Open Theses

Open Theses

  • 2020/05/06

    Who is in your opinion bubble? User profiling for controversial topics discourse

    Bachelor Thesis, Master Thesis

    Social media have been a place for very passionate discussions regarding controversial topics such as nuclear energy or gun control. With NLP techniques for automated estimation of sociodemographic characteristics of users based on their language, we can analyze the average social profiles of groups with different opinion. How diverse are these ‚bubbles‘ and what are their characteristic properties? Which properties are characteristic for users vulnerable to fake news? Can we, in turn, use this information to improve fake news detection and/or opinion classification? go

    Supervisor: Prof. Dr. Iryna Gurevych

    Announcement as PDF

  • 2020/05/06

    Multi-hop reasoning for HowTo questions

    Bachelor Thesis, Master Thesis

    Recently, large pre-trained language models have shown great success on a variety of applications in NLP. Using these models, it has already been shown, that it is possible to find the correct answer to questions, that require reasoning over multiple sentences of the input text. Preferably, however, we would like to have computer systems, that are able to aggregate different information on their own in a meaningful ways, that help people with their problems. In this thesis we like to investigate, how well models can reason over the (causal) implications of multiple steps of HowTo instructions from WikiHow. Specifically, we want to investigate, how well language models are able to re-construct the order (when required) of multiple step-wise instructions, leading to a specific goal (e.g. “How to build a house?”) go

    Supervisors: Prof. Dr. Iryna Gurevych, Max Glockner, M.Sc.

    Announcement as PDF

  • 2020/05/06

    Changes in expressed opinion strength when facing disagreement

    Bachelor Thesis, Master Thesis

    Social media have been a place for very passionate discussions regarding controversial topics. Do these discussions have any effect or are we all stubborn victims of our filter bubbles? Using data from social media conversations of the same users over several months, the aim of this project is to find if social media users ever weaken or strengthen the intensity with which they communicate their opinion to others, and if so, what are the possible interactions causing such a change. go

    Supervisor: Prof. Dr. Iryna Gurevych

    Announcement as PDF

  • 2020/05/06

    Claim Validation on Fiction Books

    Bachelor Thesis, Master Thesis

    “Fake news” has been a major threat for democratic societies. Existing approaches developed to fight this problem still fail on real world data. One big challenge is that apparently current systems leak gold information from their pre-trained process. The goal of this thesis is to build a system capable of verifying information that is not available in Wikipedia. go

    Supervisors: Prof. Dr. Iryna Gurevych, Gisela Vallejo, M.Sc.

    Announcement as PDF

  • 2020/05/06

    Learning to Generate High-quality Dialogue

    Master Thesis

    Applications of open-domain conversational agents are becoming widespread. However, training such agents to generate high-quality responses is still a big challenge as the quality of responses depends on various factors. Recent methods train agents directly by gold responses from training sets. These methods have been shown generating low-quality responses at evaluation. In this thesis, we propose to train a function that quantifies the quality of the generated responses by a deep preference learning method. Then, we use this function as a reward estimator in a reinforcement learning model to train agents. go

    Supervisors: Prof. Dr. Iryna Gurevych, Dr. Mohsen Mesgar

    Announcement as PDF

  • 2020/05/06

    Choose your words carefully!

    Frame Analytics in Natural Language Processing

    Bachelor Thesis, Master Thesis

    Framing is about how we organize, perceive, and communicate about reality. By choosing specific words, attributes, arguments, associations, images or emotions, we aim to communicate our perspective. Carefully framing a topic can have a significant impact on the public opinion. Therefore, we are especially interested to answer the key question: What makes a frame successful? go

    Supervisors: Prof. Dr. Iryna Gurevych, Christopher Klamm , M.Sc., M.A.

    Announcement as PDF

  • 2020/05/06

    Measuring the Coherence of Dialog Responses

    Master Thesis

    Applications of conversational agents are becoming widespread. However, training such agents to generate high-quality responses is still a big challenge as the quality of responses depends on various factors. One of these factors is coherence. In this thesis, we built upon one of our existing models to measure the coherence of a response to its preceding dialog utterances using BERT-based language models.

    go

    Supervisors: Prof. Dr. Iryna Gurevych, Dr. Mohsen Mesgar

    Announcement as PDF

  • 2020/05/06

    Modernizing Scientific Communication with NLP

    Bachelor Thesis, Master Thesis

    Modern science revolves around publications. The worldwide acceleration of research and the democratization of scientific publishing have led to an unprecedented increase in publication volumes. Today anyone can put a paper on arXiv and get it cited, but how do we know if it actually is good research worth building upon? go

    Supervisors: Prof. Dr. Iryna Gurevych, Ilia Kuznetsov, Dipl.-Ling.

