









































Pa
ge

 
1



Pa
ge

 
68

American Journal of  Smart 
Technology and Solutions (AJSTS)

Public Perceptions of  Telegram’s Association with Illicit Activities: A Sentiment Analysis 
Using VADER and Machine Learning 

Rabel Catayoc1*

Volume 4 Issue 1, Year 2025
ISSN: 2837-0295 (Online)

DOI: https://doi.org/10.54536/ajsts.v4i1.4667
https://journals.e-palli.com/home/index.php/ajsts

Article Information ABSTRACT

Received: March 02, 2025

Accepted: April 05, 2025

Published: May 07, 2025

Digital communication platforms have significantly transformed social interaction and 
information dissemination, yet simultaneously present challenges related to illicit activities, 
security threats, and regulatory oversight. Telegram, a widely-used encrypted messaging 
service, has recently drawn global scrutiny due to allegations linking it to criminal enterprises, 
including identity theft, illicit drug markets, and distribution of  child exploitation materials. 
This study systematically evaluates public sentiment surrounding Telegram’s reported 
facilitation of  illegal activities, employing comparative sentiment analysis methodologies: 
VADER (Valence Aware Dictionary and Sentiment Reasoner) and a supervised machine 
learning approach (TF-IDF vectorization coupled with Logistic Regression). A corpus of  
632 reader comments from a Wall Street Journal article discussing Telegram’s controversial 
associations was analysed. VADER-based labelling identified an unexpectedly predominant 
positive sentiment (59.5%), indicating potential public scepticism toward negative media 
narratives or ideological support for encrypted platforms. The logistic regression classifier 
demonstrated robust predictive performance, with overall accuracy of  89.56%, precision 
of  91%, recall of  90%, and an F1-score of  89%, yet displayed a notable positivity bias, 
misclassifying nuanced negative commentary. Qualitative word cloud visualizations further 
highlighted distinctive lexical patterns, underscoring explicit concerns around security and 
criminality in negative comments and humour or reflective discourse in positive remarks. 
Methodologically, results expose critical limitations of  traditional lexical approaches in 
capturing subtle, implicit, or context-dependent negativity, suggesting the integration 
of  advanced context-aware modelling techniques, such as transformer-based neural 
embeddings, for enhanced precision. Practically, this analysis provides critical insights 
for platform governance, risk management strategies, regulatory frameworks, journalistic 
practices, and computational linguistics research, emphasizing the necessity for balanced 
methodological approaches to accurately gauge and respond to nuanced public sentiment 
within contentious digital discourse contexts.

Keywords
Entiment Reasoner, Logistic 
Regression, Machine Learning, 
Sentiment Analysis, Vader, Tf-Idf

1 Mindanao State University - Iligan Institute of  Technology, Philippines
* Corresponding author’s e-mail: rabelcatayoc@gmail.com

INTRODUCTION 
Over the past decade, digital communication 
platforms have substantially transformed interpersonal 
communication, information dissemination, and 
commercial interactions (Van Dijck et al., 2018; Castells, 
2013). Among the rapidly expanding range of  messaging 
services, Telegram—founded in 2013—has emerged as a 
particularly influential platform, attracting approximately 
950 million active monthly users as of  July 2024 (Team, 
2025). Telegram’s appeal stems largely from its user-
friendly interface, robust end-to-end encryption, and 
publicly professed commitment to user privacy, rendering 
it a versatile medium for both personal and professional 
exchanges (Gillespie, 2018; Baumgartner et al., 2020).
Despite its legitimate utility, Telegram has increasingly 
faced scrutiny for allegedly facilitating illicit activities 
(Kohlmann, 2024; Sexton, 2024). Emerging evidence 
indicates that the platform has become instrumental 
in enabling cybercriminal behavior, including identity 

theft, illicit drug distribution, circulation of  child sexual 
exploitation material, and the orchestration of  extremist 
activities (Europol, 2022; The Wall Street Journal [WSJ], 
2024). Telegram’s minimal content moderation policies, 
coupled with its robust encryption and the resultant 
anonymity, have been identified as pivotal factors that 
attract malicious actors to the platform (Gillespie, 2018; 
Europol, 2022).
The arrest of  Telegram’s CEO, Pavel Durov, by French 
authorities in August 2024 marked a significant turning 
point, drawing global media attention to the platform’s 
alleged role in supporting criminal enterprises (WSJ, 
2024). This high-profile incident intensified discussions 
concerning digital platform accountability, focusing on 
the ethical and regulatory responsibilities of  technology 
providers in moderating user-generated content to 
prevent the proliferation of  illegal activities (Gorwa, 
2019; Suzor, 2019).



Pa
ge

 
69

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

In this evolving context, understanding public sentiment 
toward Telegram is critically important, given its influence 
on user trust, regulatory responses, and corporate 
reputation (Pfeffer et al., 2014; Liu, 2015). Sentiment 
analysis—a computational methodology employing 
natural language processing techniques—provides 
an effective tool for systematically assessing public 
perceptions and societal concerns as expressed in textual 
form (Cambria, Das, Bandyopadhyay, & Feraco, 2017). 
By analyzing reader-generated comments on news articles 
addressing Telegram’s purported association with criminal 
activities, scholars can derive nuanced insights into 
public attitudes and discursive trends, thereby informing 
business strategy and regulatory policy formulation 
(Stieglitz & Dang-Xuan, 2013; Mostafa, 2013).
This study systematically applies sentiment analysis 
methodologies—specifically, the Valence Aware 
Dictionary and sEntiment Reasoner (VADER) and 
a supervised machine learning approach (TF-IDF 
vectorization with Logistic Regression)—to reader 
comments from a recent Wall Street Journal article 
examining Telegram’s contentious role in illicit activities. 
The research aims to elucidate prevailing public sentiment, 
offering critical implications for Telegram’s reputation 
management, policy strategies, regulatory considerations, 
and broader sociotechnical discourse.

LITERATURE REVIEW
As digital platforms have gained substantial user bases 
worldwide, there has been growing attention to the 
way users interact, interpret, and respond to media 
representations of  digital companies and their societal 
impacts. Particularly, understanding public perceptions 
of  criminal activities linked to technology platforms has 
critical implications for business practices, regulatory 
scrutiny, and corporate reputation (Pfeffer, Zorbach, 
& Carley, 2014). Sentiment analysis has emerged as a 
valuable analytical tool for systematically evaluating public 
perceptions and interpreting their potential impacts on 
business decisions and policymaking (Liu, 2012). This 
literature review explores prior research employing 
sentiment analysis methods relevant to online public 
discourse about technology platforms and illegal activities. 
It specifically focuses on the business implications for 
Telegram, a widely used messaging and social media app 
identified recently as a primary marketplace for criminal 
transactions (Wall Street Journal, 2024).

Sentiment Analysis: Methods and Applications
Sentiment analysis involves computational methodologies 
designed to systematically identify and extract subjective 
information from textual datasets, evaluating public 
attitudes toward particular topics, products, or services 
(Pang & Lee, 2008; Liu, 2012). VADER (Valence Aware 
Dictionary and sEntiment Reasoner), the sentiment 
analysis model utilized in this research, has proven robust 
in evaluating short-form social media texts due to its 
lexicon-based approach, which assigns positive, negative, 

neutral, and compound sentiment scores (Hutto & Gilbert, 
2014). Scholarly studies utilizing sentiment analysis have 
underscored its reliability and accuracy in understanding 
public sentiment toward social media phenomena. 
For instance, Stieglitz and Dang-Xuan (2013) applied 
sentiment analysis to Twitter data, effectively capturing 
public emotions during political events. Furthermore, 
studies by Mostafa (2013) demonstrated sentiment 
analysis utility in extracting consumer sentiment on social 
media toward brands, products, and corporate practices, 
showing its significance for business implications.

Classification Metrics in Sentiment Analysis
Classification models are evaluated using specific metrics 
designed to capture their performance from various 
perspectives, particularly when applied to sentiment 
analysis tasks.

Precision
Precision measures the proportion of  correctly predicted 
positive observations out of  all observations predicted as 
positive, indicating the model’s reliability in its positive 
predictions (Sokolova & Lapalme, 2009). High precision 
implies minimal false positive classifications, which is 
crucial in sentiment analysis to avoid misrepresenting 
users’ sentiment (Liu, 2015). Models with high precision 
effectively minimize type I errors (incorrect positive 
labels), indicating careful and trustworthy identification 
of  the target sentiment (Aggarwal & Zhai, 2012).

Recall
Recall quantifies the model’s capacity to correctly detect 
the actual positive instances out of  all instances that 
genuinely belong to that category (Han, Kamber, & Pei, 
2011). A model with high recall effectively avoids missing 
significant instances of  the target class, which is critical 
for sentiment analysis where overlooking important 
user sentiment could lead to inaccurate conclusions and 
missed insights (Cambria et al., 2013). Models optimized 
for recall are sensitive to the subtleties of  sentiment, 
capturing the majority of  relevant sentiment-bearing 
comments.

F1-score
The F1-score provides a balanced assessment by 
combining both precision and recall into a single metric 
(Powers, 2011). Particularly useful in sentiment analysis 
contexts, the F1-score addresses the limitations of  relying 
solely on either precision or recall, offering a nuanced 
measure of  overall model performance (Cambria et al., 
2017). Given that precision and recall may individually 
vary, the F1-score provides a holistic view crucial for 
balanced evaluation, especially when class distributions 
are uneven or when balancing false positives and false 
negatives is equally important (Forman & Scholz, 2010).

Support
Support refers to the total number of  actual occurrences 



Pa
ge

 
70

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

of  each class within the dataset, indicating class 
distribution and serving as context for interpreting 
performance metrics (Pedregosa et al., 2011). 
Understanding class support is essential, especially in 
real-world sentiment analysis applications, as imbalanced 
distributions can significantly influence performance 
metrics and interpretations (Weiss & Provost, 2003). 
Proper acknowledgment of  support helps in evaluating 
if  performance metrics are consistent across different 
sentiment classes or if  disparities exist due to class 
imbalance (He & Garcia, 2009).

Accuracy and Averages (Macro and Weighted)
While accuracy offers an intuitive assessment of  
model performance—representing the ratio of  correct 
predictions to total predictions—it can be misleading 
when class distributions are imbalanced (Sokolova & 
Lapalme, 2009). Thus, macro and weighted averages are 
additionally recommended as comprehensive measures 
accounting for class distribution. Macro averaging treats 
each class equally, emphasizing balanced class-level 
performance, while weighted averaging factors in class 
support, providing metrics that reflect actual dataset 
distributions (Pedregosa et al., 2011).

Telegram: A New Platform for Digital Criminal 
Activity
Telegram has rapidly become popular, with around one 
billion users, for its simplicity, functionality, and its stance 
toward user privacy and data confidentiality (Wall Street 
Journal, 2024). However, its minimal moderation policies 
and encrypted communications have inadvertently created 
an environment conducive to illicit activities, including 
identity theft, child exploitation, weapons smuggling, 
and drug trafficking (Kohlmann, 2024; Sexton, 2024). 
Research has shown that platforms combining ease of  
access, encryption, and light moderation can attract 
illicit users, negatively affecting corporate reputation, 
attracting regulatory attention, and causing market-value 
losses (Gillespie, 2018). The Wall Street Journal (2024) 
specifically highlighted Telegram’s transformation into 
a favored platform among criminal entities, triggering 
public concerns and potential regulatory consequences. 
The public exposure of  Telegram’s unintended uses 
can significantly affect its corporate image, impacting 
customer trust and potentially undermining its market 
positioning (Gillespie, 2018; Kohlmann, 2024).

