A Technical seminar report
On
IOT Based Underground Mine Environment Monitoring System
Submitted in partial fulfillment of the requirement for the award of degree
Bachelor Of Technology
In
Computer Science and Engineering
By
M.Pavan Simha Reddy
(21EG105D27)
Under the Esteemed guidance of
Dr.V.Rama Krishna
Department Of Computer Science and Engineering
ANURAG UNIVERSITY
Venkatapur (V), Ghatkesar (M), Medchal (D), T.S-500088
Year 2024-25
Table of Contents
Sno Topic
1. About the Research Paper
2. Why This Topic is Important
3. Main Objectives of the Report
4. Understanding from the Paper
5. Explanation of Technical Details
6. Diagrams
7. Conclusions and Future Scope
8. Additional Research Papers
9. References
1. About the Research Paper
Social media sites such as Twitter(now X) have changed the way we share information,
but they have also amplified the dissemination of objectionable content. Firstly,
While meeting this challenge, "VMCA-Trans," a new deep learning model, comes on
board. It is based on a multimodal model, which merges computer vision and natural lan-
guage processing to analyze images and text. VMCA-Trans is aware that one-method so-
lutions have limitations, therefore, it employs sophisticated transformer models and
a new "Offend Tweet" dataset to identify objectionable content such as hate
speech as well as obscenity. Secondly this approach highlights the necessity of im-
proved AI in an attempt to create a safe online environment and respond to the escalat-
ing challenges of digital debates.
2. Why This Topic is Important?
Social media platforms like X (formerly Twitter) have transformed the way people com-
municate and access information. However, they have also amplified the spread of objec-
tionable content, including hate speech, misinformation, and obscenity. As digital spaces
continue to shape public discourse, ensuring a safe online environment becomes a press-
ing challenge. Traditional moderation methods often struggle with the sheer volume and
complexity of content, making automated solutions essential.
To address this issue, VMCA-Trans, a new deep learning model, introduces an advanced
approach to content moderation. Unlike conventional models that rely on either text or
image analysis alone, VMCA-Trans adopts a multimodal strategy by integrating com-
puter vision and natural language processing. This allows it to analyze images and textual
content simultaneously, improving the accuracy of detecting harmful material. Addition-
ally, the model utilizes the "Offend Tweet" dataset, which enhances its ability to identify
and filter objectionable content.
The significance of this topic lies in the growing need for improved AI moderation to
tackle the challenges of digital interactions. In recent years, the rise of misinformation,
cyberbullying, and online extremism has highlighted the limitations of existing modera-
tion techniques. VMCA-Trans represents a step forward in addressing these challenges,
offering a scalable, intelligent solution that can adapt to evolving online threats.
By leveraging cutting-edge transformer models, VMCA-Trans not only enhances content
moderation but also contributes to maintaining the balance between free speech and re-
sponsible online discourse. As social media continues to evolve, AI-driven solutions like
VMCA-Trans will play a crucial role in ensuring a safer and more inclusive digital envi-
ronment for users worldwide.
3. Main Objectives of the Report
• Examine the Impact of Social Media on Information Sharing
Analyze how platforms like X (formerly Twitter) have transformed communica-
tion while also amplifying objectionable content.
• Highlight the Challenges of Online Content Moderation
Discuss the difficulties of managing large volumes of digital content, including
hate speech, misinformation, and obscenity.
• Introduce VMCA-Trans as an Innovative Solution
Present the deep learning model VMCA-Trans and explain how it enhances con-
tent moderation using advanced AI techniques.
• Explain the Multimodal Approach
Describe how VMCA-Trans integrates computer vision and natural language pro-
cessing to analyze both images and text for better content detection.
• Discuss the Role of the “Offend Tweet" Dataset
Explore how the dataset improves the model’s ability to detect objectionable con-
tent effectively.
• Evaluate the Importance of AI in Digital Safety
Emphasize the necessity of AI-driven moderation to create a safer and more re-
sponsible online environment.
• Address Ethical and Practical Considerations
Examine the balance between free speech and content moderation, ensuring AI
does not lead to excessive censorship.
• Assess Future Implications of AI in Content Moderation
Predict how advancements in AI-driven moderation can shape the future of digital
communication and platform policies.
4. Understanding from the Paper
The VMCA-Trans model identifies offensive content on Twitter through a multi-step
method. First, it collects tweets that include images and text and enriches the text context
with OCR and image captions. Next, it uses improved ViT and BERT models to pull out
features from the text and images and convert them into numerical values. The central
idea, "Value Mixed Cross Attention," unites these features so that the model can under-
stand how visual and textual elements are connected. This is done through query, key,
and value matrices to identify attention scores, which mark important portions of both
types of information.
VMCA-Trans is a deep-learning framework designed to detect offensive content in social
media posts using computer vision and natural language processing. An ablation study
determined that six transformer layers are optimal for detecting binary classes like "Hate
Speech" and "Targeted Insult." Fewer layers lead to under-fitting, lowering accuracy,
while additional layers cause overfitting, reducing performance by approximately 1-2%.
Comparative studies on the Offend Tweet dataset demonstrate that VMCA-Trans sur-
passes state-of-the-art multi-modal classifiers, including ITIN, DMAF, TVMF, and Hate-
CLIPper, across categories such as "Derogatory Commentary" and "Hate Speech." Al-
though ISSUES achieves slightly better recall and AUC scores, VMCA-Trans maintains
higher accuracy and precision in most evaluations. For "Profane Language" detection,
Hate-CLIPper outperforms VMCA-Trans by a small margin. However, in "Targeted In-
sult" detection, VMCA-Trans surpasses existing classification networks in accuracy, pre-
cision, recall, and F1-score, with only a slight AUC advantage held by Hate-CLIPper.
