Digital Commons Network^™

Image Captioning Using Reinforcement Learning, Venkat Teja Golamaru

Master's Projects

Image captioning is a crucial technology with numerous applications, including enhancing accessibility for the visually impaired, developing automated image indexing and retrieval systems, and enriching social media experiences. However, accurately describing the content of an image in natural language remains a challenge, particularly in low-resource settings where data and computational power are limited. The most advanced image captioning architectures currently use encoder-decoder structures that incorporate a sequential recurrent prediction model. This study adopts a typical Convolutional Neural Network (CNN) encoder Recurrent Neural Network (RNN) decoder structure for image captioning, but it has framed the problem as a sequential decision-making task. …

Application Of Knowledge Graph Techniques On Textbooks, Yutong Yao

Master's Projects

Textbooks are written and organized in a way that facilitates learning and understanding. Sections like glossary terms at the end of a textbook provide guidance on the topic of interest. However, it takes manual effort to create the index terms in the glossary that highlight the key referenced terminologies and related terms. Knowledge graphs, which have been used to represent and even reason over data and knowledge, can potentially capture textbook’s important terms, concepts, and their relations. Popular since the initial introduction by Google Knowledge Graphs (KGs), they combine graph and data to capture and model enormous amounts of relational …

Context Aware Neural Machine Translation Using Graph Encoders, Saurabh Kale

Master's Projects

Machine translation presents its root in the domain of textual processing that focuses on the usage of computer software for the purpose of translation of sentences. Neural machine translation follows the same idea and integrates machine learning with the help of neural networks.Various techniques are being explored by researchers and are famously used by Google Translate, Bing Microsoft Translator, Deep Translator, etc. However, these neural machine translation techniques do not incorporate the context of the sentences and are only determined by the phrasesor sentence structure. This report explores the neural machine translation technique dedicated to context-aware translations. It also provides …

Comparative Analysis Of Transformer-Based Models For Text-To-Speech Normalization, Pankti Dholakia

Master's Projects

Text-to-Speech (TTS) normalization is an essential component of natural language processing (NLP) that plays a crucial role in the production of natural-sounding synthesized speech. However, there are limitations to the TTS normalization procedure. Lengthy input sequences and variations in spoken language can present difficulties. The motivation behind this research is to address the challenges associated with TTS normalization by evaluating and comparing the performance of various models. The aim is to determine their effectiveness in handling language variations. The models include LSTM-GRU, Transformer, GCN-Transformer, GCNN-Transformer, Reformer, and a BERT language model that has been pre-trained. The research evaluates the performance …

Untraining Gender Bias: An Eye-Tracking Study, Ripujit S. Bamrah

Master's Projects

In recent years, social cognitive theory has emphasized the role of cognitive processes in shaping perceptions and behavior related to gender bias. By examining the impact of targeted training interventions, this study seeks to better understand the influence of such processes on decision-making in the context of character selection. This human-computer interaction study explores the potential of intervention-based training to untraining gender bias in character selection. With an increasing need to address gender bias in various domains, understanding the impact of gender-based training becomes crucial. According to our hypothesis, exposure to masculine characters would boost people’s preference for female- intellectualized …

Document-Level Machine Translation With Hierarchical Attention, Yu-Tang Shen

Master's Projects

Machine translation (MT) aims to translate texts with minimal human involvement, and the utilization of machine learning methods is pivotal to its success. Sentence-level and paragraph-level translations were well-explored in the past decade, such as the Transformer and its variations, but less research was done on the document level. From reading a piece of news in a different language to trying to understand foreign research, document-level translation can be helpful.

This project utilizes a hierarchical attention (HAN) mechanism to abstract context information making document-level translation possible. It further utilizes the Big Bird attention mask in the hope of reducing memory …

Ml-Based User Authentication Through Mouse Dynamics, Sai Kiran Davuluri

Master's Projects

Increasing reliance on digital services and the limitations of traditional authentication methods have necessitated the development of more advanced and secure user authentication methods. For user authentication and intrusion detection, mouse dynamics, a form of behavioral biometrics, offers a promising and non-invasive method. This paper presents a comprehensive study on ML-Based User Authentication Through Mouse Dynamics.