Implications of  Sentiment Analysis for Telegram’s 
Business
Sentiment analysis provides a methodological framework 
through which businesses like Telegram can systematically 
evaluate the public’s reaction to their portrayal in media 
and the ensuing narrative around illicit activities. Negative 
sentiment toward companies can harm brand reputation, 
deter investors, and invite strict regulatory scrutiny 
(Mostafa, 2013; Pfeffer et al., 2014).

Companies increasingly utilize sentiment analysis to 
monitor their public reputation in real-time, enabling 
timely interventions to mitigate negative narratives (Hutto 
& Gilbert, 2014). Given the substantial reputational 
risks evident in Telegram’s recent association with 
illegal markets and data breaches, systematic sentiment 
analysis provides valuable insights. Such analytical 
findings can guide Telegram’s strategic management 
and operational decision-making, including moderating 
policy adjustments, user engagement strategies, and 
regulatory compliance approaches (Gillespie, 2018; 
Kohlmann, 2024). Moreover, sentiment analysis also 
allows identification of  key themes in user-generated 
content, enabling Telegram’s management team to 
better understand public concerns, systematically 
address these issues, and communicate effectively with 
stakeholders, thereby enhancing corporate transparency 
and accountability (Stieglitz & Dang-Xuan, 2013).

Regulatory and Ethical Considerations
Telegram’s positioning amid allegations related to 
facilitating criminal activity, such as identity theft and 
child exploitation, presents significant ethical, legal, and 
regulatory challenges. Businesses perceived as tolerating 
or inadequately addressing such activities can face 
substantial fines, reputational damage, and consumer 
attrition (Gillespie, 2018). Regulatory authorities 
worldwide increasingly require technology firms to 
demonstrate proactive measures to counteract illegal 
activities and harmful content (Sexton, 2024). Sentiment 
analysis outcomes highlighting public negativity towards 
Telegram can underscore the urgency for the firm to 
intensify moderation practices, reporting procedures, and 
collaboration with external watchdog organizations such 
as the Internet Watch Foundation (IWF) and the National 
Center for Missing and Exploited Children (NCMEC) 
(Wall Street Journal, 2024; Sexton, 2024).
Ultimately, prior literature clearly underscores the 
significance of  sentiment analysis as a powerful analytical 
method to gauge public perceptions and their business 
implications. The case of  Telegram emphasizes the 
necessity of  deploying robust sentiment analysis tools like 
VADER to analyze public sentiment toward corporate 
practices, particularly concerning critical ethical, legal, 
and reputational issues.
For Telegram, the practical implications derived 
from sentiment analysis can translate into substantial 
business actions, notably improving moderation efforts, 
transparency initiatives, stakeholder communications, 
and regulatory compliance. Ultimately, robust 
sentiment analyses can assist businesses like Telegram 
in understanding, managing, and mitigating significant 
reputational risks arising from associations with 
illegal activities, thus contributing to improved long-
term sustainability, ethical practices, and corporate 
responsibility.



Pa
ge

 
71

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

MATERIALS AND METHODS
Data Collection
The dataset used in this study consists of  reader 
comments extracted from a Wall Street Journal article 
examining Telegram’s role as a prominent platform 
utilized by criminals for illicit activities. The article 
highlights Telegram’s perceived role in facilitating criminal 
activities, including identity theft, drug trafficking, and 
exploitation by paedophile rings. Reader comments were 
systematically gathered from the Wall Street Journal’s 
digital publication platform, forming a corpus suitable for 
sentiment analysis aimed at uncovering public opinion on 
the topic.

Data Preprocessing
To prepare the text data for analysis, a series of  
preprocessing steps were applied. Each comment was:

• Tokenized using NLTK’s word_tokenize function to 
split the text into individual word units.

• Lowercased to ensure consistency in word matching.
• Filtered by removing standard English stopwords 

using NLTK’s stopword list, which reduced noise and 
improved the focus on sentiment-bearing terms.

• Lemmatized using the WordNet Lemmatizer, 
standardizing words to their base or dictionary form (e.g., 
“running” → “run”).
This preprocessing pipeline ensured that the textual 
input was clean, normalized, and ready for effective 
vectorization and modelling.

Stage 1: Sentiment Labelling (VADER Lexicon 
Analysis)
To generate sentiment labels for supervised learning, 
each comment was analysed using the Valence Aware 
Dictionary and Sentiment Reasoner (VADER), a lexicon 
and rule-based sentiment analysis tool optimized for social 
media and short text. VADER produces a compound 
score for each comment, ranging from -1 (most negative) 
to +1 (most positive). Labels were assigned as follows: 

text into numerical feature vectors that reflect both word 
frequency and uniqueness across the dataset.

• Logistic Regression as the classification algorithm, 
chosen for its simplicity, robustness, and strong 
performance in text classification tasks.
The model was trained using an 80/20 train-test split, 
with the TF-IDF vectorizer fit on the training data and 
applied consistently to the test set.

Evaluation Metrics
Model performance was evaluated using standard 
classification metrics:

Table 1: Sentiment Labelling (VADER Lexicon Analysis)
Sentiment Code Score
Positive sentiment 1 compound score ≥ 0
Negative sentiment 0 compound score < 0

These labels were treated as ground truth for training the 
machine learning model.

Stage 2: Model Training (TF-IDF + Logistic 
Regression)
Following sentiment labelling, a supervised learning 
model was trained to classify sentiment directly from the 
comments. The modelling pipeline involved:

• TF-IDF vectorization of  the comments to convert 

Table 2: Evaluation Metrics
Classification Description
Accuracy Proportion of  total correct predictions.
Precision Proportion of  correctly predicted 

positive comments to all predicted 
positives.

Recall Proportion of  correctly predicted 
positives to all actual positives.

F1 Score Harmonic mean of  precision and 
recall.

A confusion matrix was also generated to visualize 
true positives, true negatives, false positives, and false 
negatives, offering a comprehensive view of  classification 
performance.

Visualization Techniques
To complement the numerical evaluation, visualizations 
were created to aid interpretation of  results:

• Bar charts displayed the distribution of  predicted vs. 
actual sentiments.

• Confusion matrix heatmaps illustrated classification 
performance.

• Word clouds were generated for both positive and 
negative classes to highlight the most frequently used 
terms in each sentiment group.

RESULTS AND DISCUSSION
This section presents the results of  the following: text 
preprocessing, sentiment distribution, Model Evaluation: 
TF-IDF + Logistic Regression, Comparison of  Actual 
vs Predicted Sentiment Counts, Evaluation of  Model 
Predictions, and Word Frequency Patterns by Sentiment 
Category.

Text Preprocessing
Prior to model training and sentiment classification, all 
reader comments were preprocessed to improve the 
quality and consistency of  the input data. This process 
included:



Pa
ge

 
72

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

The cleaned and lemmatized text formed the basis for the 
TF-IDF vectorization and subsequent machine learning 
classification. This step ensured that irrelevant syntactic 
noise was minimized, and that semantically meaningful 
patterns were retained for sentiment learning.

Sentiment Distribution
To assess public reaction to a Wall Street Journal article 
investigating Telegram’s increasing role as a platform 
frequently utilized for illicit activities, a sentiment analysis 
was performed on 632 reader comments using a VADER-
based approach. VADER (Valence Aware Dictionary and 
sEntiment Reasoner) is a lexicon- and rule-based sentiment 
analysis tool optimized for analyzing informal, social 
media-style text. It was used to classify each comment 
into either positive sentiment (1) or negative sentiment (0) 
based on the compound polarity score. Table 4 and Figure 
1 show the resulting distribution was as follows:

Table 3: Text Processing
Process Description
Lowercasing All text was converted to lowercase to ensure uniformity in word representation. For example, 

words like “Telegram,” “telegram,” and “TELEGRAM” were normalized to a single lowercase 
token (“telegram”). This standardization is critical for eliminating case-based redundancy and 
ensuring that semantically equivalent tokens are treated identically by downstream algorithms.

Tokenizing Each comment was segmented into individual word units using NLTK’s word_tokenize function. 
Tokenization enables granular analysis by breaking sentences or phrases into smaller, discrete 
components (tokens). For instance, the sentence “Telegram is not secure.” would be split into the 
tokens [“Telegram”, “is”, “not”, “secure”, “.”], allowing for syntactic and lexical processing on a 
word-by-word basis.

Stop word 
removal

Common English stop words—such as “the,” “and,” “is,” and “was”—were removed using 
NLTK’s built-in stopword corpus. These words typically carry low semantic weight and are 
unlikely to contribute meaningfully to sentiment polarity. Their removal streamlines the feature 
space, reduces dimensionality, and allows more important sentiment-bearing terms to dominate 
the analysis.

Lemmatization Using NLTK’s WordNet Lemmatizer, each word token was reduced to its canonical or base 
form (lemma). For example, “running,” “ran,” and “runs” were all normalized to “run.” Unlike 
stemming, lemmatization leverages linguistic context and a vocabulary dictionary to yield 
grammatically correct base forms. This enhances semantic coherence and improves the model’s 
ability to generalize across morphological variations of  the same word.

of  child abuse material. This result may initially appear 
counterintuitive given the negative framing of  the article. 
However, it aligns with prior observations that public 
discourse around privacy-centric platforms often reveals 
polarized sentiment, where user loyalty, ideological leanings, 
or distrust of  traditional media override the framing of  the 
original reporting (Marwick & Lewis, 2017).
The high rate of  positivity suggests that many 
commenters either disputed the framing of  the article 
as sensationalist or one-sided, defended Telegram as a 
platform prioritizing privacy and freedom of  speech, or 
expressed distrust in regulatory narratives or mainstream 
media accounts of  digital platforms.
These findings have several important implications to 
Telegram, policymakers, business analysts and platform 
strategists, and journalists.
The sentiment trends identified among reader 
comments suggest a strong base of  public support, 
at least within the sampled population. This sustained 
positive sentiment provides strategic reinforcement 

Table 4: Sentiment Distribution of  Reader Comments
Sentiments Code Freq Percentage
Positive Sentiment 1 376 59.5%
Negative Sentiment 0 256 40.5%

This distribution indicates that the majority of  reader 
responses to the article carried a positive or at least non-
negative tone, suggesting either skepticism of  the article’s 
framing or support for Telegram as a platform, despite its 
controversial associations. 
The majority of  reader comments expressed a positive 
sentiment, despite the article’s explicit emphasis on 
Telegram’s alleged role in facilitating criminal activities 
such as identity theft, drug sales, and the distribution 

