
In Web news information, the position of the feature word 
on the document is different, and the importance of the 
document ability expressed is also other. The feature 
word level describes this characteristic. If the feature 
word TE appears n times in the document, as in formula 2, 
the feature word level score
k
T is the k-th appearance 
level of the feature word T in the document.  
( )
ln
1
1
2
k
n
T
k
score T
=
=

                 (2) 
The weight of the feature word is  
( ) ( )
k
w T score T idf =              (3) 
In the case of calculating the weighted similarity between 
the query Q and a specific news feature word group D, 
due to the large amount of calculation and time overhead 
of the conventional VSM similarity, the ratio of the 
weight value of the QD cross part to the sum of the QD 
weight value is used for calculation (4). 
( )
( ) ( )
( )
( ) ( )
1
11
,
QD
qk dk
k
QD
qk dk
kk
w T w T
sim Q D
w T w T
=
==
+
=

+





     (4) 
2.4  Semantic grouping vector space 
calculation method  
For the vector space model, the conventional method of 
semantic grouping vector space calculation is to calculate 
the cosine similarity between vectors. The semantic 
grouping vector space of user u and Web news media d 
can be defined as:  
 
( ) ,
ud
Sim u d
ud

=

             (5) 
 
Regarding the probability model, the cosine similarity of 
vectors cannot be calculated by self-connection[6-7]. The 
following propositions are proposed to express the 
diversity of user interests.  
Proposition 1. Assuming that user you have conditions 
independent of multimedia digital archive d in the 
predetermined classification model  
12
, , ,
n
C c c c = , 
the probability that multimedia digital archive d 
recommends to user u is:  
 
( ) ( )
( ) ( )
( ) 1
n
jj
j
j
p c u p c d
p u d p u
pc
=
=

    (6) 

Web News Media Retrieval Analysis Integrating with Knowledge…                      Informatica 48 (2024) 41–54   45                  
Proof: It can be known from the total probability formula,  
( )
( ) ( )
1
,,
n
jj
j
p u d p u d c p c
=
=

     (7) 
Assuming that the user u exists independently in the 
multimedia digital file d under condition C, 
so
( ) ( )
,
jj
p u d c p u c = , and 
then
( ) ( ) ( )
,
j j j
p u d c p u c p d c = is obtained. 
Therefore, formula (7) can be transformed into  
( )
( ) ( ) ( )
1
,
n
j j j
j
p u d p u c p d c p c
=
=

   (8)
 
Get 
( )
( ) ( ) ( )
( )
1
n
j j j
j
p u c p d c p c
p u d
pd
=
=

   (9) 
The purpose is to transform the semantic grouping vector 
space problem of the probability model into a situation of 
seeking conditional probability, presenting the diversity 
of user interests.  
The adopted system has the memory of recording the 
user's search history and clicks and continues to search 
for the data information source of the user's operation 
behavior model. The system automatically completes this 
coherent operation, and the user experience is not 
disturbed. First, the user's historical search information in 
the browser is saved to learn the user's interest, and then 
the user's interest in the search information through the 
user's operation on the search results. Add time stamps to 
the data of interest. Therefore, update the points of 
interest that users need to be more interested. In the user 
interest model, the design process is shown in Figure 2.  
 
Figure 2: Design process of user modeling. 
After completing the Chinese word segmentation with the 
IK Analyzer Chinese, the vector space model is 
constructed, and the weight 
i
w can be obtained through 
the frequency of occurrence of keywords 
i
k in the 
document through the calculation formula of TF-IDF. 
idf
i i i
w tf =            (10) 
Among them, 
i
tf represents the frequency of the 
keyword 
i
k appearing in all generated texts and 
id
i
f represents the frequency of
i
k in the reverse order of 
all generated texts. The calculation method is as follows:  
idf log
i
N
n
=               (11) 
 
 
 
 
 

46   Informatica 48 (2024) 41–54                                                                     W. Xiong                                                              
Where N is the number of generated texts, and n is the 
number of all texts containing the keyword
i
k .  
The method of calculating the rights of keywords 
i
k can 
be adjusted as follows.  
t
ii
w w e
−
=                (12) 
( ) ( ) ( ) ( )
1 1 2 2
, , , , , ,
ii
d k w k w k w    =   (13) 
When comparing the data of the model and the document 
( )
12
, , ,
n
X x x x that the user is interested 
in ( )
12
, , ,
n
W w w w , the size of θ is evaluated by 
calculating the angle θ of the vector and 
( )
12
, , ,
n
X x x x composition, which is inversely 
proportional to the degree of user concern. The smaller 
the θ, the higher the correlation between this file and the 
user's interests and preferences. The calculation formula 
is as follows.  
( )
1
22
11
, cos
n
ii
i
nn
ii
ii
XW
sim X W
XW

