Frontiers in Computing and Intelligent Systems 
ISSN: 2832-6024 | Vol. 3, No. 3, 2023 

 

38 

Quantum-based Detection of Higly Semantically Similar 
Social bot 

Yulin Liu* 

Cyberspace Security College of Chengdu University of Information Technology, Chengdu, Sichuan 610225, China 

* Corresponding author Email: 1045734823@qq.com 

 

Abstract: Social bot often publish content that has a high degree of similarity in text or semantics. Combining this feature, 

this paper designs a method to detect social bot. First, the user-published text is used as the input to use sentiment analysis and 

data cleaning to predict the content. Processing and classification, and then adding the quantum similarity algorithm in the 

emerging quantum discipline in recent years to the structural system of social bot detection to perform similarity clustering, 

which greatly improves the operating efficiency of the system, and then obtains the corresponding user attribute characteristics 

for artificial intelligence. Intelligent algorithm classification, on the collected real dataset, marked two social bot groups for 

machine learning classification and detection, and listed two traditional similarity algorithms for comparison, the results show 

that the quantum similarity results Compared with the two traditional similarity algorithms, it has improved, and the average 

accuracy improvement was about 2%. 

Keywords: Artificial intelligence; Social bot; Detection; Machine learning. 

 

1. Introduction 

In recent years, with the development of Internet 

technology and the substantial improvement of living 

standards, people's tools for obtaining information have 

gradually changed from the previous TV, newspapers, etc. to 

the current network application platform. As far as the current 

social platforms are concerned, such as Twitter, Facebook and 

Sina weibo makes it easier for users to express their own 

opinions and understand other people's opinions through the 

Internet, and provides convenience for users to contact family 

and friends, so that social platforms have gained a lot of traffic 

and attention, and this trend will continue to accelerate in the 

future. In addition to the increasing number of users, it has 

brought many hidden dangers, among which the public 

opinion influence of social bot is a very serious problem. 

A class of users controlled by artificial software programs 

is called social bot. At first, social bot was used to provide 

services to users, such as automatically replying to user 

questions. However, the rise of malicious social bot has a 

negative impact on other users and public opinion security 

and even countries. Security poses a great threat. 

Manipulators mainly manipulate malicious social bot to do 

three things out of profit: (1) Focus on employer users. In 

order to make the account look popular and well-known, 

employer users will hire operators to follow them and make 

their own fans. A large amount of expansion, which may 

cause other normal users to pay attention to employer users. 

(2) A large number of false comments, false advertisements 

and rumors are released to mislead other normal users. For 

example, during the new crown pneumonia, malicious social 

bot posted more negative content on the Internet than humans, 

thus amplifying people's emotions and causing panic, and 

even leading the blame to China [1]. (3) The third category is 

a very dangerous category. The operators are malicious social 

bot participating in national politics. The most famous one is 

the 2016 US election. About 19% of related tweets are from 

social bot. These Social bots not only affect electoral politics, 

but also the economy and society [2]. For social bot, 

especially malicious ones, the research work of detecting 

them can not only lead the correct public opinion orientation, 

but also further promote the stability of social order, so the 

research on its detection work is very necessary. 

This paper mainly focuses on the detection of highly 

semantically similar types of social bot, which must satisfy a 

specific topic, and the topic meets two characteristics (1) has 

value, which means that it has research value, and its topic 

can change the direction of public opinion and the safety of 

public opinion to a certain extent. cause impact. Such as the 

new crown virus, Huawei 5G. (2) It is contagious. The literal 

meaning here is that the topic can spread widely. Because of 

the high semantic similarity, the experiment will use the 

similarity algorithm for text clustering to filter normal users. 

The traditional similarity algorithm is effective in text, but 

lacks in semantics and cannot be taken into account. 

This paper proposes to use the quantum similarity 

algorithm [3] that takes both semantic and textual similarity 

into account in the emerging quantum disciplines for 

experiments, and selects two traditional similarity algorithms 

for comparative experiments. The results show that the 

quantum similarity results Compared with the two traditional 

similarity algorithms, it has improved. 

2. Literature References 

At present, the mainstream social bot detection is based on 

machine learning and graph detection methods. 

