Abschlussarbeiten in Data Science

Abschlussarbeiten in Data Science werden zweimal im Jahr durch den Prüfungsausschuss in einem zentralen Verfahren vergeben. Außerhalb dieses Prozesses können wir nur in seltenen Ausnahmefällen Themen vergeben.

Externe Abschlussarbeiten betreuen wir nur in Ausnahmefällen, falls die Aufgabenstellung gut zu den Forschungsthemen der Professur passt. Fragen Sie dazu bitte gezielt bei Prof. Schenkel nach, wenn Sie einen Vorschlag für ein Masterarbeitsthema haben, dass Sie außerhalb der Universität bearbeiten wollen.

Beispiele für in unserer Arbeitsgruppe abgeschlossene Masterarbeiten in Data Science

[MA] Reconstruction of Argumentation Graphs

Argumentation can be understood as the activity of using arguments to convince, agree, or disagree people with people about a point of view. In our daily lives, argumentation is one of the most common behaviors in applying natural language. For example, social media users would respond to controversial topics using their stances and opinions. The collection and analysis of user ideas are critical to studying social phenomena and trends. However, it is hard to analyze all collected arguments since processing enormous data size needs much time and human costs, which is undesirable. This requires more efficient methods. A possible solution might be the research in computational argumentation because computers can handle numerous data efficiently. Besides social phenomenon analysis, other areas such as business and linguistics also benefit from studying argumentation.

Computational argumentation is a growing research field that yield many new methods in this area. This work is inspired by a study investigating in transforming natural language texts to argument graphs. In this thesis, we base on the previous studies and explore deep into the steps of each part, including classifying major claims, inferring relations between statements, and constructing argument graphs, and investigate in approaches for improvement. We propose a new method in major claim classification, which is to find the statement describing the core idea of the discussion, and obtain an excellent enhancement. Moreover, we introduce state-of-the-art methods to estimate the relations between arguments. We suggest six methods in the step of argument graph construction, which also give satisfactory results. There are some limitations to our research. We discuss them and explore some possible further improvements for achieving a better result in the future studies.

[MA] Fine-tuning a Transformer model for Multilingual document semantic similarity

An information retrieval system’s purpose is to return results that are relevant to the user’s query. Information relevant to the user’s request may not exist in the user’s native language in some instances. It’s also possible that the user can read papers in languages other than his or her native tongue but has trouble forming inquiries in them. The primary goal of Multilingual Information Extraction is to locate the most relevant information accessible, regardless of the query language.

Artificial intelligence (AI) has become an increasingly popular research field in recent years. Similarly, Natural Language Processing (NLP) has become an important point of discussion. Neural networks, do exceptionally well in this field. The speed and performance of neural networks dealing with diverse NLP tasks have been greatly enhanced due to a variety of effective learning methods and technologies.

The recent advances in NLP transfer learning have resulted in powerful models, mostly from the tech giants like Google, Facebook, Microsoft, etc. which perform well on NLP tasks in the general domain. In this thesis, we are going to fine-tune multilingual transformer models for the domain of engineering data both in English and German Languages. Hence, we need a language independent model - which can able to learn it’s parameters (weights and bias) of any language-specific features. First, we will describe how multilingual transfer is implemented, with the focus on state-of-the-art transformer models. Then, in the methodology part, we leverage our engineering domain data of English-German languages to fine-tune multilingual transformer models.

[MA] Towards (semi) automated literature-based complete transformer-based MCQ generation model for data base related field deployment

- kein Abstract verfügbar -

[MA] Creating and implementing a pipeline for retrieving and ranking arguments by relevance and quality for controversial questions

Abstract: Argumentation is considered to be a foundational discipline. Initially, its objectives are to foster critical thinking and logical reasoning, to reach a resolution when people disagree, persuade or convince others of a particular viewpoint or position, and also can be a tool for knowledge exchange.