Figure 3: Sentiment Distribution of  Reader Comments



Pa
ge

 
73

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

for Telegram’s positioning as a neutral and privacy-
respecting communication platform. Amid growing 
regulatory scrutiny and negative portrayals in mainstream 
reporting, this public backing may enable Telegram to 
maintain credibility among its core user base and even 
leverage public advocacy to counteract reputational 
risks. Moreover, such positive sentiment trends could 
help Telegram justify policy stances on privacy and 
encryption, emphasizing user rights and autonomy in 
communications.
The public resistance to negative portrayals of  Telegram 
captured by the sentiment analysis underscores substantial 
challenges policymakers may face when aligning policy 
initiatives and regulatory responses with broader public 
perceptions. Specifically, policymakers must carefully 
balance their enforcement priorities, especially concerning 
encrypted platforms, against evident user concerns about 
privacy and autonomy. This divergence implies a necessity 
for policymakers to adopt more nuanced, evidence-based 
approaches to regulation and to engage proactively with 
the public, clearly communicating regulatory objectives 
and their rationale. Sentiment analyses such as this 
provide valuable insights, highlighting areas where public 
understanding or support for regulatory interventions 
may require further cultivation or clarification.
From a strategic and analytical standpoint, the distribution 
of  sentiment emphasizes the significant advantage 
of  incorporating sentiment insights into reputation 

management strategies, communication planning, and 
audience segmentation efforts. Understanding user 
attitudes towards Telegram offers critical business 
intelligence, particularly in determining brand positioning, 
user acquisition, and long-term retention strategies. 
Positive sentiment that endures despite external 
reputational pressures may suggest robust brand loyalty 
and resilience, which platform strategists can harness to 
strengthen user engagement, increase user advocacy, and 
mitigate reputational crises. Such sentiment insights thus 
provide actionable intelligence for managing reputational 
risk and ensuring long-term user relationships.
For journalists and media professionals, the observed 
mismatch between article framing and reader sentiment 
suggests a pressing need for more comprehensive 
engagement with audience perception metrics and 
sentiment feedback loops. Particularly when reporting on 
contested or controversial technologies, journalists may 
benefit from systematically analysing audience sentiment 
and engagement data to better understand the reception 
and impact of  their reporting. Such insights could 
encourage journalists to adopt more nuanced framing, 
more effectively anticipating and addressing potential 
audience skepticism or resistance. Ultimately, leveraging 
sentiment analysis as part of  journalistic practice can 
enhance public trust, reader engagement, and the 
credibility of  media reporting.

Table 5: Sentiment Analysis Implications
Stakeholders Implications
Telegram Public sentiment can strategically reinforce Telegram’s neutral positioning despite 

reputational challenges and regulatory scrutiny.
Policymakers Resistance in public sentiment highlights the complexities policymakers face aligning 

technology regulation with user attitudes toward privacy-focused platforms.
Business Analysts and 
Platform Strategists

Positive audience sentiment underscores the importance of  integrating sentiment metrics 
into strategic brand management and user-retention planning.

Journalists A disconnect between journalistic framing and public response emphasizes the necessity 
for media practitioners to utilize audience perception insights in technology reporting.

This sentiment distribution serves as a foundational result 
in evaluating how machine learning-based sentiment 
analysis can surface important sociotechnical dynamics in 
public discourse surrounding digital platforms.

Model Evaluation: TF-IDF + Logistic Regression
To assess the performance of  the supervised machine 
learning model, a Logistic Regression classifier was 
trained on TF-IDF–transformed features derived from 
the preprocessed comments. The data was randomly 
split into training and testing sets using an 80/20 ratio. 
After training, the model was used to predict sentiment 
labels for the full dataset. These predictions were then 
compared with the original VADER-based sentiment 
labels in Table 4 to evaluate classification accuracy. The 
confusion matrix in Table 6 presents a breakdown of  
correct and incorrect predictions.

Table 6: Confusion Matrix: TF-IDF + Logistic Regression
Classification Outcomes Number of  Observations
True Positives (TP) 369
True Negatives (TN) 189
False Positives (FP) 67
False Negatives (FN) 7

Figure 2 visualizes the confusion matrix summarizing 
the model’s sentiment classification performance on 632 
reader comments. The results indicate that 369 comments 
were correctly classified as positive (true positives), and 
189 were correctly classified as negative (true negatives). 
However, 67 comments that were actually negative were 
misclassified as positive (false positives), while only 7 
positive comments were misclassified as negative (false 
negatives). These distributions demonstrate that the 



Pa
ge

 
74

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

model more frequently errs on the side of  positivity.
The classification report provides additional insight into 
model performance across precision, recall, and F1-score:
Performance metrics further reinforce this observation. 
The model achieved an overall accuracy of  88%, with 
precision scores of  96% (0.96) for negative sentiment and 
85% (0.85) for positive sentiment. This means that of  all 
comments the model predicted as negative, 96% were 
truly negative suggesting that the model rarely mislabels 
positive comments as negative; and of  all comments your 
model predicted as positive, 85% were actually positive 
showing that model occasionally mislabels negative 
comments as positive respectively (Aggarwal & Zhai, 
2012). 

Table 7: Classification Report
Sentiment Precision* Recall** F1-Score Support
Negative (0) 0.96 0.74 0.84 256
Positive (1) 0.85 0.98 0.91 376
Accuracy — — 0.88 632
Macro Avg 0.91 0.86 0.87 632
Weighted Avg 0.89 0.88 0.88 632

* – Out of  all comments that my model labeled as positive (or negative), how many were actually correctly labeled?
**  – Out of  all comments that are actually positive (or negative), how many did my model successfully identify?

Recall scores varied more substantially—while positive 
sentiment achieved a remarkably high recall of  0.98, 
negative sentiment lagged behind at 0.74. This suggests 
that the model successfully captured 98% of  all truly 
positive comments, meaning it missed very few positive 
comments, and the model successfully identified 74% of  
all truly negative comments which indicates that it missed 
some negative comments. 
The F1-scores, which balance precision and recall, were 
0.84 for negative sentiment which reflects good but not 
perfect balance between precision and recall for negative 
comments. There’s some room for improvement, mainly 
due to lower recall; and 0.91 for positive sentiment 
which reflects a strong overall performance for positive 
comments, achieving good balance and high accuracy. 
These findings suggest that while the classifier is generally 
reliable, it is significantly more confident and consistent 
when identifying positive sentiment, potentially due to 
the more explicit or straightforward language used in 
positive comments. In contrast, negative sentiment may 
be expressed with greater subtlety or linguistic complexity, 
leading to higher misclassification rates.

Implications
The model’s asymmetrical performance—high recall 
for positives and lower recall for negatives—suggests 
that user-generated negative comments may carry more 
complex, ambiguous, or implicit linguistic patterns. This 
aligns with prior findings in affective computing literature, 
where negativity is often expressed through irony, sarcasm, 

or culturally specific cues (Cambria et al., 2017). In the 
context of  this study, the model’s superior performance on 
positive sentiment has both methodological and practical 
implications. From a methodological standpoint, the high 
recall for positive sentiment suggests that the TF-IDF + 
Logistic Regression pipeline is effective in capturing the 
lexical patterns and features commonly associated with 
supportive or favorable expressions. However, the lower 
recall for negative sentiment implies that certain critical 
linguistic cues—such as sarcasm, subtle criticism, or 
context-dependent negativity—may not be fully captured 
by surface-level textual features alone. This highlights 
a limitation of  bag-of-words models when applied to 
sentiment detection, especially in domains where opinions 
are nuanced or emotionally complex.
The observed skew toward positive sentiment 
classifications could result in the masking of  important 
critical or negative feedback if  similar sentiment analysis 
models are operationally employed for real-time content 
moderation or in sentiment tracking dashboards. 
Specifically, an overly optimistic bias in automated 
sentiment classification could lead to underestimating 
user dissatisfaction, concerns, or complaints, ultimately 
limiting Telegram’s ability to accurately gauge user 
experience and responsiveness to platform policies. 
Consequently, Telegram may miss vital opportunities for 
improvement or intervention, potentially weakening their 
overall strategy for addressing user sentiment effectively.
From a business and risk management perspective, the 
model’s observed bias towards positivity indicates a 

Figure 2: Confusion Matrix: TF-IDF + Logistic Regression



Pa
ge

 
75

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

significant analytical blind spot. Overestimating positive 
sentiment may lead analysts to underestimate underlying 
reputational or compliance-related risks, hindering early 
detection and management of  potential public relations 
issues, regulatory noncompliance, or controversies that 
negatively impact brand perception. Therefore, analysts 
and strategists must critically account for the model’s 
positive skew and seek supplementary measures or 
qualitative insights to counterbalance and enhance the 
robustness of  risk assessment frameworks.
For regulatory authorities and stakeholders concerned 
with platform governance, the model’s potential 
underrepresentation of  negative sentiment carries 
substantial implications for monitoring and addressing 
harmful or controversial content. If  negativity or critical 
discourse is systematically underreported or inadequately 
represented by the sentiment model, regulators may not 
fully comprehend public concern, discomfort, or backlash 
toward problematic or disputed platform practices. Thus, 
regulators should advocate for analytical transparency 

and accuracy in automated sentiment assessment 
tools, ensuring a reliable basis for policy formulation, 
enforcement priorities, and broader public accountability 
mechanisms. 
These analytical results underline the critical need within 
computational linguistics and natural language processing 
research communities for sentiment classification models 
that demonstrate greater sensitivity and accuracy in 
capturing negative expressions within public discourse. 
The apparent positive bias underscores known linguistic 
challenges, such as subtlety, irony, sarcasm, implicit 
negativity, or complex sentiment cues, highlighting 
ongoing areas for improvement in sentiment modelling. 
Computational linguists are therefore encouraged to 
prioritize developing nuanced, context-aware models 
that better reflect the multifaceted nature of  negative 
sentiment expression, particularly when analyzing 
contentious topics or emotionally charged online 
communication environments.

Table 8: Model Evaluation Implications
Stakeholders Implications
Telegram as a platform This skew may obscure critical feedback if  similar models are used operationally for 

content moderation or sentiment dashboards.
Business analysts and risk 
managers

the model's tendency to overpredict positivity implies a potential blind spot in detecting 
reputational or compliance-related risks.

Regulatory stakeholders an underrepresentation of  negativity could have implications for monitoring harmful 
content or evaluating user sentiment toward controversial content.

Computational linguists these results emphasize the necessity of  building sentiment models sensitive to the 
nuances of  negative expression in public discourse.

Practically, the model’s conservative stance on detecting 
negativity could affect how public sentiment is interpreted 
by stakeholders such as platform regulators, journalists, 
and technology companies. Underestimating negative 
feedback may lead to an overly optimistic assessment of  
user attitudes toward Telegram’s role in illicit activities, 
thereby distorting public discourse or policy priorities. 
Future iterations of  the model may benefit from 
incorporating more context-aware approaches, such as 
transformers or neural embeddings, to more accurately 
capture the semantics of  critical or disapproving content.
Comparison of  Actual vs Predicted Sentiment Counts
To further evaluate the performance of  the trained logistic 
regression classifier, a comparison was made between 
the original sentiment labels (assigned by the VADER 
analyser) and the predicted sentiment labels generated by 
the TF-IDF + Logistic Regression model.
The comparison between actual and predicted sentiment 
labels in Table 9 and Figure 4 demonstrates a clear 
tendency of  the logistic regression classifier to favour 
positive sentiment predictions. Specifically, the model 
predicted fewer negative comments (192 predicted versus 
256 actual) and more positive comments (440 predicted 
versus 376 actual), indicating an observable imbalance 
toward positive sentiment classifications.