2.1. Machine learning based detection 

Jorge [4] et al. used a single-class classifier to detect social 

bot, and the accuracy rate exceeded 0.89, and proved by 

experiments that in the face of the diverse types of social bot 

and the status quo of continuous anti-detection technology 

changes, single-class detection is better than binary 

classification. better adaptability. Due to the dataset 

imbalance problem, BinWu[5] et al. proposed an improved 

conditional generative adversarial network to expand the 

dataset, that is, combining Wasserstein distance with gradient 

penalty and clustering algorithms as a new data augmentation 



 

39 

method, In order to improve the accuracy of social bot 

detection, experiments show that the F1 score is 97.56%. 

Feng et al. [6] proposed a method called BotFlowMon to 

detect social bot from the perspective of network data traffic. 

The experimental results show that BotFlowMon has an 

accuracy rate of 96.1% and can see social bot in an average 

of 0.71 seconds. traffic. 

Compared with traditional machine learning detection 

methods, deep learning detection methods are more suitable 

for processing big data. Wu et al. [7] designed a social bot 

detection structure based on deep neural network and active 

learning. 30 features were extracted from four types of tweet 

content, tweet comments, likes and other interactive 

information, and time information to detect social bot, 

including 9 self-created features, and an active learning 

method was designed to effectively expand data labels, the 

accuracy rate reached 98.87%. Sneha Kudugunta et al. [8] 

proposed a deep neural network based on a contextual long-

term short-term memory (LSTM) architecture, which takes 

both the tweet content and the user's personal information as 

input, and uses fewer features to detect social bot, but the 

accuracy of the results is good, exceeding 96%. Greeshma 

Lingam [9] et al. designed a deep Q-network model, which 

takes the content attributes of tweets, user personal 

information content attributes and social relationship 

attributes as input and deep Q-learning model. Combined, 

based on the Q-value function to detect social bot, the overall 

accuracy rate has reached 93%. 

2.2. Graph-based detection methods 

In graph-based detection, social bot are generally referred 

to as sybils and are often used to forge or create multiple 

identities to launch malicious campaigns. Gao et al. [10] 

designed a framework, SybilFrame, which is mainly based on 

graphs and combined with machine learning as an auxiliary 

method to detect Sybil. SybilFrame is a multi-stage 

classification method. The first stage extracts useful 

information on the dataset, the second stage is the post-

inference layer. Experimental results on large-scale datasets 

show that SybilFrame performs much better than previous 

methods, but the disadvantage is that the accuracy depends on 

the external classifier. 

The classic graph-based detection Sybil method believes 

that Sybil is only connected to a small number of normal users, 

and this view does not hold in real social networks. In order 

to deal with this problem, Zhang [11] et al. developed 

SybilSAN, a two-layer model, It is intended to make full use 

of the interaction information and connections between 

normal users and Sybil, and propose a more practical Sybil 

attack model. Based on the attack model, three random walk-

based algorithms are coupled to detect Sybil, and SybilSAN 

also has Stronger detection robustness, but performance 

degrades when the Sybil population is not concentrated and 

dispersed. 

El-Mawass [12] et al. designed a hybrid method to detect 

social bot, combining graphs and traditional supervised 

learning classification, where graphs are not constructed 

using user interaction connections on social networks, but the 

application of user usage The program is similar to the graph, 

and then the Markov random field model is constructed on 

this basis. The experiments show that the detection of the 

supervised learning combined with the Markov random field 

model in the paper has improved the accuracy and recall rate 

of the detection only by supervised learning. 

3. Experiment 

3.1. Work process 

Figure 1 shows the specific working steps of the system. 

1. The input is the original text. It is a user-published 

content field in the data set crawled by the crawler, 

which can be mainly divided into two types, one is to 

reply to other people's tweets, and the other is to non-

reply to the tweets. 

2. The second step is preprocessing, the first of which 

is data cleaning. The purpose is to obtain as much as 

possible the complete content posted by the user from 

the original text to prepare for subsequent operations. 

The main cleaning contents are: 

(1) Remove @user, @ is a function 

provided by Twitter for users. For example, if user 

A @user B in a tweet, user B can see reminders and 

tweet jumps in "message", which contributes to the 

entire sentence the value is small and needs to be 

filtered. 

(2) Remove the URL, the original text will 

have a URL at the end. 