This project proposes a novel framework integrating sophisticated techniques such as embeddings extraction using Transformer models with cutting-edge machine learning algorithms such as Recurrent Neural Networks (RNN). The project aims to accurately identify users based on their distinct mouse behavior and detect unauthorized access …

Steganographic Capacity Of Selected Machine Learning And Deep Learning Models, Lei Zhang

Master's Projects

As machine learning and deep learning models become ubiquitous, it is inevitable that there will be attempts to exploit such models in various attack scenarios. For example, in a steganographic based attack, information would be hidden in a learning model, which might then be used to gain unauthorized access to a computer, or for other malicious purposes. In this research, we determine the steganographic capacity of various classic machine learning and deep learning models. Specifically, we determine the number of low-order bits of the trained parameters of a given model that can be altered without significantly affecting the performance of …

Automated Short-Answer Grading And Misconception Detection Using Large Language Models, Nazmul H. Kazi

UNF Graduate Theses and Dissertations

As education technology continues to evolve, the domains of Automatic Short-Answer Grading (ASAG) and Automated Misconception Detection (AMD) stand at the forefront of innovative approaches to educational assessment. We explore the transformative potential of Large Language Models (LLMs) in revolutionizing these critical areas. Leveraging the remarkable capabilities of LLMs in semantic inference, contextual understanding, and transfer learning, we embark on a comprehensive journey to enhance both ASAG and AMD. On ASAG, we illuminate the efficacy of transfer learning by fine-tuning RoBERTa Large, a state-of-the-art LLM, on task-related corpora, e.g. the Multi-Genre Natural Language Inference (MNLI) corpus. The model's adaptability across …

Practical Ai Value Alignment Using Stories, Md Sultan Al Nahian

Theses and Dissertations--Computer Science

As more machine learning agents interact with humans, it is increasingly a prospect that an agent trained to perform a task optimally - using only a measure of task performance as feedback--can violate societal norms for acceptable behavior or cause harm. Consequently, it becomes necessary to prioritize task performance and ensure that AI actions do not have detrimental effects. Value alignment is a property of intelligent agents, wherein they solely pursue goals and activities that are non-harmful and beneficial to humans. Current approaches to value alignment largely depend on imitation learning or learning from demonstration methods. However, the dynamic nature …

Murder On The Vr Express: Studying The Impact Of Thought Experiments At A Distance In Virtual Reality, Andrew Kissel, Krzysztof J. Rechowicz, John B. Shull

Philosophy Faculty Publications

Hypothetical thought experiments allow researchers to gain insights into widespread moral intuitions and provide opportunities for individuals to explore their moral commitments. Previous thought experiment studies in virtual reality (VR) required participants to come to an on-site laboratory, which possibly restricted the study population, introduced an observer effect, and made internal reflection on the participants’ part more difficult. These shortcomings are particularly crucial today, as results from such studies are increasingly impacting the development of artificial intelligence systems, self-driving cars, and other technologies. This paper explores the viability of deploying thought experiments in commercially available in-home VR headsets. We conducted …

Spam Comments Detection In Youtube Videos, Priyusha Kotta

Master's Projects

This paper suggests an innovative way for finding spam or ham comments on the video- sharing website YouTube. Comments that are contextually irrelevant for a particular video or have a commercial motive constitute as spam. In the past few years, with the advent of advertisements spreading to new arenas such as the social media has created a lucrative platform for many. Today, it is being widely used by everyone. But this innovation comes with its own impediments. We can see how malicious users have taken over these platforms with the aid of automated bots that can deploy a well-coordinated spam …

Tiny Language Models Enriched With Multimodal Knowledge From Multiplex Networks, Clayton Fields, Osama Natouf, Andrew Mcmains, Catherine Henry, Casey Kennington

Computer Science Faculty Publications and Presentations

Large transformer language models trained exclusively on massive quantities of text are now the standard in NLP. In addition to the impractical amounts of data used to train them, they require enormous computational resources for training. Furthermore, they lack the rich array of sensory information available to humans, who can learn language with much less exposure to language. In this study, performed for submission in the BabyLM challenge, we show that we can improve a small transformer model’s data efficiency by enriching its embeddings by swapping the learned word embeddings from a tiny transformer model with vectors extracted from a …

A Structured Narrative Prompt For Prompting Narratives From Large Language Models: Sentiment Assessment Of Chatgpt-Generated Narratives And Real Tweets, Christopher J. Lynch, Erik J. Jensen, Virginia Zamponi, Kevin O'Brien, Erika Frydenlund, Ross Gore

VMASC Publications

Large language models (LLMs) excel in providing natural language responses that sound authoritative, reflect knowledge of the context area, and can present from a range of varied perspectives. Agent-based models and simulations consist of simulated agents that interact within a simulated environment to explore societal, social, and ethical, among other, problems. Simulated agents generate large volumes of data and discerning useful and relevant content is an onerous task. LLMs can help in communicating agents' perspectives on key life events by providing natural language narratives. However, these narratives should be factual, transparent, and reproducible. Therefore, we present a structured narrative prompt …