This bias aligns closely with the earlier observation 
regarding class-specific recall metrics (as previously 
presented in Table 7), where the model achieved high 
recall for positive sentiment (0.98) but comparatively 
lower recall for negative sentiment (0.74). Such results 
reinforce the notion that the classifier finds positive 
linguistic patterns easier or clearer to detect, potentially 
due to the typically straightforward, explicit lexical 
characteristics associated with positive sentiment.

Table 9: Actual and Predicted Sentiment Distribution
Sentiments Number of  observations
Actual (632)
Negative comments 256
Positive comments 376
Predicted (632)
Negative comments 192
Positive comments 440

Conversely, the systematic underprediction of  negative 
sentiment highlights a methodological limitation of  
employing bag-of-words based TF-IDF vectorization. 
This result suggests that subtlety, irony, sarcasm, and 



Pa
ge

 
76

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

implicit forms of  criticism—which are characteristic of  
negative sentiment—may not be adequately captured by 
purely lexical and frequency-based textual representations 
(Cambria et al., 2017; Liu, 2015). Therefore, future 
improvements may require incorporating more context-
sensitive modelling approaches—such as deep learning 
methods, neural embeddings, or transformer-based 
architectures—that can better interpret linguistic nuances 
and emotional complexity present in critical or negative 
user-generated comments.
In practical terms, the observed bias toward positivity 
could have meaningful implications for stakeholders. If  
used operationally, such a model might systematically 
underestimate user dissatisfaction or critique, thereby 
influencing sentiment dashboards, content moderation 
systems, or strategic decision-making based on public 
feedback.
Thus, from both methodological and applied perspectives, 
the skewed distribution illustrated in this analysis 
emphasizes the importance of  adopting more nuanced 
computational approaches to sentiment analysis tasks, 
particularly within socially or emotionally charged digital 
discourse contexts.

Each bar is labeled with the total number of  comments 
per class (0 = negative, 1 = positive). The model predicted 
more positive comments than were actually labeled, 
indicating a bias toward classifying sentiment as positive.

Implications
The classifier’s observable bias toward predicting 
positive sentiment indicates a methodological 

limitation inherent in TF-IDF–based logistic regression 
approaches. Specifically, the frequent misclassification 
or underrepresentation of  negative sentiment suggests 
that current lexical and frequency-based methods may 
inadequately capture subtle, context-dependent, or 
implicit expressions of  negative opinions. To address 
these challenges, future sentiment modelling efforts 
should incorporate more advanced computational 
linguistic frameworks, such as contextualized embeddings 
or transformer-based approaches, that better account for 
complex linguistic phenomena including sarcasm, irony, 
or implicit negativity (Cambria et al., 2017; Liu, 2015).
The model’s positive prediction skew has practical 
ramifications for real-world sentiment monitoring and 
decision-making scenarios. For platform providers like 
Telegram, relying exclusively on similar predictive models 
for content moderation, user-experience assessment, or 
reputation management could lead to an overly optimistic 
interpretation of  user sentiment. This in turn may obscure 
critical insights into genuine user concerns, dissatisfaction, 
or risks, potentially hindering timely and appropriate 
responses to emerging issues or user grievances.
The underrepresentation of  negative sentiment could 
inhibit effective detection of  critical user feedback, 
limiting the platform’s ability to accurately gauge and 
respond to user sentiment trends. Consequently, potential 
issues might escalate unnoticed, compromising overall 
user satisfaction and retention. An overprediction of  
positive sentiment poses a risk to accurate assessment 
and management of  reputational or compliance-related 
concerns. It may lead to misinformed strategic decisions 
by underestimating user dissatisfaction or overlooking 
emerging controversies and risks.
Regulators relying on similar sentiment analytics tools 
might underestimate the prevalence and intensity of  
negative user perceptions toward controversial issues, 
thus impacting the effectiveness of  their oversight and 
policy interventions. These findings underscore the 
critical need for developing and adopting models that 
better interpret nuanced linguistic cues associated with 
negative sentiment. Researchers are thus encouraged to 
advance context-aware and semantically sophisticated 
analytical techniques that reflect the intricacies of  user-
generated negative commentary.

Figure 3: Sentiment Distribution: Actual vs Predicted

Table 10: Actual and Predicted Sentiment Implications
Stakeholder Implications
Methodological 
Implications

TF-IDF-based models inadequately capture nuanced negativity, indicating a need for 
more context-sensitive methods in sentiment analysis.

Practical Implications Overly positive predictions may obscure critical user feedback, compromising effective 
sentiment monitoring and response.

Telegram (Platform 
Management)

Underestimating negative sentiment could prevent the timely identification of  emerging 
user dissatisfaction or concerns.

Business Analysts & Risk 
Managers

Excessive positivity may result in overlooked reputational and compliance risks, 
impairing strategic decision-making accuracy.



Pa
ge

 
77

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

In conclusion, the observed sentiment prediction 
bias illustrates critical methodological and practical 
limitations. Future efforts in sentiment analysis should 
aim for a balanced representation of  nuanced negative 
expressions, thereby enabling more accurate, insightful, 
and responsive analysis of  public discourse in digital 
communication environments.

Evaluation of  Model Predictions
To further assess the performance of  the logistic 
regression model trained using TF-IDF features, we 
compared its predictions to the original sentiment labels 
produced by the VADER analyzer. Each comment was 
categorized into one of  four interpretation types. 
Table 11 provides a detailed explanation of  classification 
outcomes generated by the sentiment analysis model 
(TF-IDF vectorization with Logistic Regression). It 
categorizes the model’s predictions into four distinct 
outcomes: True Positive (TP), True Negative (TN), False 
Positive (FP), and False Negative (FN). The True Positive 
category, representing comments correctly predicted as 

positive, has the highest frequency with 375 observations, 
indicating strong model performance in identifying 
positive sentiment accurately. The True Negative 
category, referring to correctly identified negative 
comments, includes 191 observations, also showing 
robust performance for negative predictions.
Conversely, the False Positive category—comprising 
65 comments incorrectly classified as positive despite 
being labeled negative—illustrates a tendency of  the 
model toward positivity bias. The False Negative 
outcome was notably low, with only a single comment 
misclassified as negative despite its true positive label, 
reflecting minimal risk of  overlooking genuinely positive 
feedback. Collectively, these outcomes underscore the 
model’s overall high accuracy but also highlight its 
asymmetrical performance, particularly its propensity to 
err on the side of  positive sentiment classification. Such 
detailed classification metrics are valuable for identifying 
specific areas for methodological improvements and for 
understanding the practical implications of  deploying this 
model in real-world sentiment monitoring contexts.

Table 12 presents and visualize in Figure 6 the distribution 
of  the logistic regression model’s prediction outcomes in 
comparison with the actual sentiment labels (generated 
using the VADER analyzer). As illustrated, the largest 
frequency occurred in the True Positive (correctly 
predicted positive) category with 375 comments, 
followed by the True Negative (correctly predicted 
negative) category at 191 comments. The False Positive 
category, representing comments incorrectly identified 
as positive despite their true negative labeling, had a 
notable presence with 65 occurrences. Conversely, the 
False Negative category exhibited minimal representation 

with only a single occurrence, indicating a very low rate 
of  incorrectly identifying positive comments as negative.
Examining specific examples of  misclassification 
provides further insight into the limitations of  the model. 
For instance, the comment “Company’s fault someone 
stupid downloaded personal data,” which was genuinely 
negative (true label: negative), was erroneously classified 
as positive (false positive). Additionally, the extremely 
brief  and contextually ambiguous comment “‘s really” 
was similarly misclassified as positive, despite its original 
negative label.

Regulatory Stakeholders Underrepresentation of  negativity can hinder regulatory insight into public concerns, 
reducing effectiveness in policy and oversight.

Computational Linguists 
& NLP Researchers

Findings emphasize the importance of  developing advanced models sensitive to subtle 
linguistic cues associated with negative expressions.

Table 11: Explanation of  Classification Outcomes for Model Predictions
Prediction outcome Interpretation Frequency

✅ True Positive (TP) Correctly predicted positive comment 375

✅ True Negative (TN) Correctly predicted negative comment 191

❌ False Positive (FP) Incorrectly predicted positive when the true label was negative 65

❌ False Negative (FN) Incorrectly predicted negative when the true label was positive 1

Table 12: Example of  Misclassified Comments
Comment True Label Predicted
“Company’s fault someone stupid downloaded personal data.” 0 (Negative) 1 (False Positive)
“‘s really.” 0 (Negative) 1 (False Positive)



Pa
ge

 
78

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

These misclassifications exemplify common challenges 
faced by lexically driven sentiment analysis models, 
particularly in interpreting the nuanced context, implicit 
negativity, or the brevity frequently encountered in user-
generated comments. Such errors underscore the need 

for more sophisticated, context-aware models capable 
of  accurately capturing subtle or implicit expressions 
of  negative sentiment commonly occurring in digital 
discourse.

These cases highlight challenges in interpreting context 
or brevity, common in reader-generated content.
The evaluation of  the logistic regression model trained 
on TF-IDF features, compared against sentiment labels 
originally produced by the VADER analyzer, reveals 
distinct strengths and limitations. As presented in 
Tables 11 and 12, and visualized in Figure 6, the model 
demonstrated strong performance in correctly classifying 
positive sentiment (375 True Positives) and negative 
sentiment (191 True Negatives), reflecting overall high 
predictive accuracy. However, the presence of  65 False 
Positives—negative comments incorrectly classified as 
positive—illustrates a clear positivity bias in the model’s 
classification approach. Notably, the occurrence of  
False Negatives was minimal, with only one instance of  
misclassification. Specific misclassified examples, such 
as the nuanced negative comment, “Company’s fault 
someone stupid downloaded personal data,” and the 
brief, ambiguous expression “‘s really,” further illustrate 
common difficulties encountered by lexically-oriented 
sentiment analysis methods in accurately interpreting 
subtlety, brevity, irony, or implicit negativity commonly 
found in reader-generated comments. These observations 
collectively highlight the practical necessity and 
methodological importance of  adopting more advanced, 
context-sensitive analytical frameworks, particularly in 
sentiment analysis tasks involving nuanced or implicitly 
expressed opinions.

Implications
The observed classification outcomes and 

misclassifications carry several critical methodological and 
practical implications for sentiment analysis applications:
The positivity bias reflected by the model’s higher rate 
of  false positives suggests limitations inherent to lexically 
driven TF-IDF approaches, which fail to adequately 
capture subtle linguistic nuances, contextually embedded 
negativity, or ambiguous user-generated expressions. 
Consequently, sentiment models should integrate more 
contextually sophisticated approaches, including neural 
network-based or transformer models, which are capable 
of  better representing nuanced linguistic features such 
as irony, sarcasm, implicit negativity, and brevity that 
characterize authentic digital discourse. For real-world 
monitoring applications—such as content moderation or 
sentiment dashboards—this positivity bias may result in 
an underestimation of  critical user feedback, potentially 
limiting the accuracy of  strategic decisions made by 
platforms, businesses, or regulatory bodies. The model’s 
conservative approach toward negative sentiment may 
inadvertently lead stakeholders to overlook emerging 
concerns or grievances expressed subtly or implicitly by 
users.
Telegram’s reliance on similarly biased sentiment 
classification models could obscure critical or dissatisfied 
user feedback, thereby negatively impacting the 
platform’s ability to accurately gauge user perceptions, 
intervene effectively in user concerns, or strategically 
respond to emerging user dissatisfaction. Such oversight 
may ultimately compromise user trust, retention, and 
satisfaction. The positivity bias may represent a blind spot 
in identifying reputational or compliance-related risks. 