=
==
==
   
   
   


    (14) 
2.5  Web news media retrieval analysis  
In multimedia digital archives, user Web news media 
retrieval is completed by three stages: collecting and 
analyzing user data, constructing user interest model, and 
updating user interest model. Suppose Web News Media 
wants to obtain user information. In that case, it must first 
get the user's obvious information, such as the user's 
registered account number, age, education, occupation, 
unit, keywords of interest, etc[8-9]. The user can modify 
and reply to this significant information, to gradually 
improve the information. However, some users are 
unwilling to provide accurate registration information due 
to personal privacy or time issues. To solve this problem, 
you can set up implicit information to extract the user's 
knowledge. The test will remove the bookmark of the 
keyword searched for, download and save the files, etc. 
According to the bookmarks maintained by the user, the 
downloaded and saved document information, the user's 
long-term concern, research field, and other issues can be 
determined, which can become important sources of 
information for establishing a model.  
Because the user's interest is not fixed, it is necessary to 
establish an update mechanism when building a model to 
remove the forgotten topics in time to add new content, 
calculate the weight of the user's interest, and rank them 
according to the proportion of the weight. People's 
forgetting value is the trend of forgetting from the 
beginning and gradually becoming late. In the interest 
model system, the weight of the keyword of interest is 
multiplied by the update time, the weight of the phrase is 
sorted, and the forgotten interest topics are deleted. 
Complete the tracking of the effective behavior of the 
user and harvest a new keyword to recalculate the 
proportion. If the weight exceeds the threshold, it is 
added to the user interest model to complete the model 
update. It can be proved from Proposition 1 that 
according to the results of the ranking query based on the 
recommended ratio, the semantic grouping vector space 
model calculation can be used to query the media digital 
archive users. Based on p(u) of inequality (9), since p(u) 
does not interfere with the results of the recommendation 
probability, according to this method, the detailed 
explanation of the retrieval calculation of the multimedia 
digital archive users is performed.  
Algorithm 1. Web news media retrieval analysis 
algorithm based on knowledge recognition of semantic 
grouping vector space model.  
Input: domain classification model, user interest model, 
retrieval keywords, search engine, output: multimedia 
digital archive users' Web news media search results.  
(1) According to the search keywords, the search engine 
is used to generate a preliminary search result set X.  
(2) Set the number of iterations i=0.  
(3) For the i-th Web news media in the set X, formula (1) 
is used to calculate the probability distribution in the 
field's classification model.  
(4) Equation (9) is used to calculate the probability that 
the multimedia digital file i is recommended to the 
current user and added to the list Y .  
(5) If the multimedia digital file i is the last multimedia 
digital file in the set X, go to (6); otherwise, set i=i+1 and 
return to (3).  
(6) Sort and output the multimedia digital files according 
to the probability in the list Y in descending order.  
Because the algorithm is actually based on another search 
engine, for each multimedia digital archive of search 
results, the probability distribution in the domain 
classification model must be calculated. This has a great 
impact on the performance of the algorithm. If the search 
engine calculates the probability distribution in the 
domain classification model of each Web news media in 
advance, the performance of the algorithm will be 

Web News Media Retrieval Analysis Integrating with Knowledge…                      Informatica 48 (2024) 41–54   47                  
significantly improved to meet the needs of real-time 
processing.  
3  Web news media retrieval analysis 
process 
3.1  Web news media retrieval  
The four parts of the browser plug-in, personal manager, 
user model learner, and information personalized searcher, 
constitute the experimental system; as shown in Figure 3, 
it is the browser plug-in that provides users with 
convenient tools. After the user logs in and register 
information, the browser plug-in can be used to complete 
the Web news media retrieval of multimedia digital 
archives—no need to log in to the server. In addition, the 
browser plug-in mainly collects the user's personal 
information and transmits it to the server. The personal 
manager is used to manage the user's personal 
information, hobbies, and bookmarks through the 
personal manager. The purpose of tracking user behavior 
is to learn user interests. The information Web, a news 
media retriever, can complete the user's query and 
recommendation in the multimedia numbers calculated 
by the semantic grouping vector space model.  
 