Object Detection And Image Categorization By Transferring Commonsense Knowledge With Premises And Quantifiers, Irina Chernyavsky

Theses, Dissertations and Culminating Projects

Domestic, or household robots, are autonomous robots designed to make our home-life easier by performing chores and mundane tasks such as cleaning, or cooking. Currently domestic robots are specialized to complete a specific task and, therefore, are confined by factors such as mobility, size, and complexity. With the fast development of computer vision and robotics, the need for more compact, advanced and multi-task robots has emerged. Therefore, the robot needs to be multi-functional, able to discern the environment and the tasks. The aim of this paper is to categorize images in domestic robots as relevant to the culinary, laundry, vacuum …

Structured Dialogue State Management For Task-Oriented Dialogue Systems, Anh Duong Trinh

Doctoral

Human-machine conversational agents have developed at a rapid pace in recent years, bolstered through the application of advanced technologies such as deep learning. Today, dialogue systems are useful in assisting users in various activities, especially task-oriented dialogue systems in specific dialogue domains. However, they continue to be limited in many ways. Arguably the biggest challenge lies in the complexity of natural language and interpersonal communication, and the lack of human context and knowledge available to these systems. This leads to the question of whether dialogue systems, and in particular task-oriented dialogue systems, can be enhanced to leverage various language properties. …

Towards An Improved Understanding Of The Concept Of Style And Its Implications For Textual Style Transfer, Somayeh Jafaritazehjani

Doctoral

The concept of linguistic style denotes that many aspects of text can vary while maintaining a same source core semantic meaning. For example, a message may be written in a formal or informal style. The textual style transfer problem aims at generating a paraphrase of a given text by modifying its style while preserving its content. To the best of our knowledge, within the literature on textual style transfer, there is no standard widely accepted definition of the concept of style. Moreover, very few works have investigated the characteristics of language styles. Therefore, previous research, as far as our knowledge …

Finer Details Of Language Modeling: Text Segmentation, Working Within Resource Limits, And Watermarking, Evan Gordon Lucas

Dissertations, Master's Theses and Master's Reports

Language modeling is a vast sub-field of natural language processing and this work focuses on solving some specific problems within that field. Technically, the work falls into a number of sub-categories within natural language processing; how to segment texts, improving sparse transformer performance for summarization tasks, character level models for dialect determination, watermarking of large language models, and a general method of incorporating minimal human feedback for continual or online learning. Despite touching on many small areas, they all connect as being related to the very general problem of handling sequential data. Language and text can be thought of as …

Anti-American Stance In Turkey: A Twitter Case Study, Gowri Prathap, Alex Korb, Luke Palmieri, Ekrem Kaya, Saltuk Karahan, Hamdi Kavak

School of Cybersecurity Faculty Publications

The availability of social media and biased actors exacerbated Anti-American and Anti-Western views to extremes. In this paper, we report our efforts in analyzing anti-American views on Twitter. We have collected over three years of Turkish tweets related to the US, translated them into English, and analyzed these tweets using various computational social science tools. We found that Turkish tweets related to the US are significantly negative, and emotions reflect disgust and anger. Furthermore, we found that the source of the negative views stems from political actors like Trump or Biden rather than general hatred. Our results shed light on …

Fake News Detection Using Natural Language Processing, Fabiolla Mayrink Costa, Rael Guimaraes

ICT

Nowadays with the advance of technologies we have vast access to any sort of information. We are able to use our phone/computer to access the news of any part of the world. It is great to keep us informed about everything that is happening around the world. It is also a powerful tool used for companies while making strategic business decisions. The biggest issue is that technology can and is being used to manipulate people/companies by propagating fake news. Fake news can mislead people's perceptions while forming opinions on a determined subject. It can also have a big impact on …

Data-Driven Strategies For Disease Management In Patients Admitted For Heart Failure, Ankita Agarwal

Browse all Theses and Dissertations

Heart failure is a syndrome which effects a patient’s quality of life adversely. It can be caused by different underlying conditions or abnormalities and involves both cardiovascular and non-cardiovascular comorbidities. Heart failure cannot be cured but a patient’s quality of life can be improved by effective treatment through medicines and surgery, and lifestyle management. As effective treatment of heart failure incurs cost for the patients and resource allocation for the hospitals, predicting length of stay of these patients during each hospitalization becomes important. Heart failure can be classified into two types: left sided heart failure and right sided heart failure. …