Figure 4: Interpretation of  Model Predictions vs. Actual Sentiment Labels



Pa
ge

 
79

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

Overlooking negative sentiment due to misclassification 
can hinder the timely detection and response to potential 
issues, controversies, or public relations risks, thus 
potentially damaging the organization’s reputation or 
compliance standing.
If  regulatory authorities utilize such sentiment analysis 
models to monitor online discourse, underrepresentation 
of  negativity can significantly distort the accuracy of  
regulatory assessments. Misclassification may cause 
regulators to underestimate public dissatisfaction 

or concern, undermining effective policymaking, 
intervention, and oversight. The presented data reinforces 
the need to prioritize development of  sentiment analysis 
models that robustly address linguistic complexity 
in negative expressions. Computational linguists are 
thus encouraged to refine methodologies—employing 
advanced neural models, semantic embeddings, and 
transformer architectures—to overcome existing 
limitations and better capture the full spectrum of  
linguistic expression in sentiment analysis tasks. 

Table 13: Evaluation of  Model Predictions Implications
Stakeholder Implications
Methodological The positivity bias underscores the limitations of  lexically driven TF-IDF models, 

necessitating context-sensitive sentiment approaches.
Practical A model's tendency toward positivity may result in the systematic underestimation 

of  critical or negative user feedback.
Telegram (Platform 
Management)

Underestimating negative sentiment could hinder Telegram’s ability to detect and 
address emerging user concerns effectively.

Business Analysts and Risk 
Managers

Positivity bias in sentiment modeling creates analytical blind spots, potentially 
masking reputational or compliance-related risks.

Regulatory Stakeholders Underrepresentation of  negative sentiment can compromise regulatory oversight 
and the accuracy of  policy interventions.

Computational Linguists and 
NLP Researchers

Misclassifications highlight the necessity for developing sentiment models capable 
of  capturing linguistic subtleties and complexities inherent in negative expressions.

Overall, these implications emphasize the necessity 
of  employing sophisticated analytical techniques, 
complementary qualitative insights, and contextual 
linguistic awareness in sentiment classification to 
improve practical accuracy and methodological reliability, 
particularly in sentiment analysis scenarios involving user-
generated digital communication.

Word Frequency Patterns by Sentiment Category
To further explore the thematic content of  reader 
sentiment, two word clouds were generated to visualize the 
most frequently occurring terms in positive and negative 
comments. This qualitative visualization complements the 

quantitative performance metrics by highlighting the types 
of  language that characterize different sentiment classes.
The word cloud for positive comments (Figure X) 
prominently features terms such as “funny,” “agree,” 
“think,” and “good,” suggesting that readers who viewed 
the article more favorably often expressed humor, 
agreement, or general reflection. These comments may 
reflect support for regulatory action or acknowledgment 
of  the app’s broader appeal despite criminal misuse.
These visualizations provide additional context to the 
sentiment classification model by revealing the lexical 
patterns that differentiate user attitudes in response to 
the article.

Figure 7: Commonly used Words for Positive Comments



Pa
ge

 
80

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

In contrast, the word cloud for negative comments 
(Figure Y) is dominated by terms like “criminals,” “data,” 
“identity,” “illegal,” and “scared,” indicating a strong 
concern around security breaches, digital exploitation, 

and Telegram’s alleged facilitation of  illicit activity. The 
presence of  emotionally charged words in this category 
underscores a clear disapproval and unease among users.

The qualitative analysis depicted in Figures 7 and 8 
complements quantitative metrics by revealing distinctive 
lexical patterns in sentiment-classified comments. 
Specifically, the positive comments word cloud (Figure 
7) prominently highlights terms such as “funny,” “agree,” 
“think,” and “good,” signifying reader engagement 
characterized by humor, agreement, reflective discourse, 
and general positivity toward the platform. This suggests 
readers positively disposed toward Telegram may express 
broader support or appreciation, possibly perceiving 
benefits of  regulatory attention or viewing the platform’s 
positive attributes independently of  reported illicit 
activities. In contrast, the negative comments word cloud 
(Figure 8) emphasizes terms such as “criminals,” “data,” 
“identity,” “illegal,” and “scared,” explicitly reflecting 
heightened concerns regarding data security, digital 
threats, and Telegram’s potential role in enabling criminal 
activity. Emotionally charged language in negative 
sentiment comments highlights clear user apprehension 
and explicit disapproval about the platform’s associated 
risks.

Implications
The clear distinction between positive and negative 
lexical patterns indicates the effectiveness of  qualitative 
text visualization techniques, such as word clouds, in 
supplementing quantitative modeling efforts, particularly 
by contextualizing underlying emotional dimensions 
that purely numerical methods might overlook. Lexical 
patterns from negative sentiment illustrate pressing 
concerns about privacy, security, and illicit activity, 
highlighting critical areas where Telegram must proactively 
address user apprehensions to effectively manage public 
perceptions and mitigate reputational risk. Explicitly 
negative terms (“identity,” “illegal,” “scared”) suggest 

heightened reputational or compliance risks that business 
analysts should systematically monitor, reflecting the 
need for rigorous risk management strategies that address 
user security concerns directly. Frequent mentions of  
crime-related terminology among negatively classified 
comments underscore urgent public demand for effective 
regulatory intervention to ensure platform accountability, 
protect users from exploitation, and uphold digital safety 
standards. The stark lexical contrast between positive and 
negative comments reaffirms the necessity for advanced 
computational methodologies sensitive to distinctively 
emotional, domain-specific vocabulary, thus enhancing 
model precision in capturing nuanced sentiment 
expressions. In sum, qualitative visualization of  lexical 
patterns significantly enriches quantitative sentiment 
analyses, providing valuable insights for methodological 
refinement, strategic management, regulatory oversight, 
and computational linguistics advancement.

Model Performance Metrics
To evaluate the accuracy of  the machine learning classifier 
(TF-IDF vectorization + Logistic Regression), standard 
performance metrics were computed using the predicted 
sentiment labels (Analysis) and the ground-truth labels 
(sentiment) generated by the VADER-based approach.
The model achieved a high degree of  overall accuracy, 
correctly predicting the sentiment classification for 
approximately 89.56% of  all reader comments. The 
precision score was 0.9100, indicating that the majority 
of  comments predicted as positive or negative were 
labeled correctly. The recall score was 0.8956, suggesting 
that the model was effective at capturing the majority of  
true sentiment classes. Finally, the F1-score, a harmonic 
mean of  precision and recall, was 0.8922, confirming 
the model’s strong and balanced performance across 

Figure 8: Commonly used Words for Negative Comments



Pa
ge

 
81

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

sentiment categories.
These results reinforce the model’s suitability for 
sentiment classification tasks in real-world online 
discussions, particularly in socially sensitive contexts like 
Telegram’s association with criminal activities.

Table 14 and Figure 9 summarize the key performance 
metrics for the logistic regression classifier using TF-IDF 
features to predict sentiment labels derived from VADER-
generated ground-truth labels. The model demonstrated a 

Table 14: Model Performance Metrics Table
Metric Score
Accuracy 0.8956 (90%)
Precision 0.9100 (91%)
Recall 0.8956 (90%)
F1 Score 0.8922 (89%)

high overall accuracy of  approximately 89.56%, signifying 
its strong capability to correctly classify reader comments 
according to their sentiment. A precision score of  0.9100 
(91%) indicates a high level of  confidence that comments 
classified as either positive or negative by the model were 
indeed labeled correctly. The recall score of  0.8956 (90%) 
underscores the model’s effectiveness in identifying the 
majority of  the sentiment classes accurately within the 
dataset. Furthermore, the F1-score of  0.8922 (89%), 
as a balanced metric combining precision and recall, 
confirms robust and well-balanced overall performance. 
Collectively, these metrics strongly suggest the classifier’s 
appropriateness and reliability for practical sentiment 
classification applications, particularly in sensitive digital 
contexts—such as discussions involving Telegram and 
its alleged association with criminal activities—where 
accurate and nuanced interpretation of  user-generated 
discourse is critical.

Implications of  Model Performance Metrics
The classifier’s strong overall performance (accuracy: 
89.56%, precision: 91%, recall: 90%, and F1-score: 89%) 
confirms the effectiveness of  TF-IDF vectorization 
combined with Logistic Regression for accurately capturing 
lexical patterns indicative of  sentiment. However, 
despite strong metrics, there remains a methodological 
implication that certain nuanced linguistic features—such 
as implicit negativity or context-dependent subtleties—
may still not be fully captured, necessitating continued 
refinement and potential integration of  advanced, 
context-aware modeling approaches.
The high reliability indicated by these metrics suggests 
the model’s suitability for practical deployment in 
sentiment monitoring tools or dashboards, especially 
in contexts involving socially or emotionally sensitive 
topics. Nevertheless, stakeholders must remain aware 
of  potential limitations—particularly related to subtle, 
implicit, or nuanced negative sentiment—that might 
be systematically underrepresented despite overall 
strong performance. The robustness of  these metrics 

indicates that sentiment classification models could 
significantly aid Telegram’s management in real-time 
monitoring of  user sentiment, allowing for proactive 
and informed interventions. Nevertheless, Telegram 
should integrate qualitative or supplementary analyses 
to ensure comprehensive coverage of  user concerns, 
especially those communicated through subtle linguistic 
expressions that may evade quantitative detection. 
High precision and recall rates offer business analysts a 
reliable tool for assessing user sentiment accurately, thus 
improving strategic and operational decision-making 
processes. However, analysts should remain cautious, 
complementing such models with qualitative analysis 
or expert judgment to capture nuanced expressions of  
potential risks and reputational threats.
For regulators, the demonstrated model reliability 
supports informed monitoring and policy-making 
concerning online discourse and platform governance. 
Nonetheless, regulators should acknowledge potential 
blind spots in automated sentiment classification, 
particularly regarding subtle negative commentary that 

Figure 7: Mode Performance Metrics



Pa
ge

 
82

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

Table 14: Model Performance Metrics Table
Stakeholder Implications
Methodological High model accuracy validates TF-IDF and Logistic Regression but highlights the 

necessity for advanced approaches to capture nuanced negative expressions.
Practical Robust overall performance supports operational deployment for sentiment analysis, 

though stakeholders should remain cautious of  subtle sentiment complexities.
Telegram (Platform 
Management)

Reliable sentiment analysis metrics enable effective user monitoring; however, 
supplementary qualitative analyses are needed to detect subtle concerns.

Business Analysts and 
Risk Managers

Strong precision and recall metrics enhance informed decision-making, yet 
complementary analysis remains critical for nuanced risk detection.

Regulatory Stakeholders High classifier reliability aids regulatory oversight but necessitates awareness of  potential 
blind spots in identifying subtly negative user feedback.