 
Figure 3: System architecture. 
Our system can track the behavior of guests; it is 
distributed on the edge of the client and server and will 
not affect the customer's reading and system 
performance. 
3.2  News page judgment and related 
information extraction information 
extraction  
Through retrieval and learning, the link weights and node 
offsets of each layer are obtained, and the network 
initialization is completed. The reverse conduction 
algorithm (BP)is adopted, and the deep trust network 
model monitored from top to bottom is fine-tuned to 
overcome the shortcomings of local optimization and 
long search time. Although the performance of the deep 
belief network model shows strong characteristic learning 
ability, from the above principles, Internet search requires 
a large amount of sample data to generate more parameter 
values. On the other hand, based on the problem of Web 
news media recommendation search, there is a lack of a 
large amount of sample data, and it is found that the 
generation of a large number of parameters takes a long 
time, which is not good for practical applications. During 
the calculation process, the feature vector is used to 
represent the web page, and if the keyword weight We are 
determined by the TF*IDF method, and the term item is 
determined to be a named entity, the weight value shall be 
appropriately enhanced. The specific definition is as 
follows.  
 , 
,
e
e
e
is named en i tity
idf e is
df
ot e
W
s
e
hr
  
=


    (15) 
Among them, α is the weighting factor, which is 5 in this 
experiment.  
Finally, if m word items with large weight values are 
selected to generate web page feature vectors and 
applications. The number of shared term items in the two 
webpage feature vectors is used as the basis for judging 
similarity. If the number of shared terms is greater than 
the threshold, the two web pages are similar.  
After determining the reprinting or similar relationship, 
relevant information is extracted and recorded. The 
leading information recorded is the reprinted website, the 
source site of the reprinted website, the number of 
responses to the reprinted website, and the time of the 
news release. The reprinted website and the source site 
here are only records of the reprinting relationship, not 
the finalized accurate source site and reprinted website. 
The last source site will be determined in the next step.  
3.3  Judgment of news reprinting 
relationship and calculation of authority of 
news source sites  
()
  
     
  
reprint times
News reprint rate denoted as Trams
source website clicks
=
(16) 
However, since the reprinting relationship of Web news 
has two types, direct reprinting, and indirect reprinting, 
the source site cannot be determined initially, and the two 
attribute values of all nodes are initialized to 1 in the 
entire network layer. Then, in pt →qt, the website pt 
describes that the news of the website qt is reproduced. 
The content quality attribute value and the reprint 
attribute value are calculated using the following 
repetitive formula. The attribute value of all web pages is 
normalized to 1 when each iteration is completed.  
( ) ( )
01
qt pt
A pt A qt
→
=

         (17) 

48   Informatica 48 (2024) 41–54                                                                     W. Xiong                                                              
( ) ( )
10
pt qt
A pt A qt
→
=

          (18) 
( )
( )
( ) ( )
0
0 1
2
2
0
pt
A pt
A pt
A pt

=




    
(19) 
( )
( )
( ) ( )
1
1 1
2
2
1
pt
A pt
A pt
A pt

=




      (20) 
Iteratively update the attributes of each node 
( )
0
A pt ( )
1
A pt according to the above formula.  
The extracted reprinting information is used first to 
extract the relationship between news reprinting sites, 
calculate the authority value of each reprinting site, and 
use the website with the most reprinting times as the 
source site, including the relationship between direct 
reprinting and indirect reprinting, and that authority value 
is treated as the value of the reprint rate of the news.  
3.4   Calculation of new response rate  
The response rate (denoted as Rep) directly reflects 
people's reaction to Web news. usually  
  
 
  
amount of responses
Response rate
number of clicks
=
    (21) 
Observation results show that most news pages only 
provide the number of answerers rather than the number 
of clicks/viewers. The number of clicks/views on the 
page is stored on the page server-side and cannot be 
obtained through simple capture and information 
extraction. Based on a large number of observations, a 
response rate ratio is summed up based on the relative 
number of news responses, and this ratio is used as the 
news response rate. Here, the number of responses is the 
total of the number of responses from the source site and 
the number of responses from the reprint website[8-9]. 
Figure 4 shows the distribution of the number of news 
responses.  
 