(3) Remove non-English language tweets, 

because quantum similarity is only for English, so 

remove non-English sentences. 

(4) Remove irrelevant characters, such as 

"*", "[", "&" and other strange characters. 

(5) According to the manual observation of 

this topic tweet, users often regard # keywords, 

such as #HuaWei, #5G, #HUAWEI, etc. as part of 

the sentence, so there is no need to remove the #tag, 

just remove the "#" character. Can. 

(6) Remove stop words, and filter stop 

words with the help of the functions provided by 

the word segmentation library. Usually, stop words 

have less value to sentences, and filtering can 

improve the efficiency of the program. 

3. Sentiment analysis. Take the sentence output in 2 

as input, call the VADER [13] library for sentiment 

analysis, and output the sentiment of each sentence, 

including positive, negative and neutral sentiments. 

4. Participle clauses. The sentence output in 2 is used 

as input, and the public NLTK [14] library is used to 

segment the sentence according to the part of speech set 

in the library. 

5. Quantum similarity clustering. The purpose of 

similarity clustering is to screen social bot into the 

identification link as much as possible, and eliminate 

normal users, which can greatly enhance the running 

speed of the system and improve the recognition 

accuracy. The outputs of 3 and 4 are used as input, and 

then clustering is performed to obtain several group 

categories, calculate the texts one by one for similar 

grouping. There are the following two clustering results. 

(1) The number of sentences in the group is 

less than one-tenth of the number of users. This 

value is not large, and it only represents the 

opinions of a small number of people, so it is 

eliminated. 

(2) The number of sentences in the group is 

one tenth higher than the number of users, reserved. 

6. Feature extraction. Find the user corresponding to 

the sentence from the output in 5, and extract the 

selected features as the output. 



 

40 

7. Machine learning identification, take the output in 

6 as input, and use the logistic regression algorithm in 

the sklearn library to identify social bot. 

 

Figure 1. Detection process 

3.2. Quantum Similarity [3] 

In recent years, quantum disciplines have gradually 

developed. Among them, quantum theory has a good 

mathematical modeling framework, which is combined in the 

fields of economic theory, psychology and information 

science, and there are some attempts in natural language in 

artificial intelligence. The quantum similarity algorithm is 

also known as the sentence similarity calculation method 

based on quantum theory. It uses the properties of quantum 

theory in terms of superposition, interference, entanglement 

and incompatibility to establish a sentence quantum state 

modeling model, and then introduces the concept of fidelity. 

To judge whether two quantum states are close, that is, to 

judge whether two sentences are similar. The main process is 

to decompose the two sentences into state functions, and then 

use the density matrix to represent the meaning of a single 

complete sentence, and then further obtain the fidelity 

calculation expansion between them, and then represent the 

word as a semantic word vector (Word Embedding) Do the 

calculation to get the similarity value of the two sentences. 

A natural language sentence is composed of words with 

different parts of speech, such as nouns, verbs, prepositions 

and adjectives, etc. This paper decomposes sentences by parts 

of speech and forms a set }...,{ i21 =W , where i  

is a set of words with different parts of speech,

}...,{ 21 ji  = , where j  represents different words of 

the same part of speech, for i , where the word vectors are 

superimposed 
j

j|  , a Hilbert space χ whose basis 

vectors are orthogonal to each other is obtained, as shown in 

Table 1.  

Table 1. Part of speech classification of natural language sentences 

Part of 

Speech 
n v adj conj adj n 

Sentence 

1 
I like sunny and honest boys 

Sentence 

2 
I like handsome and cheerful boys 

The Hilbert space of the adjective in sentence 1 is Consists 

of sunny|   and honest|  superimposed. The whole 

sentence can be expressed as "I like sunny boys" and "I like 

honest boys". 