Computational Linguists 
and NLP Researchers

While lexical models exhibit strong predictive accuracy, their inherent linguistic 
limitations call for research into context-sensitive computational methodologies.

could inform nuanced policy decisions. The strong 
performance metrics highlight the value of  lexical models 
but also underscore their inherent limitations. Researchers 
are encouraged to further explore integrating advanced 

linguistic methodologies (such as neural embeddings or 
transformer models) to address limitations related to 
capturing subtlety, irony, or contextually complex negative 
expressions in sentiment analysis.

Overall, while this model demonstrates substantial 
accuracy and reliability for real-world sentiment analysis, 
stakeholders must remain cognizant of  inherent limitations 
and pursue complementary analytical techniques and 
approaches for comprehensive interpretation of  public 
sentiment.

Discussion
Text Preprocessing
The text preprocessing methods implemented in this 
study were critical in ensuring accurate sentiment 
classification and robust model performance. The 
preprocessing pipeline, which included lowercasing, 
tokenization, stop word removal, and lemmatization, 
played a significant role in enhancing data consistency and 
semantic interpretability. Converting text to lowercase 
effectively mitigated variability caused by case sensitivity, 
standardizing semantically identical words into unified 
tokens. This step was essential for avoiding redundancy 
and ensuring that terms such as “Telegram,” “telegram,” 
and “TELEGRAM” did not dilute or distort lexical 
patterns recognized by downstream sentiment classifiers.
Tokenization further contributed to the model’s granularity 
and interpretative accuracy by segmenting user-generated 
content into discrete lexical units. By enabling word-
level analyses, tokenization facilitated precise sentiment 
attribution and more accurate feature extraction, allowing 
models to distinctly recognize sentiment cues from 
individual tokens. The removal of  common English stop 
words similarly played a critical methodological role by 
reducing textual noise and dimensionality. Excluding low-
value linguistic elements—such as articles, prepositions, 
or conjunctions—allowed sentiment-bearing words 
to be more prominently weighted in the subsequent 
TF-IDF vectorization stage. This approach aligns well 
with established sentiment analysis literature, which 

emphasizes the advantage of  minimizing unnecessary 
linguistic content that could introduce ambiguity or dilute 
the overall feature relevance (Aggarwal & Zhai, 2012; Liu, 
2015).
Finally, the lemmatization procedure notably strengthened 
the semantic coherence and interpretability of  the text 
data. By systematically converting morphological variants 
into standardized base forms, lemmatization substantially 
enhanced the model’s ability to generalize beyond 
surface-level lexical variation. This facilitated the accurate 
aggregation of  related terms—such as “running,” “ran,” 
and “runs”—thereby enhancing feature representation 
consistency and boosting overall predictive reliability 
(Jurafsky & Martin, 2009). Such linguistic normalization 
is particularly critical in nuanced sentiment analysis 
scenarios, like the Telegram-related discussions explored 
here, where accurate interpretation of  morphological 
variations can meaningfully influence sentiment polarity 
outcomes.
Overall, the text preprocessing methods adopted in this 
study significantly contributed to the classifier’s robust 
performance (accuracy ~89.56%). Future research could 
explore the potential benefits of  integrating more advanced 
preprocessing steps, such as context-aware embeddings 
or deep linguistic models, to better address nuanced 
linguistic structures, implicit sentiment, and subtleties 
inherent in user-generated digital communication. This 
expanded methodological repertoire might further 
enhance sentiment classification accuracy, particularly 
within socially sensitive or linguistically complex online 
discourse contexts.

Sentiment Distribution
The sentiment distribution derived from reader 
comments, as analyzed by the VADER-based sentiment 
classification, provides valuable insights into public 



Pa
ge

 
83

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

perceptions surrounding Telegram, particularly in relation 
to its portrayal in mainstream media. Interestingly, despite 
the negative framing of  Telegram’s association with 
criminal activities such as identity theft, drug trafficking, 
and exploitation, the sentiment analysis revealed a 
predominant positivity among reader comments, with 
approximately 59.5% classified as positive versus 40.5% 
negative. This notable prevalence of  positive sentiment 
may initially seem counterintuitive given the explicitly 
critical portrayal of  Telegram in the article. However, 
this finding resonates closely with existing scholarship 
emphasizing polarized public responses toward digital 
platforms associated with privacy, encryption, and 
user autonomy. Such platforms frequently engender 
divided public opinion, often shaped significantly by 
users’ ideological leanings, platform loyalty, or inherent 
skepticism towards traditional media narratives (Marwick 
& Lewis, 2017).
The evident divergence between media framing 
and audience sentiment underscores a substantial 
sociotechnical dynamic that stakeholders—such as digital 
platforms, policymakers, business strategists, and media 
professionals—must carefully navigate. For Telegram, 
strong positive sentiment indicates a solid foundation of  
user support, potentially insulating the platform against 
reputational damage and regulatory scrutiny, thereby 
reinforcing its public position as a privacy-respecting 
communication medium. Policymakers, however, 
face pronounced challenges, as public resistance to 
negative framing and subsequent policy interventions 
illustrates the complex relationship between regulatory 
enforcement, public perception, and platform autonomy. 
Consequently, policymakers should approach regulation 
with greater sensitivity to user sentiment, perhaps 
engaging in proactive public dialogue to communicate 
clearly the rationale and objectives underlying platform 
governance decisions.
For business analysts and platform strategists, the 
sustained positive user sentiment signals a strategic 
opportunity to deepen user engagement, brand loyalty, 
and resilience to reputational crises. Incorporating 
comprehensive sentiment analysis into reputation 
management frameworks thus provides actionable 
insights, enabling proactive strategic responses aligned 
with genuine user attitudes. Meanwhile, the evident 
disconnect between journalistic framing and reader 
sentiment highlights a critical area for professional 
reflection among journalists and media organizations. 
Specifically, sentiment analysis offers a powerful 
feedback mechanism to gauge the impact of  journalistic 
narratives on public opinion. Integrating sentiment 
analytics into journalistic practice can improve audience 
engagement and enhance credibility by encouraging more 
nuanced, responsive reporting practices, particularly on 
controversial technological issues.
Overall, this sentiment distribution analysis illustrates how 
sentiment modelling techniques—such as the lexicon-
based VADER analyzer—can significantly contribute to 

understanding and interpreting complex public discourse 
around digital platforms. Such analyses not only inform 
methodological refinement in computational sentiment 
studies but also carry direct implications for the practical 
management of  public perception, policy development, 
strategic communication, and journalistic integrity in the 
digital age.

Model Evaluation: TF-IDF + Logistic Regression
The evaluation of  the Logistic Regression classifier using 
TF-IDF-transformed features revealed notable insights 
into the strengths and limitations of  lexical sentiment 
analysis approaches in classifying public sentiment. The 
confusion matrix (Table 6 and Figure 2) demonstrates 
that the model achieved substantial predictive accuracy, 
correctly classifying 369 comments as positive (True 
Positives) and 189 as negative (True Negatives). 
However, it also exhibited an asymmetrical performance 
with a noteworthy positivity bias—misclassifying 67 
negative comments as positive (False Positives), while 
only misclassifying 7 positive comments as negative 
(False Negatives). This asymmetry, while affirming the 
model’s overall reliability, underscores a pronounced 
methodological limitation wherein nuanced expressions 
of  negative sentiment are frequently missed or incorrectly 
interpreted.
The classification report (Table 9) further reinforces 
these findings through precision, recall, and F1-score 
metrics. The model exhibited high precision (96%) in 
negative sentiment classification, suggesting that when it 
did predict negativity, it was highly reliable. Nevertheless, 
the significantly lower recall for negative sentiment 
(74%) compared to positive sentiment recall (98%) 
signals a clear methodological challenge in capturing 
subtler forms of  negativity. Such challenges align closely 
with established literature in sentiment analysis, which 
frequently identifies implicit negativity, irony, sarcasm, 
and contextually embedded critiques as complex linguistic 
constructs often inadequately captured by surface-level 
lexical analysis (Cambria et al., 2017).
From a methodological standpoint, these results highlight 
limitations inherent in bag-of-words approaches such 
as TF-IDF, which rely primarily on explicit lexical 
cues rather than sophisticated semantic interpretation. 
Thus, the model’s robust identification of  positive 
sentiment likely reflects the clearer lexical and syntactic 
patterns commonly associated with explicit agreement 
or approval. In contrast, negative sentiment frequently 
involves linguistic subtlety and complexity, explaining the 
observed positivity bias. Future research directions could 
address this limitation by integrating context-sensitive 
approaches, such as transformer-based language models 
or neural embeddings, which can better interpret subtle 
linguistic signals and implicit semantic content.
Practically, the observed positivity skew has substantial 
implications for stakeholders. For Telegram, deploying 
similar sentiment models operationally—such as in 
content moderation or user-sentiment dashboards—may 



Pa
ge

 
84

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

inadvertently underestimate user dissatisfaction, obscuring 
critical feedback and inhibiting effective responsiveness. 
For business analysts and risk managers, positivity bias 
represents a potential analytical blind spot, potentially 
compromising the timely detection of  reputational and 
compliance risks. Regulatory stakeholders should similarly 
remain cautious, recognizing that a sentiment model biased 
toward positivity could diminish the visibility of  user 
concerns or critical perspectives essential for informed 
policymaking and oversight. Finally, computational 
linguists and NLP researchers are encouraged to address 
these biases and improve sentiment models by developing 
nuanced methods capable of  recognizing linguistically 
complex negative expressions, ultimately ensuring more 
balanced and accurate sentiment assessments in public 
discourse.
In conclusion, this model evaluation indicates strong 
overall reliability for lexical sentiment analysis using TF-
IDF vectorization and Logistic Regression. However, 
the asymmetrical performance pattern emphasizes 
the need for methodological enhancements to better 
capture nuanced negativity. Future sentiment analysis 
models should therefore adopt advanced computational 
methods, ensuring comprehensive and contextually 
accurate interpretation of  public sentiment, particularly 
within emotionally charged or contentious digital 
communication contexts.

