Fakultät Kommunikation und Umwelt
Refine
Document Type
- Master Thesis (37) (remove)
Language
- English (37)
Is part of the Bibliography
- no (37)
Keywords
- COVID-19 (2)
- Machine learning (2)
- NLP (2)
- PM2.5 (2)
- AI Hallucinations (1)
- AI integration (1)
- API-Driven Architecture (1)
- ASP.NET MVC (1)
- Agile (1)
- Air pollution (1)
Institute
Despite the remarkable achievements of large transformer-based pre-trained models like BERT, GPT in several Natural Language Processing (NLP) tasks including Sentiment Analysis (SA), challenges are still present for subdued source languages like Nepali. Nepali language is written in Devanagari script, has complex grammatical structure and diverse linguistic features. Due to the absence of balanced datasets, and computational resources for Nepali, achieving optimal result with the latest architecture remains challenging. For this reason, publicly available NLP modelsfor Nepali are very less, making research in this area difficult. This paper attempts to addressthis gap through the use of pre-trained transformer models specially tailored for Nepali from Hugging Face including BERT, DistilBERT, ALBERT, and DeBERTa for sentiment analysis in Nepali tweets on relatively balanced datasets. The models are trained on large Nepali datasets and optimized for NLP tasks involving Devanagari scripts. To evaluate the model’s performance, various tokenization strategies are investigated in order to capitalize on transformer-based embedding with the SoftMax function and confusion matrix. The outcomes of models are compared using the same datasets. The study’s results shows that DistilBERT achieved the highest accuracy rate of 88% in Nepali sentiment analysis tasks, followed by BERT and DeBERTa at 83% and 80%, respectively. However, ALBERT showed a low accuracy of 70%. The result of this approach shares valuable viewpoints for the field of sentiment analysis in diverse linguistic contexts.
The popularity of using transformer-based models like Bidirectional Encoder Representations from Transformers (BERT) for various Natural Language Processing (NLP) tasks is increasing rapidly. Unfortunately, the research is very limited for low-resource languages like Nepali. This study examines the utilisation of pre-trained BERT-based NERNepal for Named Entity Recognition (NER) tasks in Nepali text. The main goal is to investigate the efficiency of the NERNepal model, which has been pre-trained and fine-tuned on Nepali corpus data for NER. This research provides new insights by evaluating the NERNepal model on two distinct datasets. It addresses unique linguistic challenges specific to Nepali. The study also offers a detailed analysis of the model's strengths and weaknesses. By focusing on diverse datasets, this study shows how adaptable the model is and how its performance varies. These aspects have not been explored extensively before. The EverestNER dataset is one of the largest human-annotated datasets in Nepal so far, and the Nepali_NER dataset is also BIO-annotated for the NER task, which helped to compare the model’s prediction. The performance was better on the Nepali_NER dataset in comparison with another selected dataset. The EverestNER dataset contained many complex words for ORG connected with many tokens for a single entity with different contextual meanings and different annotations for the same word in different tokens as per context. Because of this, it created more confusion for the prediction, especially for the ORG entity. It had a similar issue with another dataset as well, but the label annotation was better in comparison. Furthermore, the research tries to clarify the difficulties and constraints related to utilising pre-trained BERT models for NER in low-resource languages such as Nepali. The study focuses on research areas on the efficacy of the model and its performance on two different datasets. Post-training with the EverestNER train dataset was attempted, but due to computational resource limitations, only a maximum of 3 epochs was possible, which did not improve the evaluation. It also has implications for enhancing language processing tools for Nepali. The results show the ability of a pre-trained BERT-based model to improve NER skills for Nepali text. However, further study is required to overcome the current obstacles to identifying complex words. The availability of high computational resources and the possibility of combining other NER approaches with a transformer-based model could increase the performance and robustness of the model.
Keywords: Named Entity Recognition, Natural Language Processing, BERT, low resources language
The recent development of generative artificial intelligence (AI) has inspired some game developers to try and integrate it into games. Anomalous behavior by generative AI or users seems undesirable and effort is being made to find ways on how to prevent anomalous responses by generative AI. However, these responses are a unique trait of generative AI and in our project, we explore strategies on how the negative effects of anomalous user and AI behavior can be mitigated, instead of trying to prevent them from appearing. We want to leverage their uniqueness as an intended game mechanic and see in what context these strategies can work. To achieve these goals, we researched already existing integrations of generative AI features in games and entertainment as well as possible connections to game mechanics and systems that are not related to AI. Based on this, we created 4 prototypes of conversational games with generative AI characters, which allowed the AI to create anomalous responses and did not limit the player freedom of what they may input. In our 36 participatory design sessions, we found that anomalous user and AI behavior can have a positive influence on the player's experience if a game's story is set up in a way that it can allow new and diverging story threads, that are created by the player and AI, to be integrated and if there are supplementary game mechanics, that allow the player to follow the intended story and complete the game.