    Announcement as PDF

  • 2020/05/06

    Argument Paraphrasing for Author Obfuscation

    Master Thesis

    Conversational agents shall be able to provide realistic arguments in an opinionated discussion. The training data for learning such natural argumentation are ideally real statements provided publicly by human speakers. However, such opinionated statements often incorporate very specific stylistic elements of an individual speaker with a distinct personality. Being able to reidentify such opinion source through a conversational agent’s response is ethically controversial. Therefore, the aim of this project is to explore to which extent agents can learn to present arguments and opinions in a neutral manner that doesn’t allow to track down the original training data authors. go

    Supervisor: Prof. Dr. Iryna Gurevych

    Announcement as PDF

  • 2020/05/06

    Multi-Modal Commonsense Reasoning

    Master Thesis

    When we as humans reason about our world, we use information from multiple modalities (vision, sound, text, smell, etc.) to reach conclusions. These conclusions are often based on implicit experiences of our past and can be regarded as common sense reasoning. This thesis has the ambition of understanding what commonsense entails and how we can train an AI that is able to perform true reasoning. go

    Supervisors: Prof. Dr. Iryna Gurevych, Jonas Pfeiffer, M.Sc.

    Announcement as PDF

  • 2020/02/10

    Summarizing debates on large scale

    Bachelor Thesis, Master Thesis

    Argumentation Mining deals with the automatic extraction of argumentative structures from natural text. Given the recent advances in this field, we are able to identify large numbers of arguments supporting or opposing a controversial topic (e.g. “nuclear energy”). To make this information digestible for users, we would like to aggregate the arguments by grouping similar arguments together and present the most prominent ones to the user. go

    Supervisors: Prof. Iryna Gurevych, Tilman Beck, M.Sc.

    Announcement as PDF

  • 2020/02/10

    Entity and Relation Extraction from Chatlogs

    Bachelor Thesis, Master Thesis

    Dialogues accompany us on a daily basis in our lives; be it on social media like Facebook and Twitter, or instant messengers like WhatsApp. Recent advances in Deep Learning led to an increasing popularity of virtual assistants like Amazon’s Alexa or Apple’s Siri and the development of increasingly sophisticated chatbots. However, current systems only work well in very pre-defined settings. The need for large amounts of training data impaired with informal speech makes it difficult to train neural networks which generalize well to new domains. A key challenge is to recognize entities and to identify their relations across different dialog turns.

    The goal of this thesis is to investigate several approaches for entity recognition in chatlogs as well as finding and identifying relevant links between different entities. go

    Supervisors: Prof. Iryna Gurevych, Ji-Ung Lee, M.Sc.

    Announcement as PDF

  • 2020/02/10

    Generating Creative Language

    Bachelor Thesis, Master Thesis

    We would like to be able to explore the area of creative argument generation: can we generate creative, humorous, and insightful arguments automatically? Despite recent advances in natural language processing, natural communication with artificial intelligence remains difficult. This is in part because AI systems generally struggle with using and understanding creative language. Creative language (including metaphors, similes, humor and more) is particularly challenging for natural language generation (NLG), the research area that works to automatically generate natural, fluent text. In order to develop friendly, and convincing AI, we need to be able to incorporate creative language into our models of language generation. go

    Supervisors: Prof. Iryna Gurevych, Kevin Stowe, PhD

    Announcement as PDF

  • 2019/06/13

    Generating Text from Graph-based Data

    Bachelor Thesis, Master Thesis

    Recently, graph-to-sequence models have been applied to the task of text generation from structured data. Usually, these models incorporate Graph Neural Networks (GNNs) as graph encoders for learning effective graph representations. For instance, given a knowledge graph, we are interested in verbalizing a text that reproduces the information contained in the graph. The goal of this project is to develop and research deep learning techniques to generate textual information from graph-based data, such as meaning representations and knowledge graphs. go

    Supervisors: Prof. Iryna Gurevych, Leonardo Filipe Rodrigues Ribeiro, M.Sc.

    Announcement as PDF

  • 2019/04/09

    Online learning for interactive annotation

    Master Thesis

    The goal of this thesis therefore is to research and evaluate online learning algorithms, i.e. algorithms that can be used to update models by only training on newly incoming data, reducing the need to retrain from scratch. go

    Supervisors: Prof. Iryna Gurevych, Jan-Christoph Klie, M.Sc.

    Announcement as PDF

  • 2018/11/28

    Addressing The Speaker Consistency Issue in Conversational Agents

    Master Thesis

    Supervisors: Prof. Iryna Gurevych, Dr. Mohsen Mesgar

    Announcement as PDF

  • 2018/03/01

    Using Coreference Resolution in Question Answering and How to Make the Best of It

    Master Thesis

    Coreference resolution is the task of finding different expressions that refer to the same entity. The availability of coreference information is beneficial for several applications including question answering. In this project, we want to examine the effect of state of the art supervised and unsupervised coreference resolvers in question answering systems by answering the following questions:

    How good should a coreference resolver be in order to improve the performance of a QA system?