These results highlight the effectiveness of VMCA-Trans in enhancing classification ac-
curacy, reinforcing its role in ensuring safer online spaces. Despite its success, future im-
provements are needed, such as refining classification models, enabling real-time analy-
sis, and developing real-world applications for detecting offensive content in social media
feeds. The study underscores the necessity of continuous advancements in AI-driven
moderation to address the evolving landscape of online discourse and digital safety.
5. Explanation of Technical Details
• Text-Based Analysis (2000s): Early detection relied on NLP, keyword matching,
sentiment analysis, and basic ML but struggled with context, sarcasm, and en-
coded messages.
• Rule-Based Systems (2000s): Utilized dictionaries and predefined rules but
lacked adaptability to evolving language, leading to high false positives.
• Machine Learning Era (2010s): Models like SVM, Naive Bayes, and Logistic
Regression improved detection of overt content but lacked deeper contextual un-
derstanding.
• Deep Learning Revolution (2010s): CNNs, RNNs, and LSTMs enhanced pattern
recognition and context processing, supported by embeddings like Word2Vec and
GloVe.
• Multimodal Analysis (2010s): Combined computer vision and NLP to detect
harmful content, advancing hate speech research.
• Transformer Models (2020s): Advanced architectures like BERT and RoBERTa
improved contextual understanding, handling nuances and long-range dependen-
cies.
• VMCA-Trans (2020s): Introduced specialized models and datasets for real-time
detection of offensive content, pushing the boundaries of multimodal content
moderation.
Fig 1.1 : Evolution of Technology
Fig 1.2: A Twitter post containing text and an im-
age undergoes processing to enhance semantic
context. OCR and image captions are extracted
using Tesseract and ViT-GPT2. The VMCA-Trans
model overview is provided in Fig. 2. Experimen-
tal analysis evaluates Offend Tweet’s compo-
nents separately, testing image and text modali-
ties with standard vision and text classifiers.
6. Diagrams
7. Conclusions and Future Scope
• Superior Accuracy: Multimodal analysis improves the detection of hidden offen-
sive content.
• Contextual Understanding: Attention mechanisms interpret image-text relation-
ships for better content classification.
• Automation: Streamlines the content moderation process, reducing manual ef-
fort.
• Customization: The Offend Tweet dataset enables fine-tuning for specific use
cases.
• Data Bias Challenges: Datasets may reflect societal biases, impacting model
fairness.
• Computational Cost: High resource requirements make large-scale deployment
challenging.
• Interpretability Issues: The "black box" nature of deep learning models compli-
cates understanding and debugging.
• Evolving Content: Requires continuous updates to remain effective against new
forms of offensive content.
• Contextual Limitations: Struggles with sarcasm, coded language, and nuanced
expressions.
• Ethical Concerns: Potential risks of misuse in surveillance or censorship applica-
tions.
8. Additional Research Papers
1. Arnisha Akhter, Uzzal Kumar Acharjee, Md Alamin Talukder, Md Manowarul Is-
lam, and Md Ashraf Uddin. 2023. A robust hybrid machine learning model for
Bengali cyberbullying detection in social media. Natural Language Processing
Journal 4 (2023), 100027.
2. Mohammed Ali Al-Garadi, Kasturi Dewi Varathan, and Sri Devi Ravana. 2016.
Cybercrime detection in online communications: The experimental case of cyber-
bullying detection in the Twitter network. Computers in Human Behavior 63
(2016), 433–443.
3. Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra,
C Lawrence Zitnick, and Devi Parikh. 2015. VQA: Visual question answering. In
Proceedings of the IEEE International Conference on Computer Vision, 2425–
2433.
4. Abayomi Bello, Sin-Chun Ng, and Man-Fai Leung. 2023. A BERT framework to
sentiment analysis of tweets. Sensors 23, 1 (2023), 506.
5. Vedadri Yoganand Bharadwaj, Vasamsetti Likhitha, Vootnoori Vardhini, Adari
Uma Sree Asritha, Saurabh Dhyani, and M Lakshmi Kanth. 2023. Automated Cy-
berbullying Activity Detection using Machine Learning Algorithm. In E3S Web
of Conferences, Vol. 430. EDP Sciences, 01039.
9. References
• Feiran Huang, Xiaoming Zhang, Zhonghua Zhao, Jie Xu, and Zhoujun Li. 2019.
Image–text sentiment analysis via deep multimodal attentive fusion. Knowledge-
Based Systems 167 (2019), 26–37.
• Emad Kebriaei, Ali Homayouni, Roghayeh Faraji, Armita Razavi, Azadeh Shak-
ery, Heshaam Faili, and Yadollah Yaghoobzadeh. 2023. Persian offensive lan-
guage detection. Machine Learning (2023), 1–21.
• Douwe Kiela, Suvrat Bhooshan, Hamed Firooz, Ethan Perez, and Davide Testug-
gine. 2019. Supervised multimodal bitransformers for classifying images and text.
arXiv preprint arXiv:1909.02950 (2019).
• Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet
Singh, Pratik Ringshia, and Davide Testuggine. 2020. The hateful memes chal-
lenge: Detecting hate speech in multimodal memes. Advances in Neural Informa-
tion Processing Systems 33 (2020), 2611–2624.
• Ankur Kumar. 2022. The Illustrated Image Captioning using transformers.
ankur3107.github.io (2022).