Here, the part-of-speech set W is represented as a state 

function i +++= ...| 21 , and a weight is assigned 

to each Hilbert space (here represented by a complex value 

ii

i ep


=  , the value is optional), the weight can be 

understood It is assigned for each part of speech to contribute 

differently to the sentence as a whole, then the mathematical 

expression of the function is as follows 

  == i

i

i
i

ie
mm

  |
1

|p
1

|
i

i

     (1) 

In the formula, 
m

1
  is the normalization coefficient, 

which ensures that the normalization 1| =  ,Where 

n
1

  in = 
j

j

n
 |1

  is also a normalization 

coefficient, It can be expressed as 

2

)|)|(|
2

|)|(|
2

(

|

433211

)1(

4

3

2

1

+++++

=

andehonestsunny
e

likeeboysI
e i

i
i

i







 

2

)|)cheerful|handsome(|
2

|)|(|
2

(

|

433211

)2(

4

3

2

1

+++++

=

ande
e

likeeboysI
e i

i
i

i







 
The density matrix is obtained by the outer product of the 

state vectors, expressed as 

.||
1

|
1

|
1

|| *

,

*
jij

ji

i

j

jji

i

is pp
m

p
m

p
m

 ===   (2) 

And then get their fidelity calculation formula 

|,|)1(|| )1()1(2)1()1()1(

iiip  ==
 

|,|)2(|| )2()2(2)2()2()2(

jjjp  ==      (3) 

Then their fidelity is calculated as 
2)2()1()2(2)1(2)2()1( |||)( === jiji pprTF   (4) 

The similarity between them can be obtained by the inner 

product 
 )2()1( | ji 

 of the two right losses. But this is a 

complex value, but if it is multiplied by its own complex 

conjugate, A real value can be obtained, which is the 

definition of fidelity. Represent each word in it as a semantic 

word vector, then calculate the cosine similarity of the two 

words, accumulate and average through the loop iteration, 

you can get the similarity value of the two sentences, and then 

perform nonlinear transformation to determine whether the 

two sentences are similar, gather sentences judged to be 

similar together to form a group.  

3.3. Figure 

(1) Status number: The total number of tweets, 

replies and retweets. 

(2) Favorite number: The total number of favorite 

tweets. 

(3) Verifier: Verification is mainly a function set 

by Twitter in order to allow some users to be better 

recognized by other users. Only after verification can 

there be a verification label. Social bot cannot pass the 

verification. 

(4) Ratio of followers to following: Followers 

and following represent the number of followers and 



 

41 

followers, respectively. This can also be understood as 

the relationship between users. Normal users will not 

pay attention to social bot, and social bot will pay 

attention to normal users for profit or disguise. 

Therefore, social bot usually has a higher fpf value than 

normal users. 

(5) Retweet number: The number of tweets 

retweeted by other users in practice. Usually, tweets 

published by normal users are rarely retweeted by other 

users, so the ret value of social bot will be higher than 

that of normal users. 

(6) The original amount: The original proportion 

of user tweets. Normal users use Twitter to express their 

views on life dynamics or events, while social bot will 

retweet other people’s tweets for this purpose, so usually 

The Originalp value for normal users will be higher. 

(7) Quantity: In the selected event or topic, for 

the purpose of interest, usually social bot will tweet and 

retweet more than normal users on the event topic, so 

usually the quantity value of social bot will be higher 

than that of normal users. higher. 

(8) Registration Date: Calculation formula: the 

total number of days from the registration time to the 

present time/the number of days in a year (the 

experimental value is 366). Usually, the registration 

time of social bot will be related to the topic of the event, 

such as registration after the event occurs. So usually the 

regtime value of social bot will be lower than that of 

normal users. 

4. Experimental results 

4.1. DataSet 

The social platform selected in this paper is Twitter. The 

data set is based on the API interface provided by the platform 

and crawled by crawlers. The selected topic is "Huawei + 5G", 

with a total of 36,677 pieces of data. The selected fields are 

the original post link, content and key. Words, post time, 

author, author link, likes, platform, replies, retweets, and 

author location. The user information data fields are the 

number of pictures and texts, the number of likes, the number 

of fans, the number of followers, the number of original 

tweets, the total number of tweets, whether to authenticate, 

and the registration time. We select two types of social bot 

groups for experimental comparison in our experiments: (a) 

the exact type of tweet replication (b) the type of tweets with 

the same semantics and roughly similar types of tweet texts. 

Both social bot groups were manually labeled in the dataset. 