Comparison of  Actual vs Predicted Sentiment Counts
The comparison between actual and predicted sentiment 
counts from the logistic regression classifier trained on 
TF-IDF-transformed features provides critical insights 
into both the methodological strengths and limitations 
of  the approach. The classifier exhibited a notable bias 
toward positive sentiment, systematically predicting a 
higher number of  positive comments (440) compared 
to the actual positive count (376), while correspondingly 
underestimating negative sentiment (192 predicted versus 
256 actual). This positivity skew aligns closely with 
earlier classification metrics, particularly the recall scores 
that highlighted significantly higher recall for positive 
sentiment (0.98) compared to negative sentiment (0.74). 
These observations suggest that positive sentiment 
expressions typically feature more explicit, straightforward 
linguistic patterns, which are easier for lexical-based 
TF-IDF models to detect consistently. Conversely, the 
underrepresentation of  negative sentiment indicates a 
substantial methodological shortcoming in capturing 
nuanced linguistic cues, subtle criticism, irony, and 
implicit forms of  negativity.
From a methodological perspective, the identified 
bias underscores significant limitations inherent in 
traditional bag-of-words approaches like TF-IDF, which 
rely predominantly on surface-level lexical features and 
frequency counts. While effective in capturing explicit 
sentiment expressions, these approaches often fail to 
adequately interpret the context-dependent subtleties and 
complexity inherent in negative user-generated comments 

(Cambria et al., 2017; Liu, 2015). Therefore, enhancing 
model accuracy—particularly for negative sentiment—
would require integration of  advanced contextualization 
methods such as neural embeddings, transformer-based 
architectures, or deep learning techniques that offer 
deeper semantic understanding of  linguistic nuances.
Practically, the identified positivity bias bears substantial 
implications for operational sentiment monitoring in 
real-world digital platforms. For Telegram, the systematic 
underestimation of  negative sentiment could obscure 
critical user feedback, grievances, or dissatisfaction, 
potentially compromising timely intervention, user 
satisfaction, and trust. Similarly, business analysts and 
risk managers relying on such sentiment analysis tools 
may underestimate reputational and compliance risks, 
misinforming strategic decisions and hindering proactive 
risk mitigation. Regulatory stakeholders may also be 
adversely affected, as this positivity bias could diminish 
the accuracy and effectiveness of  regulatory assessments, 
particularly around sensitive or contentious platform 
practices. Consequently, these stakeholders must approach 
sentiment analysis results cautiously, supplementing lexical 
models with qualitative or additional analytical methods 
to achieve comprehensive sentiment understanding.
Furthermore, computational linguists and natural 
language processing researchers should interpret these 
findings as a clear mandate for continued innovation. 
This bias highlights the necessity for more nuanced, 
contextually sensitive modeling approaches capable of  
accurately interpreting the implicit, subtle, and often 
culturally specific cues of  negative sentiment expressions. 
Advancing analytical techniques beyond traditional lexical 
frameworks—towards sophisticated, contextually aware 
computational linguistics—will enable more accurate, 
balanced, and insightful sentiment analysis outcomes, 
particularly within emotionally charged digital discourse 
contexts.
In conclusion, while the logistic regression classifier 
demonstrates overall robust predictive reliability, 
its positive sentiment skew identifies a significant 
methodological and practical limitation. Addressing this 
limitation through the adoption of  advanced semantic 
modeling approaches promises more balanced, accurate, 
and practically valuable sentiment analyses for digital 
communication platforms and stakeholders.

Evaluation of  Model Predictions
The detailed evaluation of  the Logistic Regression model 
utilizing TF-IDF features, compared against original 
sentiment labels generated by the VADER analyzer, reveals 
key methodological strengths and critical limitations. The 
analysis categorizes model predictions into four distinct 
outcomes: True Positives (TP), True Negatives (TN), 
False Positives (FP), and False Negatives (FN). Notably, 
the model demonstrated strong overall predictive 
accuracy, correctly identifying positive sentiment in 375 
instances and negative sentiment in 191 instances. This 
robustness in correctly classifying sentiment underscores 



Pa
ge

 
85

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

the general effectiveness of  TF-IDF–based logistic 
regression approaches for explicit sentiment recognition. 
However, the presence of  65 False Positive predictions—
negative comments incorrectly classified as positive—
highlights a significant positivity bias within the model’s 
classification approach. Conversely, the occurrence of  
False Negatives was minimal, with only one such instance 
observed, reflecting minimal risk in overlooking positive 
user-generated feedback.
An examination of  specific misclassified examples 
further illuminates inherent methodological challenges. 
Comments like “Company’s fault someone stupid 
downloaded personal data,” a negatively intended critique 
incorrectly predicted as positive, exemplify the difficulty 
lexically oriented sentiment models encounter in 
accurately interpreting implicit negativity, subtle criticism, 
or nuanced linguistic context. Similarly, brief  and 
ambiguous comments like “‘s really” further emphasize 
these limitations, where brevity and lack of  explicit lexical 
cues impede accurate classification. These instances 
underscore a critical shortcoming in lexically driven 
models such as TF-IDF, demonstrating their inadequate 
handling of  subtle, contextually embedded negativity, 
ambiguity, and implicit expressions prevalent in authentic 
digital communication.
The positivity bias observed has substantial 
methodological implications, calling attention to 
the inherent constraints of  frequency-based lexical 
approaches when interpreting user-generated content. 
Such limitations advocate strongly for the integration of  
more advanced computational methods—particularly 
deep learning models, context-aware neural embeddings, 
or transformer-based architectures—that possess greater 
semantic understanding and can better account for 
linguistic subtleties, irony, brevity, and implicit expressions 
of  negative sentiment.
Practically, the implications of  the model’s positivity 
bias are significant for real-world sentiment monitoring. 
Telegram, as a platform provider, risks systematically 
underestimating genuine user dissatisfaction or critical 
feedback if  similarly biased models are operationalized. 
This could compromise accurate detection and timely 
management of  user concerns, potentially affecting user 
satisfaction and retention adversely. Business analysts and 
risk managers could likewise misinterpret sentiment data 
due to this positivity skew, possibly overlooking subtle 
but critical reputational or compliance-related risks. 
Regulatory stakeholders relying on such models might 
similarly underestimate negative sentiment, reducing their 
effectiveness in policy oversight and intervention. For 
computational linguists and natural language processing 
researchers, these findings reinforce the necessity 
of  developing models capable of  nuanced semantic 
interpretation, motivating research into sophisticated, 
contextually sensitive linguistic methodologies.
Overall, these insights emphasize a clear need to 
complement lexically driven sentiment analyses with 
advanced semantic modeling techniques and qualitative 

interpretation strategies. Ensuring comprehensive and 
balanced detection of  nuanced negative sentiment 
expressions will significantly enhance methodological 
robustness, practical utility, and stakeholder trust in 
computational sentiment analysis tools—particularly 
within sensitive, nuanced digital communication contexts.

Word Frequency Patterns by Sentiment Category
The qualitative exploration of  lexical patterns through 
word clouds (Figures 5 and 6) offers meaningful 
contextual insight, complementing the quantitative 
sentiment classification results. The visualization distinctly 
highlights differences in language usage between positive 
and negative comments, providing deeper understanding 
into user sentiment. Specifically, the positive sentiment 
word cloud (Figure 5) prominently features words such 
as “funny,” “agree,” “think,” and “good,” indicating 
that readers positively disposed toward Telegram 
frequently employed language associated with humor, 
agreement, reflective discourse, and explicit positivity. 
These lexical choices likely signify broader support, 
either for Telegram’s foundational principles (privacy 
and autonomy) or for regulatory approaches towards the 
platform, notwithstanding its controversial associations. 
Such qualitative insights reinforce previous quantitative 
findings of  prevalent positivity, underscoring user loyalty 
or skepticism toward media portrayals of  Telegram’s 
negative associations.
Conversely, the word cloud representing negative 
comments (Figure 6) emphasizes emotionally charged 
and security-focused terms, such as “criminals,” “data,” 
“identity,” “illegal,” and “scared.” The prominence of  
these words underscores explicit user concerns about 
Telegram’s potential role in facilitating illicit activities, 
including data breaches, identity theft, and digital 
exploitation. These lexically explicit negative expressions 
indicate clear apprehension and disapproval, highlighting 
a substantial segment of  public sentiment focused on 
security, privacy violations, and digital threats associated 
with the platform.
Methodologically, these visualizations highlight the 
efficacy of  qualitative text analysis techniques such as 
word clouds, which significantly enrich purely quantitative 
analytical approaches. By revealing the nuanced, 
underlying emotional and contextual dimensions of  
sentiment, these methods provide essential supplementary 
context to quantitative models, thus enhancing the overall 
interpretative accuracy of  sentiment analyses. Practically, 
the stark distinction in lexical patterns emphasizes critical 
areas for platform management—such as Telegram—to 
proactively address user concerns, especially in security, 
privacy, and trust. Failing to address these highlighted 
negative concerns could compromise user satisfaction, 
reputation, and long-term platform sustainability.
From a risk management perspective, explicitly negative 
terminology (“illegal,” “identity,” “scared”) suggests 
significant reputational or compliance risks, which 
analysts and strategists must rigorously monitor. Such 



Pa
ge

 
86

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

qualitative insights reinforce the necessity of  integrated, 
contextually nuanced risk assessment frameworks, 
enabling more precise management responses to public 
concerns or controversies. For regulatory stakeholders, 
the frequency of  crime-associated terms in negative user 
comments underscores urgent demands for accountability 
and effective oversight. Thus, policymakers are advised 
to closely align regulatory strategies with these publicly 
articulated user apprehensions, addressing platform 
accountability and protecting user rights effectively.
For computational linguists and natural language 
processing (NLP) researchers, the pronounced lexical 
distinction between positive and negative comments 
further emphasizes the methodological importance of  
advanced sentiment analysis techniques. Models sensitive 
to emotional vocabulary, domain-specific language, 
and subtle linguistic nuance will enhance precision and 
interpretative accuracy. Researchers should thus prioritize 
developing sophisticated computational methodologies, 
particularly contextually aware semantic models, to 
accurately capture the complexities of  nuanced sentiment 
expressions.
In conclusion, qualitative lexical analyses, as illustrated 
by word clouds, significantly enhance quantitative 
sentiment analyses by illuminating underlying emotional 
and thematic user sentiments. Integrating qualitative 
techniques with advanced quantitative methods provides 
critical insights for methodological refinement, strategic 
reputation management, informed regulatory oversight, 
and ongoing computational linguistic research—
particularly within contentious or socially sensitive digital 
discourse contexts.

Model Performance Metrics
The performance metrics summarized in Table 14 and 
Figure 7 indicate robust effectiveness of  the logistic 
regression classifier using TF-IDF vectorization for 
sentiment analysis tasks. The model achieved an overall 
accuracy of  approximately 89.56%, precision of  91%, 
recall of  90%, and an F1-score of  89%, collectively 
demonstrating its reliable predictive capability for 
accurately categorizing sentiment in reader-generated 
comments. High precision indicates substantial 
confidence in the model’s sentiment predictions, 
reflecting accuracy in distinguishing between positive and 
negative commentary. Similarly, strong recall underscores 
its effectiveness in correctly identifying most sentiment 
classifications present within the dataset. The balanced 
F1-score further confirms the model’s capability to 
effectively integrate both precision and recall into 
consistently reliable classification performance.
Despite these strong metrics, methodological limitations 
inherent in TF-IDF vectorization and logistic regression 
warrant critical examination. The comparatively lower 
recall for negative sentiment (74%, as detailed previously) 
implies that nuanced linguistic features—such as subtle 
criticism, implicit negativity, sarcasm, or contextually 
complex expressions—remain inadequately captured 