During the past years, it is noticeable that the e-commerce industry has emerged drastically, offering accessibility to a variety of products to the customer where they can buy products from the comfort of their home. However, it is associated with a lot of new challenges for e-commerce businesses, particularly in understanding and managing the customer behavior patterns with the rising online product returns. Since the accessibility of online shopping has risen, assessing the critical factors related to product return and prediction has become a really challenging task for e-commerce vendors. This study also aims to segment the customers based on customer behavior prediction into two categories, i.e., high return risk and low return risk customers, and further develop strategies to reduce the online returns.
This study is rather structured into four integral parts, where each part provides the comprehensive analysis. The first aspect involves analyzing and identifying the customer behavior patterns leading to product returns. The second aspect predicts the online product return based on selected features and the third is to segment the customers based on online product return prediction and categorize them into high return risk, and low return risk customers. The final step involves development of strategies to reduce the product return based on the intense analysis conducted.
To achieve this meaningful research outcome, data analysis is conducted to understand the customer behavior patterns, and a Random Forest feature selector is used to identify the customer behavior patterns that lead to product return. Based on the identified features, classification models were applied to classify and predict whether the customer is going to return product or not. Furthermore, in these seven classification models such as Logistic Regression, Ada Boost, Decision Tree, Naive Bayes, XG Boost, K-Nearest Neighbors, and Random Forest were implemented and used to compare the performance of the classification models to find out the best performing model. Lastly, the segmentation of customers is carried out based on the online product prediction using Logistic regression with classification threshold method into high return risk and low return risk customer categories.
The results obtained help in understanding the customer behavior and reduce the online product return by developing the strategies. This study will eventually help the online businesses in reducing returns which will also enhance the customer satisfaction.
Data physicalizations are physical objects that encode data. The concept comes from data visualization, except for in a physicalization the data is presented as 3D tangible objects. Research shows that physical representation of data and being able to interact with it improve understanding and memorability of data, which supports the embodied cognition thesis, according to which cognition is supported by the body and the physical world and developmental psychology research, suggesting that manipulation of physical objects can promote understanding and learning. Alongside traditional physicalizations of statistical data, there also exist physicalizations of personal data, for example heartrate patterns, DNA data, one’s emotional states during the course of the week, etc. Research shows that such physicalizations, often embodied into everyday objects to promote familiarity and intuitive interaction, stimulate self-reflection as well as social interactions by telling others the stories behind the data. Some researchers thought about using data physicalizations as gifts and created example of such gifting artifacts. However, the design process of these artifacts was not based on the user research; nor were the physicalization gifts evaluated in any way. This is the gap that we are addressing in this master thesis, and the goal of our project was therefore to create and evaluate a personal data physicalization present using the human centered design approach. We started our work by conducting user research: performing interviews and defining the important characteristics of the future physicalisation gift. It was followed by a brainstorming session of possible design solutions and creation of the prototype. The prototype was then tested with users, addressing understandability of the data presented in the physicalisation and subjective perception of the gift. Our documentation of the findings and the research process provides a new step in the research of personal data physicalisations and lays out new topics for future investigations.
This thesis centers on enhancing the humanoid Pepper robot with Artificial Intelligence-driven information delivery capabilities in an educational setting using Large Language Model at student
service centers. In this work, Pepper was connected to GPT-Engine to act as a voice assistant robot.
With the help of GPT-Engine Pepper can answer questions about the University and its study programmes. The enhancement involves a client-server architecture, where the client-side application
enables students to interact with Pepper via speech, and the server-side, equipped with AI functionalities, processes these interactions and generates appropriate responses. This setup significantly
augments Pepper's capabilities, allowing for more accurate and efficient information delivery. An experiment was conducted with nine student participants of Hochschule-Rhein Waal, where they
asked questions about the University and study Programmes to inform themselves about the University and Study Programmes. During the interaction, they evaluated the robot's clarity of communication, accuracy of responses, and overall user experience. The findings indicate that while the AI enhancements generally improved Pepper's ability to communicate and provide information, areas such as voice recognition and response accuracy require further improvement. This research contributes to the field of human-robot interaction by illustrating a practical approach to augmenting existing service robots with AI capabilities. It underscores the importance of advanced AI integration in enhancing the functionality of service robots in educational environments. The study provides valuable insights into the broader application of AI-enhanced robots in the Information Guide role, highlighting the ongoing need for technological advancements in AI integration and humanrobot interaction.
Floods are the most common, devastating, and frequently occurring natural disaster nowadays.