Individuals can explore arguments that either support or attack their own opinions, leveraging their personal knowledge and life experiences, but they also can use search engines (e.g., Google) accessed by the Internet. In this work, we focus on the arguments taken from the Web. The user could ask (input the query) the search engine a particular question, e.g., “Should I own a dog?” and will expect to receive an answer in the form of a list of Web pages (sorting by relevance), textual information, images, videos, news articles, and social media’s posts.

Usually, arguments for a specific question are in the text, which is a part of the Web page (also called “document”). The document may contain argumentative and non-argumentative text spans. The aim is to retrieve the documents, such that their argumentative parts are relevant to the query and highly qualified (argumentative). However, there is the issue that the retrieved documents may consist of arguments with low relevance to the query, low quality, or falsified, and there is usually no clear stance. Therefore, these documents will not satisfy the user’s expectations, or the user will use the wrong, fake, biased arguments to support the position.

The problem with search engines like Google is that users looking for reasonable arguments within a short time are required to do a significant amount of work after submitting their query. This work includes tasks such as reading pages, identifying arguments, filtering duplicates, and manually ranking them. In contrast, argument search engines aim to alleviate this burden by handling these tasks for users and presenting them with the best arguments. This proves advantageous in debates, interviews, and political discussions, as it ensures the availability of the strongest arguments for making informed decisions.

Our work was inspired by the Touché Lab Task 1 named “Argument Retrieval for Controversial Questions”, whose objective is to retrieve and rank documents by relevance to the topic, by argumentativeness of the documents (quality), and to detect their stance towards the topic. In this work, we investigate various methods and techniques for argument mining (i.e., automatic extraction of arguments from the document) and preprocessing for the purpose of working with individual arguments from the document rather than the entire text as a whole. We applied stance classification (i.e., determining whether the premise supports or attacks the specific claim) and quality prediction to get high-quality arguments 1 . To expand the search for the re-ranking model, we utilize query augmentation, which is performed with the assistance of ChatGPT. The primary objective is to optimally combine these approaches to retrieve highly relevant results with high-quality arguments and demonstrate that working with individual arguments produces better results than working with the entire text.

For our experiments and evaluation, we utilize several datasets and resources. The “ClueWeb22-B” corpus and controversial questions provided by the Touch´e Lab served as the basis for our analysis. The SNLI dataset is utilized to establish relations between claims and premises. At the same time, the “args.me” dataset is explicitly employed for stance classification. To predict the argument’s quality, we rely on the “Webis-ArgQuality-20” and “IBM-ArgQ-Rank-30kArgs” datasets.

To evaluate the effectiveness of our approach, we compare our results with the baseline of Touché Task 1. To ensure fair comparisons, we utilize manually annotated judgments as a benchmark for both our results and the baselines. Our approach demonstrates superior performance in the nDCG measurement compared to the baseline of Touché Lab Task 1 and achieves an accuracy of 0.54 for stance classification. It highlights the effectiveness and competitiveness of our approach in retrieving and ranking relevant arguments by relevance and quality, as well as classifying them by stance.

[MA] Natural Language Processing in Accounting

Abstract: This thesis offers an approach to detect booking duplicates by calculating sentence similarity as an application of Natural Language Processing. These bookings are exports of an accounting software. Among lots of other information, each booking has a booking note which is a short text written by the person who created the booking in the accounting software. The presented approach is part of a larger project in which all booking information is analyzed but in this thesis, solely the textual information of the notes is used for determining the similarity of two bookings. Several models are used for calculating the similarity of booking pairs and their results are compared. One important research objective is the comparison of the TFIDF as an application of the vector space model and language models as BERT and sentenceBERT which are using word and sentence embedding vectors. The best models achieve a F1-score of 0.6004 and an AUC-score of 0.555. Thorough analysis of True Positives, False Positives and False Negatives shows that embedding vectors not only offer advantages but other challenges are a consequence of using word embedding vectors when short texts are analyzed.

Keywords: Natural Language Processing - Duplicate Detection - Accounting - Short Texts

[MA] Automatic Fake News Detection on Tweets

 - kein Abstract verfügbar -