Figure 4: Statistics of the person-time number of 
responses. 
It can be seen from Figure 4 that the number of responses 
to most news is within 1,000 people. There is very few 
news with more than 3,000 people. According to the 
statistical rules in the figure above, the relative recovery 
rate values are shown in Table 1. As an example, the 
number of responses (0~500) indicates the range of the 
number of people who responded to this event, and the 
relative response rate indicates that the number of people 
who responded to this event is between (0~500). It is 
considered that the number of people who responded to 
this event accounted for 5% of the number of viewers. If 
there are more than 5,000 respondents, those who have 
read this report will basically give answers, and the 
relative answer rate is 100%.  
Table 1: List of relative response rates. 
Number of 
responses 
Relative 
response-rate (%) 
5000~ 100 
4500~5000 90 
4000~4500 80 
3500~4000 70 
3000~3500 60 
2500~3000 50 
2000~2500 40 
1500~2000 30 
1000~1500 20 
500~1000 10 
0~500 5 
 
 
 
 

Web News Media Retrieval Analysis Integrating with Knowledge…                      Informatica 48 (2024) 41–54   49                  
3.5  The influence of time factors on news 
ranking  
There are usually two trends in people's interest in news, 
as shown in Figure 5. The attention here is measured by 
the number of news viewers per unit of time. The first is 
the slow-growing type of interest in knowledge such as 
national policy news. The timeliness of news in these 
categories is not strong, and people's concern is slowly 
increasing with time. The other is the type that grows 
rapidly and declines. It is mainly for news on current 
events; this kind of news is very time-sensitive. People's 
attention to this kind of news has increased rapidly in a 
short period, and after some time, the attention has 
quickly dropped [10-12]. Therefore, the sorting of news 
must first be classified and judged, taking into account 
the influence of time. From this perspective, the 
importance of news is inversely proportional to the time 
of publication.  
 
 
Figure 5: News attention. 
In addition, the longer the release time, the higher the 
probability of reprinting and replying and the greater the 
number of responses and reprints. If the time factor is not 
taken into consideration, it is unfair to the newly 
published news. Therefore, when selecting parameters, 
the time factor will affect the importance of news. For 
reports with a long submission period, the number of 
responses and reprints will be reduced. Summarizing the 
above two points, combined with the definition of the 
news decay time parameter in literature [4], the definition 
of the time parameter is as follows.  
( )
( )
,
s
tt
s
D t t e
 −−
=                 (22) 
Among them is the publication time of the news. The 
determination of α depends on the recession time of the 
news category to which the news belongs. Recession time 
refers to the time from the news release to the 
intermediate experience that no one cares about, and it is 
defined here. The relationship between news and 
recession time is 
s
t
s
tt  
0 .5,
0.5,
 
 
Current affairs news
Non current
e
ws e ne



−
−
−
 =

=

=


 (23) 
Here β is the decline time of current affairs news. γ is the 
decline time of non-current affairs news.  
 3.6  Judgment of news influence  
Through the above steps, data on news reprint rate, news 
response rate, and influence factor of news source sites 
can be obtained(
S
W ), as well as the time parameter of 
the news release ( ) ,
s
D t t .  
Reprinting and replying to news are considered to be the 
recognition of news, so the Web news recognition rate 
(denoted as Rec) is defined as news recognition rate = a. 
×Reload rate+b×Recovery rate.  
In order to ensure that the authorization rate is less than 1, 
the relationship between a and b is defined as a+b=1. 
Since there is no suitable corpus, the values of a and b 
cannot be obtained through the training method so these 
decisions can be obtained according to the 80/20 rule. 
There may be a lot of people watching the news, but few 
people answer it, and even fewer people understandably 
do repost. So, I think the reprint rate can better reflect the 
influence of news. Experiments show that this definition 
method is feasible[13-14].  
Finally, combining the above information, define the 
influence of news (NF) as follows.  
( )
( ) Rep
s
tt
FS
N e W a Trans b
 −−
=    + 
  (24) 
4  Experiment results and analysis  
The performance evaluation of the information retrieval 
system is generally used as a benchmark, and the 
comprehensive evaluation rate F can also be used for 
evaluation (25)  
2 precision recall
F
precision recall

=
+
        (25) 
4.1  Experimental data set  
The top nine news information websites were tracked on 
the Chinese website rankings for a week. As many 
features based on the summary of Chinese webpages 
appeared in the algorithm research, Chinese webpages 
were still used as experimental subjects in the experiment. 
The news on the homepage of these nine websites is 
captured every hour. The list of captured experimental 
data is shown in Figure 6.  