Because of climate change, Europe is expected to see an even higher number of floods in the coming decade. As floods cannot be prevented, understanding the pattern and causes and being able to forecast could significantly reduce the losses during the flood. In July 2021, Germany faced a catastrophic flood, taking more than 180 people’s lives and causing around 40 billion euros of economic loss. To study the flooded region and the region at high risk, the Flood event 2021 has been visualized in this research. The Flood map has been created using a sentinel-1 image. In the visualization, floods in the Rhine River and Arh River can be seen, and the change in the pattern of flood from 12th July to 16th July has also been observed.
Different models exist that try to predict floods and warn as early as possible. Since Machine learning is popular nowadays in every sector, ML models have been implemented in this study to forecast the flood. The SARIMA, Random Forest, and LSTM have been implemented using historical data to predict the flood. The Random Forest model performed better than the remaining two models. The Isolation Forest model has also been implemented to classify the data into Flood and No Flood. This model performs well with the Flood classes but struggles in capturing the No Flood class.
API Driven Form Rendering
(2023)
This master’s thesis investigates the challenges and opportunities associated with dynamic form rendering in user interfaces, specifically in the context of Galeria Karstadt Kaufhof (GKK), a key player in the German retail sector. Collaborating with CODUCT Solutions GmbH, this research aims to enhance GKK’s operational efficiency and user experience in their digital transformation journey, focusing on Project Baldeney, a cloud-native platform that automates content management workflows.
The thesis begins by outlining the limitations of traditional, static methods of form rendering, highlighting the implications for development time and error frequency. It identifies the fragmentation in the existing API-driven architectures as a critical issue affecting system reliability and user experience. Against this backdrop, the work proposes a unified, API-driven framework to dynamically render forms for many user actions, such as data input and transactions.
The research takes a deep dive into existing systems—particularly Informatica’s Product Information Management (PIM) software and the new software in development Article Workbench to identify user experience issues, inefficiencies, and the complexities involved.
Through this examination, it offers a targeted solution to integrate various components of form management, like building, parsing, validation, and rendering, into a cohesive system.
The thesis aims to contribute to software development and user experience design substantially, by targeting improved development efficiency, reduced errors, and enhanced system reliability. By achieving these objectives, the research aspires to furnish GKK with
a robust solution for dynamic form rendering, thereby enhancing both user satisfaction and system performance in the fast-evolving digital landscape.
The literature review at hand addresses the following research question: “Does artificial intelligence and its future development potentially pose a significant threat to individuals’ well-being and humanity’s ethical values?”. For providing answers to the topic at hand, three hypotheses are investigated in detail out of which H0: “Artificial intelligence does pose a significant potential threat to individuals well-being and humanity’s ethical values.” is accepted in accordance with evaluation of findings gathered from 121 sources overall. Threats of artificial intelligence were identified in overall five fields of application, namely medical science, transport industry, potential militarization, surveillance as well as in interaction of artificial intelligence with consumers, for which the natural language processing model ChatGPT was investigated. Regarding medical sciences, threats include potential dehumanization effects through care robots equipped with artificial intelligence. In the transport industry, multiple issues may arise with autonomously operating vehicles while in potential military application of artificial intelligence, research is conducted in multiple countries. Regarding surveillance, artificial intelligence severely facilitates the implementation of social credit systems, as for instance seen in China. Additionally, human nature is embedded in the context of continuous development of artificial intelligence with aim of establishing a first approximation to guidelines of potentially needed limitations of further development of artificial intelligence.
It is concluded that, for implementing specific limitations on development of artificial intelligence, more research is needed for identifying the most suitable approach of limiting further development of artificial intelligence.
The modern work environment is subject to constant change with a notable shift from individual to team-based work structures. This transition necessitates an improved un-derstanding of how managerial actions influence the dynamics within these teams, to optimally align them for success. Despite the substantial research about leader effec-tiveness and team effectiveness based on organizational performance indicators, it is rather infrequent that the direct influence of leader behavior on teams is explored. This thesis aims to address this gap by answering how leader behaviors impact team pro-cesses in the context of German medium to large-scale IT-enterprises. Using a quanti-tative cross-sectional methodology, individuals working in team-based structures under direct supervision were surveyed. Analyzing data from 94 respondents via hierarchical linear regression analysis, several significant relationships were identified. In general, leader behavior has a strong positive impact on team processes. Furthermore, task-oriented leaders have the greatest positive influence on action processes, while relation-oriented leaders have the greatest positive influence on interpersonal processes. Change-oriented leaders also significantly influence action processes and interpersonal processes, albeit less strongly than their counterparts. These results provide implications for team management practices, emphasizing the importance of leader behavior for aligning team processes in the direction of success.