Hate Speech Detection In Hindi, Pranjali Prakash Bansod

Master's Projects

Social media is a great place to share one’s thoughts and to express oneself. Very often the same social media platforms become a means for spewing hatred.The large amount of data being shared on these platforms make it difficult to moderate the content shared by users. In a diverse country like India hate is present on social media in all regional languages, making it even more difficult to detect hate because of a lack of enough data to train deep/ machine learning models to make them understand regional languages.This work is our attempt at tackling hate speech in Hindi. We …

Personalized Tweet Recommendation Using Users’ Image Preferences, Shashwat Avinash Kadam

Master's Projects

In the era of information explosion, the vast amount of data on social media platforms can overwhelm users. Not only does this information explosion contain irrelevant content, but also intentionally fabricated articles and images. As a result, personalized recommendation systems have become increasingly important to help users navigate and make sense of this data. We propose a novel technique to use users’ image preferences to recommend tweets. We extract vital information by analyzing images liked by users and use it to recommend tweets from Twitter. As many images online have no descriptive metadata associated with them, in this framework, we …

Federated Learning For Protecting Medical Data Privacy, Abhishek Reddy Punreddy

Master's Projects

Deep learning is one of the most advanced machine learning techniques, and its prominence has increased in recent years. Language processing, predictions in medical research and pattern recognition are few of the numerous fields in which it is widely utilized. Numerous modern medical applications benefit greatly from the implementation of machine learning (ML) models and the disruptive innovations in the entire modern health care system. It is extensively used for constructing accurate and robust statistical models from large volumes of medical data collected from a variety of sources in contemporary healthcare systems [1]. Due to privacy concerns that restrict access …

Keystroke Dynamics And User Identification, Atharva Sharma

Master's Projects

We consider the potential of keystroke dynamics for user identification and authentication. We work with a fixed-text dataset, and focus on clustering users based on the difficulty of distinguishing their typing characteristics. After obtaining a confusion matrix, we cluster users into different levels of classification difficulty based on their typing patterns. Our goal is to create meaningful clusters that enable us to apply appropriate authentication methods to specific user clusters, resulting in an optimized balance between security and efficiency. We use a novel feature engineering method that generates image-like features from keystrokes and employ multiclass Convolutional Neural Networks (CNNs) to …

Twitter Bot Detection Using Nlp And Graph Classification, Warada Jayant Kulkarni

Master's Projects

Social media platforms are one of the primary resources for information as it is easily accessible, low in cost, and provides a high rate of information spread. Online social media (OSM) have become the main source of news information around the world, but because of the distributed nature of the web, it has increased the risk of fake news spread. Fake news is misleading information that is published as real news. Therefore, identifying fake news and flagging them as such, as well as detecting sources that generate them is an ongoing task for researchers and OSM companies. Bots are artificial …

Article Feed Recommendation For Thai Social Network Application Using Article Context Based On Deep Learning, Pannawit Athipatcharawat

Chulalongkorn University Theses and Dissertations (Chula ETD)

In recent years, social media applications have exhibited significant growth in the number of users. Enhancing content alignment with user preferences is essential in developing a sophisticated recommendation system. Neural networks have become pivotal in improving the performance of recommendation systems. However, the utilization of auxiliary information and text data remains markedly underexplored, especially in the context of Thai recommendation systems. Therefore, this study aims to bridge this gap by developing a recommendation system tailored to Thai social media applications. We focus on leveraging supplementary information and analyzing text features to gain deeper insights into user preferences while addressing the …

การเรียนรู้เชิงลึกสำหรับเข้ารหัสบัญชีการจำแนกโรคระหว่างประเทศ (ไอซีดี) จากบันทึกเวชระเบียนโดยใช้คลังข้อมูลภาษาไทยและภาษาอังกฤษ, ณัฐชา ซาซุม

Chulalongkorn University Theses and Dissertations (Chula ETD)