50   Informatica 48 (2024) 41–54                                                                     W. Xiong                                                              
0 500 1000 1500 2000
Tencent News
Sina News
Xinhua
PRC News
Sohu News
NetEase News
Dongfang News
Tom News
China News
News page volume
 
Figure 6: List of experimental web pages. 
After the internal deduplication of the website, these 
pages are classified into six types according to their 
content. The news distribution of each category is shown 
in Figure 7. Here, the strength of news timeliness is 
obtained from the attention model to which news belongs. 
0 10 20 30 40 50 60
Events/News
Business
Health
Sports/Entertainment
Military
Technology
Percentage of total (%)
 
Figure 7: Classification list of experimental web pages. 
4.2  Experimental result  
The recommended task is executed using Python 3.5. 
Numpy, Pandas, Scikit-learn, Natural Language Toolkit 
(NLTK), and Matplotlib software are required to be 
installed alongside Python to carry out the procedure 
(1) News influence ranking for the week from September 
10 to 16th, 2007  
Figure 8 shows the top 10 news and their influence values 
in the week from September 10 to 16th, 2007. Here, the 
recession time of current affairs news is defined as 72 
hours instead of 120 hours. Figure 9 shows the 
distribution of news influence values within a week.  
0.6366
0.6168
0.6168
0.6108
0.6108
0.6108
0.6093
Shangha …
Chinese …
Chinese …
Japanes …
A study …
Chinese …
The CPC …
4 5 6 7 8 9 10
Force value 0.682 0.6517 0.6517.
 
Figure 8: TOP10 news list from September 10 to 17th, 
2007. 
 
Figure 9: Distribution of news influence values collected 
from September 10, 2007, to September 16, 2007. 
(2) Designated news influence ranking  
The algorithm is also suitable for the sorting of 
designated news, giving some news on different topics, 
and using the topics of these news as keywords to search 
for relevant news pages using popular search engines. 
Select the top 50 from the search results for statistical 
calculations. The reason for ranking in the top 50 is that 
basically all reprinted pages and almost all similar pages, 
as well as the top 50 websites with news information in 
the Chinese website rankings, are included. It is sufficient 
to determine the source website of the news and the 
reprint rate of the news. After browsing many web pages, 
it can be found that all the comments of netizens are on 
the top websites. Netizens on other websites make almost 
zero responses, so selecting these pages can also get a 
more accurate news response rate value. After obtaining 
each news topic and reprinting the page, the relevant 
information is extracted to analyze the influence of each 
topic news according to the above algorithm, find the 
influence coefficient, sort according to the influence 
coefficient, and obtain the ranking result of the 
quantitative analysis. Next, we investigated the sorting 
results of these topics by multiple people. After synthesis, 
we got the sorting results of manual qualitative analysis. 
Finally, the consistency of the two results is compared. 
From the results of multiple comparisons, it is found that 
the sorting results calculated by this method are almost 
the same as the manual sorting results. The comparison 
results are shown in Figure 10 and Figure 11. Here is a 
comparison of the influence of news on non-related 
topics. Experiments show that this method is also 
applicable to related topics.  

Web News Media Retrieval Analysis Integrating with Knowledge…                      Informatica 48 (2024) 41–54   51                  
7.27
7.27
7.24
7.19
7.25
7.19
7.2
7.25
7.21
7.29
NPC …
Seventy …
Basic …
The …
The SDR …
Shandon …
Pork …
Chinese …
The …
Iraq held …
1 2 3 4 5 6 7 8 9 10
Post time
 
Figure 10: Results of manual ranking of designated news 
influence. 
7.27
7.27
7.24
7.19
7.25
7.29
7.19
7.2
7.25
7.21
0.6014
0.5774
0.0833
0.072
0.0341
0.0043
0.0042
0.0038
0.0037
0.0035
NPC …
Seventy …
Basic …
The …
The SDR …
Iraq held …
Shandon …
Pork …
Chinese …
The …
1 2 3 4 5 6 7 8 9 10
Force value Post time
 