by purely lexical approaches. Consequently, the high 
accuracy achieved may still systematically underrepresent 
certain nuanced sentiment types, specifically negative 
or implicit expressions. Thus, future methodological 
refinement should focus on incorporating advanced 
modeling approaches, such as context-sensitive neural 
embeddings, deep learning models, or transformer-based 
architectures, capable of  comprehensively interpreting 
linguistic subtleties and implicit sentiment cues.
From a practical standpoint, the demonstrated high 
reliability of  this classifier supports its applicability 
for operational deployment in sentiment monitoring 
contexts, particularly those involving sensitive issues such 
as digital privacy or criminal associations. For Telegram’s 
platform management, such reliable metrics indicate 
valuable potential for real-time sentiment monitoring, 
informing proactive user-experience interventions and 
platform responsiveness. However, recognizing inherent 
limitations in capturing subtle negativity, Telegram 
should employ supplementary qualitative analyses or 
human oversight mechanisms to ensure a comprehensive 
understanding of  user sentiment nuances.
For business analysts and risk management professionals, 
these robust precision and recall metrics significantly 
enhance decision-making accuracy and strategic 
confidence, enabling more precise monitoring of  
user sentiment trends. Nevertheless, analysts must 
remain vigilant to inherent methodological limitations, 
complementing automated sentiment analyses with 
qualitative evaluations or expert judgment to avoid 
overlooking subtle reputational or compliance-related 
risks. Regulatory stakeholders similarly benefit from 
the model’s demonstrated reliability, enabling informed 
oversight and nuanced policy formulation in response to 
public sentiment. However, regulators should be cautious 
of  potential blind spots related to subtle negative 
sentiment expression, adopting supplementary analysis 
methods to ensure comprehensive understanding of  
public discourse.
For computational linguists and NLP researchers, 
these performance results highlight the efficacy of  
lexical models such as TF-IDF logistic regression, but 
simultaneously underscore their limitations. Addressing 
these methodological shortcomings necessitates 
advancing context-sensitive computational methodologies 
that better interpret subtle linguistic complexities, irony, 
implicit negativity, and nuanced discourse features 
prevalent in authentic digital communication.
Overall, while the logistic regression classifier 
demonstrates substantial practical reliability and predictive 
accuracy, stakeholders must remain mindful of  inherent 
methodological limitations. Integrating complementary 
analytical techniques and developing advanced semantic 
modeling approaches remain critical avenues for 
improving comprehensive and nuanced sentiment 
analysis. Such advancements promise significant benefits 
for methodological rigor, operational decision-making, 
regulatory effectiveness, and computational linguistics 



Pa
ge

 
87

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

innovation, particularly in sensitive or controversial digital 
communication contexts.

CONCLUSIONS
This study systematically evaluated public sentiment 
regarding Telegram’s reported association with criminal 
activities by applying sentiment analysis methods—
specifically VADER and a TF-IDF vectorized Logistic 
Regression classifier—to reader comments sourced 
from a Wall Street Journal article. Results revealed 
notable patterns, both quantitative and qualitative, which 
carry critical methodological, theoretical, and practical 
implications. Quantitatively, VADER sentiment analysis 
classified the majority (59.5%) of  reader comments as 
positive, indicating significant public skepticism towards 
the negative media framing of  Telegram or a broader 
ideological support for privacy-centric platforms. The 
Logistic Regression classifier demonstrated robust 
performance (89.56% accuracy, precision of  91%, 
recall of  90%, and F1-score of  89%), affirming the 
utility of  lexically-based sentiment classification models 
for effectively analyzing public discourse. However, it 
exhibited a clear bias toward positivity, reflected in lower 
recall for negative sentiment (74%) and a significant 
presence of  false-positive classifications. These results 
highlight the critical methodological limitation of  
current lexical-based approaches in accurately capturing 
nuanced negative sentiment expressions, especially 
implicit negativity, subtle criticisms, sarcasm, and brevity. 
Consequently, advanced context-aware methodologies, 
such as transformer-based neural embeddings, are 
recommended for future research to enhance accuracy, 
particularly in detecting nuanced or implicitly negative 
sentiments.
Qualitative analysis, specifically word cloud visualizations, 
distinctly captured lexical patterns characterizing both 
positive and negative sentiment classes. Positive comments 
frequently included terms reflecting humor, agreement, 
or reflective discourse, suggesting supportive or skeptical 
attitudes toward negative reporting on Telegram. 
Conversely, negative comments explicitly conveyed 
heightened user concerns about Telegram’s role in security 
breaches and digital criminal activities, using emotionally 
charged and crime-specific language (e.g., “criminals,” 
“illegal,” “scared”). These qualitative insights not only 
contextualize quantitative sentiment classifications but 
also underscore significant sociotechnical dynamics 
that digital platforms, policymakers, and regulatory 
stakeholders must acknowledge.
Practically, this research carries substantial implications for 
Telegram’s platform management, business analysts, risk 
managers, regulatory authorities, computational linguists, 
and media professionals. The identified positivity bias in 
predictive models emphasizes caution in operationally 
deploying lexical-based sentiment analysis tools, which 
might systematically overlook subtle critical feedback, 
thereby limiting effective user engagement strategies, 
risk detection, and regulatory responses. Integrating 

qualitative analyses and advanced computational linguistic 
techniques alongside traditional lexical methodologies 
will thus enhance the comprehensive interpretation and 
responsiveness to public sentiment.
Ultimately, the nuanced examination provided by this 
study highlights sentiment analysis as an indispensable 
analytical framework, significantly informing the strategic 
management of  corporate reputation, stakeholder 
communication, regulatory policy development, and 
the ethical accountability of  digital platforms. Moving 
forward, continued methodological refinement and a 
balanced incorporation of  qualitative insights will further 
enable sentiment analysis to robustly support nuanced 
decision-making in sociotechnical domains, notably 
digital communication and platform governance.

REFERENCES 
Aggarwal, C. C., & Zhai, C. (2012). Mining text data. 

Springer.
Baumgartner, J., Zannettou, S., Keegan, B., Squire, M., & 

Blackburn, J. (2020). The Pushshift Telegram Dataset. 
Proceedings of  the International AAAI Conference on Web 
and Social Media, 14(1), 840–847.

Cambria, E., Das, D., Bandyopadhyay, S., & Feraco, A. 
(2017). A Practical Guide to Sentiment Analysis. Springer.

Cambria, E., Poria, S., Gelbukh, A., & Thelwall, M. 
(2017). Sentiment analysis is a big suitcase. IEEE 
Intelligent Systems, 32(6), 74–80.

Cambria, E., Schuller, B., Xia, Y., & Havasi, C. (2013). 
New avenues in opinion mining and sentiment 
analysis. IEEE Intelligent Systems, 28(2), 15–21.

Castells, M. (2013). Communication Power. Oxford 
University Press.

Europol. (2022). Internet Organised Crime Threat Assessment 
2022. Europol Public Information.

Forman, G., & Scholz, M. (2010). Apples-to-apples 
in cross-validation studies: Pitfalls in classifier 
performance measurement. ACM SIGKDD 
Explorations Newsletter, 12(1), 49–57.

Freelon, D., Marwick, A., & Kreiss, D. (2020). False 
equivalencies: Online activism from left to right. 
Science, 369(6508), 1197–1201. https://doi.
org/10.1126/science.abb2428

Gillespie, T. (2018). Custodians of  the Internet: Platforms, 
Content Moderation, and the Hidden Decisions that Shape 
Social Media. Yale University Press. 

Gillespie, T. (2018). Custodians of  the Internet: Platforms, 
Content Moderation, and the Hidden Decisions That Shape 
Social Media. Yale University Press.

Gorwa, R. (2019). What is platform governance? 
Information, Communication & Society, 22(6), 854-871.

Han, J., Kamber, M., & Pei, J. (2011). Data mining: Concepts 
and techniques (3rd ed.). Elsevier.

He, H., & Garcia, E. A. (2009). Learning from 
imbalanced data. IEEE Transactions on Knowledge and 
Data Engineering, 21(9), 1263–1284.

Hutto, C. J., & Gilbert, E. (2014). VADER: A parsimonious 
rule-based model for sentiment analysis of  social 



Pa
ge

 
88

https://journals.e-palli.com/home/index.php/ajsts

Am. J. Smart. Technol. Solutions 4(1) 68-88, 2025

media text. Proceedings of  the International AAAI 
Conference on Weblogs and Social Media, 8(1), 216-225. 

Kohlmann, E. (2024). Quoted in “How Telegram 
Became Criminals’ Favorite Marketplace,” The Wall 
Street Journal. 

Kohlmann, E. (2024). Telegram and the evolving digital 
underground. Journal of  Cybersecurity Research, 8(1), 
12–27.

Liu, B. (2012). Sentiment analysis and opinion mining. Morgan 
& Claypool Publishers. 

Liu, B. (2015). Sentiment analysis: Mining opinions, sentiments, 
and emotions. Cambridge University Press.

Liu, B. (2015). Sentiment Analysis: Mining Opinions, Sentiments, 
and Emotions. Cambridge University Press.

Marwick, A., & Lewis, R. (2017). Media Manipulation and 
Disinformation Online. Data & Society Research Institute.
https://datasociety.net/pubs/oh/DataAndSociety_
MediaManipulationAndDisinformationOnline.pdf

Mostafa, M. M. (2013). More than words: Social networks’ 
text mining for consumer brand sentiments. Expert 
Systems with Applications, 40(10), 4241-4251.

Pang, B., & Lee, L. (2008). Opinion mining and sentiment 
analysis. Foundations and Trends in Information Retrieval, 
2(1-2), 1-135. 

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., 
Thirion, B., Grisel, O., ... & Duchesnay, E. (2011). 
Scikit-learn: Machine learning in Python. Journal of  
Machine Learning Research, 12, 2825–2830.

Pfeffer, J., Zorbach, T., & Carley, K. M. (2014). 
Understanding online firestorms: Negative word-of-
mouth dynamics in social media networks. Journal of  
Marketing Communications, 20(1-2), 117-128.

Powers, D. M. W. (2011). Evaluation: from precision, recall 
and F-measure to ROC, informedness, markedness 
and correlation. Journal of  Machine Learning Technologies, 
2(1), 37–63.

Sexton, D. (2024). Quoted in “How Telegram Became 
Criminals’ Favorite Marketplace,” The Wall Street 
Journal. 

Sexton, J. (2024). Encrypted messaging and criminality: 
Challenges and responses. Digital Crime Studies Journal, 
10(2), 45-60.

Sokolova, M., & Lapalme, G. (2009). A systematic analysis 
of  performance measures for classification tasks. 
Information Processing & Management, 45(4), 427–437.

Stieglitz, S., & Dang-Xuan, L. (2013). Emotions and 
information diffusion in social media: Sentiment 
of  microblogs and sharing behavior. Journal of  
Management Information Systems, 29(4), 217-248.

Suzor, N. P. (2019). Lawless: The Secret Rules That Govern 
Our Digital Lives. Cambridge University Press.

Swire, B., Berinsky, A. J., Lewandowsky, S., & Ecker, U. 
K. H. (2021). Processing political misinformation: 
Comprehending the Trump phenomenon. Royal Society 
Open Science, 8(9), 210631. https://doi.org/10.1098/
rsos.210631

Team, B. (2025, March 20). How many people use Telegram? 
55 Telegram stats. Backlinko. https://backlinko.com/
telegram-users

The Wall Street Journal [WSJ]. (2024). Telegram CEO Pavel 
Durov Arrested in France. Retrieved from https://www.
wsj.com/articles/telegram-ceo-arrested-france

Van Dijck, J., Poell, T., & De Waal, M. (2018). The Platform 
Society: Public Values in a Connective World. Oxford 
University Press.

Wall Street Journal (2024). How Telegram Became Criminals’ 
Favorite Marketplace. Retrieved from www.wsj.com.

Weiss, G. M., & Provost, F. (2003). Learning when training 
data are costly: The effect of  class distribution on tree 
induction. Journal of  Artificial Intelligence Research, 19, 
315–354.