งานวิจัยนี้เสนอแบบจำลองสำหรับจำแนกรหัสไอซีดีแบบหลายฉลากเพื่อเป็นตัวช่วยจำแนกรหัสไอซีดี การให้รหัสไอซีดีไม่ครบส่งผลให้โรงพยาบาลไม่ได้รับเงินค่าชดเชยอย่างเหมาะสม เราจึงมุ่งเน้นที่จะช่วยโรงพยาบาลให้รหัสไอซีดีอย่างครบถ้วนในขั้นตอนการเบิกจ่ายเงินค่ารักษาพยาบาลซึ่งผลลัพธ์ที่ได้คือการสนับสนุนทางการเงินแก่โรงพยาบาล โดยปกติแล้วปัญหาการให้รหัสไอซีดีแบบหลายฉลากเป็นชุดข้อมูลบันทึกเวชระเบียนที่มีการกระจายตัวแบบหางยาวซึ่งไม่ควรละทิ้งข้อมูลใด ๆ ดังนั้น การจำแนกแบบหลายฉลากของเราจึงเป็นการรวมกันของวิธีการรวมแบบจำลองการเรียนรู้เชิงลึกทั้งสาม (Bidirectional Long-short term memory, Convolutional neural network and Transformer encoder) ด้วยการเลือกค่าทำนายสูงสุดและแบบจำลองทางสถิติคือ Multinomial Naïve Bayes ร่วมกับวิธี Binary Relevance ในส่วนของแบบจำลองการเรียนรู้เชิงลึกจะรับผิดชอบกลุ่มรหัสไอซีดีทั่วไป (โรคที่พบบ่อย) ในขณะที่แบบจำลองทางสถิติจะจัดการกับกลุ่มรหัสไอซีดีพบยาก (โรคที่ผู้ป่วยไม่ค่อยเป็น) ผลการทำนายรหัสไอซีดี-10 จากบันทึกการรักษาด้วยแบบจำลองนี้ให้ค่า Jaccard index ที่ 0.792 สำหรับกลุ่มรหัสไอซีดีทั่วไปและที่ 0.205 สำหรับกลุ่มรหัสไอซีดีพบยาก ส่วนของผลการทำนายรหัสไอซีดี-9 จากบันทึกการให้ยา (หัตถการรักษา) ให้ค่า Jaccard index ที่ 0.963 สำหรับกลุ่มรหัสไอซีดีทั่วไปและที่ 0.201 สำหรับกลุ่มรหัสไอซีดีพบยาก ซึ่งแบบจำลองนี้มีความสามารถในการแนะนำรหัสไอซีดีที่ขั้นตอนการเบิกจ่ายเงินค่ารักษาพยาบาล

การเปรียบเทียบขั้นตอนวิธีการเรียนรู้ของเครื่องและโครงข่ายประสาทเทียมสำหรับการตรวจคะแนนเรียงความภาษาไทยอัตโนมัติ, สุทธิชาย สุริยสัตย์

Chulalongkorn University Theses and Dissertations (Chula ETD)

ตัวแทนนักเรียนไทยได้คะแนนการประเมินด้านการอ่านค่อนข้างตํ่า ในการสอบประเมินสมรรถนะนักเรียนมาตรฐานสากล (Programme for International Student Assessment PISA) หลายงานวิจัยพบว่าทักษะด้านการอ่านสามารถถูกพัฒนาผ่านการเขียนเช่น การเขียนเรียงความได้ แต่การตรวจเรียงความนั้นใช้เวลาค่อนข้างนาน ระบบการตรวจคะแนนเรียงความโดยอัตโนมัติได้ถูกพัฒนาขึ้นเพื่อช่วยเหลือผู้ตรวจเรียงความ อีกทั้งยังสามารถแสดงผลการทำนายคะแนนให้กับผู้เขียนเรียงความได้ทันทีด้วย เนื่องจากงานวิจัยด้านการตรวจคะแนนเรียงความภาษาไทยโดยอัตโนมัติมีน้อยมาก งานวิจัยนี้จึงมุ่งเน้นไปที่การเปรียบเทียบแบบจำลองการให้คะแนนภาษาไทยโดยอัตโนมัติ ที่ใช้แบบจำลองการเรียนรู้ของเครื่องและการเรียนรู้เชิงลึก ที่ได้ถูกรายงานว่ามีประสิทธิภาพที่ดีในการทำนายคะแนนเรียงในความภาษาอังกฤษ งานวิจัยนี้พบว่าแบบจำลอง XGBoost มีประสิทธิภาพดีที่สุดในหลายสถานการณ์ที่ถูกพิจารณา และ แบบจำลองหน่วยความจำระยะสั้นแบบยาวที่ใช้ฟีเจอร์ที่สกัดจากข้อความโดยอัตโนมัติ มีประสิทธิภาพดีที่สุดในกลุ่มแบบจำลองการเรียนรู้เชิงลึก

Simplified Music Playlist Recommendation Using Sequential-Based Gated Recurrent Unit For Policy And Action Determination, Chanapa Channarong

Chulalongkorn University Theses and Dissertations (Chula ETD)

Recommendation systems are crucial for helping users discover music tailored to their preferences on streaming services. Reinforcement learning is a popular method for music recommendations. However, previous approaches have struggled with overfitting. The agent may become problematic for the new user context if, after a learning phase, it starts to predict actions only based on previous interactions. To overcome this limitation, previous methods require retraining, which require resetting all parameters in the agent. This study introduces the Policy GRU-RL method, which combines sequential-based learning and reinforcement learning to address overfitting without resetting all parameters. It utilizes a recurrent network by …