Figure 11: Specify the results of Web news media 
retrieval analysis ranking. 
Query speed is a metric that measures the efficiency of a 
system in processing and responding to user queries. It 
quantifies the time it takes for a system to execute a 
search or retrieval operation and deliver the relevant 
results to the user. Table 2 and Figure 12 show the query 
speed result. While comparing the proposed method 
(SGVSM - 5 sec) with the other existing methods 
(Generalized Vector Space Model (GVSM )- 10 sec, TF- 
IDF -12 sec), it shows that our proposed method is 
superior for prediction accuracy in Web News Media 
Retrieval to other methods. 
Generalized Vector
Space Model
TF-IDF Vector Space
Model
[proposed] Semantic
Vector Space Model
4
5
6
7
8
9
10
11
12
13
Query Speed (sec)
Model
 
Figure 12: Query speed 
 
 
 
 
 
Table 2: Query speed 
Model Query Speed 
(sec) 
Generalized Vector Space 
Model [20] 
10 seconds 
TF-IDF Vector Space Model 
[21] 
12 seconds 
Semantic Vector Space 
Model[proposed] 
5 seconds 
 
Calibration rate is a metric used to assess the accuracy of 
predicted probabilities in a predictive model. It measures 
how well the predicted probabilities align with the actual 
outcomes or events. Table 3 and Figure 13 show the 
Calibration rate results. While comparing the proposed 
method (SGVSM - 82%) with the other existing methods 
(GVSM – 70%, TF- IDF -68%), it shows that our 
proposed method is superior for prediction accuracy in 
Web News Media Retrieval to other methods. 
 
Figure 13: Calibration rate 
Table 3: Calibration rate 
Model Calibration 
Rate (%) 
Generalized Vector Space 
Model[20] 
70 
TF-IDF Vector Space 
Model[21] 
68 
Semantic Vector Space 
Model[proposed] 
82 
 
Accuracy is a metric that assesses the correctness of a 
proposed model's predictions by comparing them to 
observed values. The correctness of case predictions is 
measured as a percentage of complete occurrences. 
Figure 14 and Table 4 depict the comparative evaluation 
of accuracy in suggested and traditional methods. When 
compared to currently existing methods such as GVSM 

52   Informatica 48 (2024) 41–54                                                                     W. Xiong                                                              
and TF-IDF, which have accuracy values of 85% and 
82%, respectively, the suggested SGVSM achieves an 
accuracy value of 90%. Our proposed method provided 
superior results for Web News Media Retrieval. 
Generalized
Vector Space
Model
TF-IDF Vector
Space Model
[proposed]
Semantic
Vector Space
Model
64 72 80 88 96
Accuracy (%)
Model
 
Figure 14: Accuracy 
Table 4: Accuracy 
Model Accuracy (%) 
Generalized Vector Space 
Model[20] 
85 
TF-IDF Vector Space 
Model[21] 
82 
Semantic Vector Space 
Model[proposed] 
90 
 
Precision is a fundamental parameter utilized in the field 
of statistics to evaluate performance. Figure 15 and Table 
5 depict the comparative evaluation of precision in 
suggested and traditional methods. When compared to 
currently existing methods such as GVSM and TF-IDF, 
which have Precision values of 85% and 82%, 
respectively, the suggested SGVSM achieves an accuracy 
value of 90%. Our proposed method provided superior 
results for Web News Media Retrieval. 
 
Figure 15: Precision 
 
 
 
Table 2: Precision 
Model Precision (%) 
Generalized Vector Space 
Model[20] 
82 
TF-IDF Vector Space 
Model[21] 
78 
Semantic Vector Space 
Model[proposed] 
92 
 
Recall is a performance metric used in data categorization 
that represents the proportion of actual positives that a 
model correctly retrieves. The recall result is shown in 
Table 6 and Figure 16. When compared to currently 
existing methods such as GVSM and TF-IDF, which have 
Recall values of 88% and 85%, respectively, the 
suggested SGVSM achieves an accuracy value of 92%. 
Our proposed method provided superior results for Web 
News Media Retrieval. 
 
Figure 16: Recall 
 
Table 6: Recall 
Model Recall (%) 
Generalized Vector Space 
Model[20] 
88 
TF-IDF Vector Space 
Model[21] 
85 
Semantic Vector Space 
Model[proposed] 
92.4 
 
Computational time, also known as execution time or 
runtime, refers to the amount of time it takes for a 
computer program or algorithm to complete its execution. 
It is a crucial metric in evaluating the efficiency and 
performance of computational processes. The recall result 
is shown in Table 7and Figure 17. When compared to 
currently existing methods such as GVSM and TF-IDF, 
which have Recall values of 10.9 sec and 8.7 sec, 
respectively, the suggested SGVSM achieves an accuracy 
value of 6.4 sec. Our proposed method provided superior 
results for the Retrieval of news media from the web. 