4.2. Comparison of experimental results 

 

Figure 2. Test results of type a social bot team 

 

Figure 3. Test results of type b social bot team 

The experiments in this paper found that similarity 

clustering filtering is very important in identifying social bot 

or eliminating the amount of data, which can reduce a lot of 

data. Figure 2 is the detection result of the a-type social bot 

team. The accuracy histogram of the three algorithms shows 

that the accuracy of quantum similarity is about 97%, which 

is 2% higher than the other two traditional similarity 

algorithms. Figure 3 b-type the detection results of the social 

bot team show that the accuracy rate of quantum similarity is 

about 85%, which is 3% and 5% higher than the other two 

algorithms, respectively. Experiments show that the method 

we designed can accurately identify social bot, and the cross-

disciplinary combination the quantum similarity algorithm 

has better performance than the traditional similarity 

algorithm in judging the similarity of social bot texts. 

5. Conclusion 

We briefly introduce the background and meaning of social 

bot recognition, and then gives an overview of the current 

related work, and designs a detection method for social bot. 

First, the user's published text is used as input to eliminate 

data and then use sentiment analysis. Classify it and then use 

the quantum similarity algorithm to cluster and filter out the 

clusters, and finally get the user characteristics corresponding 

to the text and use the artificial intelligence algorithm for 

detection. This method combines the quantum similarity 

algorithm of quantum science and the detection of social bot 

in computer science. An interdisciplinary combination was 

carried out, two types of social bot groups were manually 

marked from the data set, and two classical similarity 

algorithms were selected for comparative experiments. 

References 

[1] Shi, W., et al. (2020). "Social bot' Sentiment Engagement in 
Health Emergencies: A Topic-Based Analysis of the COVID-
19 Pandemic Discussions on Twitter." Int J Environ Res Public 
Health 17(22). 

[2] Hagen, L., et al. (2020). "Rise of the Machines? Examining the 
Influence of Social bot on a Political Discussion Network." 
Social Science Computer Review 40(2): 264-287. 

[3] Bingqing, M. and M. Lei (2021). "Study on sentence similarity 
based on quantum theory." Journal of East China Normal 
University (Natural Science) 2021(1): 60. 

[4] Rodríguez-Ruiz, J., et al. (2020). "A one-class classification 
approach for bot detection on Twitter." Computers & Security 
91. 

[5] Wu, B., et al. (2020). "Using Improved Conditional Generative 
Adversarial Networks to Detect Social bot on Twitter." IEEE 
Access 8: 36664-36680. 

94%
95%
96%
97%
98%

quantum_sim cosine_sim jaccard

Accuracy

76%
78%
80%
82%
84%
86%

quantum_sim cosine_sim jaccard

Accuracy



 

42 

[6] Feng, Y., et al. (2020). "Towards Learning-Based, Content-
Agnostic Detection of Social Bot Traffic." IEEE Transactions 
on Dependable and Secure Computing: 1-1. 

[7] Wu, Y., et al. (2021). "A novel framework for detecting social 
bot with deep neural networks and active learning." 
Knowledge-Based Systems 211. 

[8] Kudugunta, S. and E. Ferrara (2018). "Deep neural networks 
for bot detection." Information Sciences 467: 312-322. 

[9] Lingam, G., et al. (2019). "Adaptive deep Q-learning model for 
detecting social bot and influential users in online social 
networks." Applied Intelligence 49(11): 3947-3964. 

[10] Gao, P., et al. (2015). "Sybilframe: A defense-in-depth 
framework for structure-based sybil detection." arXiv preprint 
arXiv:1503.02985. 

[11] Zhang, X., et al. (2018). Sybil Detection in Social-Activity 
Networks: Modeling, Algorithms and Evaluations. 2018 IEEE 
26th International Conference on Network Protocols (ICNP): 
44-54. 

[12] El-Mawass, N., et al. (2018). Supervised Classification of 
Social Spammers using a Similarity-based Markov Random 
Field Approach. Proceedings of the 5th Multidisciplinary 
International Social Networks Conference on - MISNC '18: 1-
8. 

[13] Hutto, C.J. & Gilbert, E.E. (2014). VADER: A Parsimonious 
Rule-based Model for Sentiment Analysis of Social Media 
Text. Eighth International Conference on Weblogs and Social 
Media (ICWSM-14). Ann Arbor, MI, June 2014. 

[14] Bird, Steven, Edward Loper and Ewan Klein (2009).Natural 
Language Processing with Python. O'Reilly Media Inc.