    Due to the limited generalization of coreference resolvers, are unsupervised systems better suited for QA?

    Would it be more beneficial for QA if we customized coreference resolvers for only a subset of noun phrases (e.g. only resolving pronouns)? go

    Supervisors: Prof. Dr. Iryna Gurevych, Dr. phil. Nafise Sadat Moosavi

  • 2018/02/19

    VerbNet Semantic Role Labeling

    Master Thesis

    Semantic Role Labeling (SRL) automatically assigns labels to the participants of a situation described in a sentence. For example, given the sentence

    Bill hits a table with a hammer

    we want to discover that “Bill” is the Agent, “table” is the Theme and “hammer” is the Instrument of the action. Whereas FrameNet- and PropBank-based SRL has received significant attention, almost no work has been done on VerbNet-based SRL, which is more linguistically involved, but also more theoretically motivated and language-independent. Our hypothesis is that using additional constraints from semantic role theory can improve the performance of SRL systems, which can be then used as a preprocessing step for question answering, machine translation and other tasks. go

    Supervisors: Prof. Dr. Iryna Gurevych, Ilia Kuznetsov, Dipl.-Ling.

    Announcement as PDF

  • 2018/02/19

    Text Analytics using Supervised Text Classification

    Bachelor Thesis, Master Thesis

    Text classification has become a popular solution to many problems brought forward by social media platforms such as Twitter. In a nutshell, supervised text classification extracts relevant information from manually classified documents and learns a model from the extracted information. The growing text analytics market heavily relies on text classification to offer services such as sentiment analysis, document categorization, scientific discovery etc. With the help of DKPro-TC, our framework for supervised text classification, you will study novel text analysis tasks and collect new insights into artificial intelligence. go

    Supervisors: Prof. Dr. Iryna Gurevych, Dr.-Ing. Johannes Daxenberger

    Announcement as PDF

  • 2018/02/19

    Non-Factoid Question Answering

    Bachelor Thesis, Master Thesis

    Automatic Question Answering (QA) is a broad research area with the ultimate goal of finding useful answers for arbitrary questions. An important challenge in QA is finding answers for complicated questions that require expert knowledge, past experience, or opinions. Even though knowledge graphs contain huge amounts of factual information, they cannot be used to answer such complicated questions. We therefore explore innovative approaches to search for relevant, complex, and unstructured information within the large number of discussions in community Question Answering platforms such as StackExchange and Quora. go

    Supervisors: Prof. Dr. Iryna Gurevych, Andreas Rücklé, M.Sc.

    Announcement as PDF

  • 2018/02/19

    Natural Language Generation

    Bachelor Thesis, Master Thesis

    In the past few decades we have witnessed a fast growth of information content in all types of mass media. As the volume of published data grows, accessing and processing information in the shortest possible time is becoming of vital importance. Natural Language Generation (NLG) is the process of mapping from some underlying representation of information to a presentation of that information in an easily accessible textual or spoken form. NLG application spectrum includes question answering and recommendation systems, personalized assistants and human-computer interfaces, document summarizers and news generation machines. The goal of this thesis is to develop novel NLG approaches, design accurate evaluation methods and use NLG techniques for solving real-world problems. go

    Supervisors: Prof. Dr. Iryna Gurevych, Yevgeniy Puzikov, M.Sc.

    Announcement as PDF

  • 2018/02/19

    Deep Learning with External Memory

    Master Thesis

    Can you read a short story and answer questions about its contents afterwards? Most likely. Can neural networks do the same? Not quite. Most of today’s neural networks can handle a large number of individual sentences, but have trouble with full stories or books. Memory-augmented neural networks contain an external memory component which can store, retrieve and modify large amounts of data. This makes them a perfect fit for understanding stories or other machine comprehension tasks. go

    Supervisors: Prof. Dr. Iryna Gurevych, Michael Bugert, M.Sc.

    Announcement as PDF

  • 2018/02/19

    Active Deep Learning for Natural Language Processing

    Bachelor Thesis, Master Thesis

    Recently, deep learning gained much popularity. However, deep learning methods suffer from data sparsity – especially in NLP, where data is often cost-intensive to label. While active incremental learning approaches cope with this, they are not very well explored for deep learning models and most previous work has been done on convolutional neural networks for computer vision tasks. Furthermore, there is still only little support for such methods in the most common deep learning libraries.

    The goal of this thesis is to develop an active learning extension for a popular deep learning framework, such as PyTorch or Tensorflow, and compare different strategies for a natural language learning task. go

    Supervisors: Prof. Dr. Iryna Gurevych, Ji-Ung Lee, M.Sc.

    Announcement as PDF