Web News Media Retrieval Analysis Integrating with Knowledge…                      Informatica 48 (2024) 41–54   53                  
Generalized Vector Space
Model
TF-IDF Vector Space
Model
[proposed] Semantic
Vector Space Model
6
7
8
9
10
11
Computational Time (sec)
Model
 
Figure 17: Computational time 
 
Model Computational 
time 
Generalized Vector 
Space Model[20] 
10.9 
TF-IDF Vector Space 
Model[21] 
8.7 
Semantic Vector Space 
Model[proposed] 
6.4 
 
Table 7: Computational time 
 
5  Implication 
The combination of Web News Media Retrieval Analysis 
and Knowledge Recognition of Semantic Grouping 
Vector Space Model has significant practical applications 
for information retrieval and analysis in the field of 
online news media. This strategy seeks to improve the 
efficiency and relevancy of web news searches by 
integrating sophisticated approaches in information 
retrieval, semantic grouping, and knowledge recognition. 
The model enhances the extraction of significant insights 
by identifying semantic relationships within the material, 
allowing for more precise categorization and collection of 
articles based on their underlying knowledge structures. 
This integration enhances a news retrieval system by 
making it more sophisticated and contextually aware. 
Additionally, it has the potential to improve the user 
experience by delivering more coherent and informative 
results. Furthermore, this approach may be utilized in 
several domains, such as journalism, research, and data 
analysis, providing a valuable instrument for effectively 
navigating and understanding the extensive and 
ever-changing realm of web-based news media. 
 
6  Discussion 
The generalized Vector Space Model may struggle to 
capture semantic nuances and ambiguity in language. 
Words with multiple meanings or contexts may be 
represented by a single vector, leading to potential 
confusion. TF-IDF Vector Space Model faces challenges 
with synonymy and polysemy, where different words may 
have similar meanings or a single word may have 
multiple meanings. When implementing a proposed 
Semantic Vector Space Model, achieving superior 
performance in capturing semantic nuances and 
addressing the challenges of ambiguity becomes possible. 
This model endeavors to represent words in a 
multidimensional space, taking into account their 
semantic relationships, thereby offering a more nuanced 
and context-aware representation. By considering the 
inherent meanings and associations between words, the 
Semantic Vector Space Model aims to enhance accuracy 
and effectiveness. 
 
7  Conclusions 
Compared with previous vector space models, the vector 
space model based on the semantic grouping of feature 
words proposed in this paper is more accurate for Web 
news information and is suitable for the retrieval of Web 
news information systems. The calibration rate of 
document query, the comprehensive evaluation rate F, 
and query speed have been significantly improved. The 
current situation is personalized services. The general 
retrieval system in the past can no longer meet the 
retrieval requirements in different environments, 
purposes and different times. This paper has carried out a 
series of research and analysis on Web news media 
retrieval. Through experiments, we can see the 
interference factors for the calculation of the semantic 
grouping vector space model. Experiments have proved 
that the accuracy of analysis has been improved, and the 
interests and needs of users can be correctly expressed so 
that the accuracy of Web news media retrieval is further 
enhanced. Keeping up with real-time updates and 
delivering the latest news can be a challenge. Systems 
might struggle to provide up-to-the-minute information 
due to processing delays or the dynamic nature of news. 
In future research, systems may focus on improving 
semantic understanding to provide more accurate and 
contextually relevant results. This could involve 
advanced natural language processing techniques, 
including sentiment analysis, entity recognition, and topic 
modeling. 
 
Acknowledgements 
2022 Hunan Natural Science Foundation "Research on 
the Construction of Teaching Innovative Team of Higher 
V ocational Teachers Based on the Background of 
Improving Quality and Cultivating Excellence"
（No.: 2022JJ60020 ） 

54   Informatica 48 (2024) 41–54                                                                     W. Xiong                                                              
Data Availability  
The data used to support the findings of this study are 
available from the corresponding author upon request. 
 
Conflicts of Interest 
The authors declare no conflicts of interest 
 
Funding Statement 
This study did not receive any funding in any form. 
References  
[1] Solainayagi, P., & Ponnusamy, R. (2019). 
Trustworthy media news content retrieval from web 
using truth content discovery algorithm. Cognitive 
Systems Research, 56(AUG.), 26-35. 
[2] Solainayagi, P., & Ponnusamy, R. (2019). 
Trustworthy media news content retrieval from web 
using truth content discovery algorithm. Cognitive 
Systems Research, 68(5), 566-588. 
[3] Kai, S., Wang, S. , &  Liu, H. . (2018). 
Understanding User Profiles on social media for 
Fake News Detection. FakeMM'18 Workshop, 
31(6), 1-7. 
[4] D Corney, Gonzalo, J., Martinez, M. ,  Poblete, B. , 
&  Valochas, A. . (2018). Recent Trends in News 
Information Retrieval, 24, 1012 –1017. 
[5] Davies R. J. (2016). Digital computing techniques in 
the manufacture and operation of engine 
management systems. Aeronautical Journal, 
79(776), 349-353. 
[6] Solainayagi, P. , &  Ponnusamy, R. . (2019). 
Trustworthy media news content retrieval from web 
using truth content discovery algorithm. Cognitive 
Systems Research, 56(8), 26-35. 
[7] Chen, H. ,  Huang, B. ,  Liu, W. Z. ,  Gao, Y. B. 
, &  Jiang, X. Y. . (2019). Python-based web news 
crawler and retrieval. Software Guide, 47(1):1-38. 
[8] [Dewandaru, A. ,  Supriana, I. , &  Akbar, S. . 
(2017). Keyword and event extraction for thematic 
map retrieval from indonesian online news site. 
Journal of Physics Conference, 38(8), 10320-10342. 
[9] George Dimitrakopoulos, Panagiotis Demestichas, 
& Vera Koutra. (2012). Intelligent management 
functionality for improving transportation efficiency 
by means of the car pooling concept. IEEE 
Transactions on Intelligent Transportation Systems, 
13(2), 424-436. 
[10] Fei-Yue Wang. (2011). Intelligent systems and 
social management. IEEE Intelligent Systems, 
26(6), 2-3. 
[11] Meesad, P. . (2021). Thai fake news detection based 
on information retrieval, natural language 
processing and machine learning. SN Computer 
Science, 2(6), 1-17. 
[12] Jing Fan, Tianyang Dong, Xinxin Guan, & Ying 
Tang. (2013). A rapid simulation system for 
decision making in intelligent forest management. 
IEEE Intelligent Systems, 28(5), 2-9. 
[13] Pe?A-Araya, V. ,  Quezada, M. ,  Poblete, B. , &  
Parra, D. . (2017). Gaining historical and 
international relations insights from social media: 
spatio-temporal real-world news analysis using 
twitter. Epj Data Science, 6(1), 25. 
[14] Abbas Rajabifard, Russell G. Thompson, & Yiqun 
Chen. (2015). An intelligent disaster decision 
support system for increasing the sustainability of 
transport networks. Natural Resources Forum, 
39(2), 83-96. 
[15] Frasincar, F., Borsje, J. and Levering, L., 2009. A 
semantic web-based approach for building 
personalized news services. International Journal of 
E-Business Research (IJEBR), 5(3), pp.35-53. 
[16] Nkongolo Wa Nkongolo, M., 2023. News 
Classification and Categorization with Smart 
Function Sentiment Analysis. International Journal 
of Intelligent Systems, 2023. 
[17] Bangyal, W.H., Qasim, R., Rehman, N.U., Ahmad, 
Z., Dar, H., Rukhsar, L., Aman, Z. and Ahmad, J., 
2021. Detection of fake news text classification on 
COVID-19 using deep learning 
approaches. Computational and mathematical 
methods in medicine, 2021, pp.1-14. 
[18] Štrimaitis, R., Stefanovič, P., Ramanauskaitė, S. and 
Slotkienė, A., 2021. Financial context news               
sentiment analysis for the Lithuanian 
language. Applied Sciences, 11(10), p.4443. 
[19] Sun, N. and Du, C., 2021. News text classification 
method and simulation based on the hybrid deep 
[20] Tsatsaronis, G. and Panagiotopoulou, V., 2009, 
April. A generalized vector space model for text 
retrieval based on semantic relatedness. 
In Proceedings of the Student Research Workshop 
at EACL 2009 (pp. 70-78). 
[21] SUBA, C., Retrieval of Information Document 
Using TF-IDF Algorithms and Vector Space Model 
Representation.