Offene Abschlussarbeiten

Measuring Vocabulary Overlap from Differentially-private Text Paraphrasing

An open question concerns how privacy mechanisms affect lexical fidelity and content retention. While most evaluations of private text generation focus on utility metrics without directly quantifying how much of the lexical material remains after privacy perturbation, the overlap between the vocabulary spaces  is an informative proxy for memorization risks (e.g., if overlap is high between original and privatized text) or utility degradation (e.g. if overlap is high between diverse topical domains).

References

Meisenbacher, S., Chevli, M., Vladika, J., & Matthes, F. (2024). DP-MLM: Differentially private text rewriting using masked language models. arXiv preprint arXiv:2407.00637.

Arnold, S. (2025). Inspecting the Representation Manifold of Differentially-Private Text. arXiv preprint arXiv:2503.14991.

A Corpus-based Study on Zero/That Complementizers in Conversational AI

In English, the most common type of object clause is introduced by the complementizer that, as in „I know that Peter will arrive soon.“ However, in most contexts this complementizer can be omitted, resulting in an asyndetic zero clause, as in „I know Peter will arrive soon.“

This alternation between that and zero complementizers has been widely studied in natural language, but much less attention has been given to how such variation manifests in computationally generated language. Investigating the distribution and conditions of zero vs. that complementizers in AI-generated text can provide valuable insights into how closely machine language approximates human usage patterns, the degree of syntactic naturalness achieved by language models, and potential stylistic biases in AI communication.

WildChat provides a corpus of one million user-agent conversations.

References

Conde-Silvestre, J. C., & Calle-Martín, J. (2015). Zero that-clauses in the history of English. A historical sociolinguistic approach (1424–1681). Journal of Historical Sociolinguistics, 1(1), 57-86.

Shank, C., Bogaert, J. V., & Plevoets, K. (2016). The diachronic development of zero complementation: A multifactorial analysis of the that/zero alternation with think, suppose, and believe. Corpus Linguistics and Linguistic Theory, 12(1), 31-72.

Uncovering Bias in the Space Allocation of Text-to-Image Generative AI 

Text-to-Image Synthesis (Rombach et al, 2022) offers creative potential but may also exhibit biases in how visual space is allocated to subjects from different demographic groups. Specifically, the placement of individuals in images, particularly in terms of spatial depth, may reflect stereotypical associations.

Since psychology posits that elements in the foreground are perceived as more dominant due to their visual accessibility (Arnheim, 1974; Leder et al., 2004), this study aims to examine the spatial depth allocation in response to demographic attributes. Specifically, prompts will contain explicit reference to multiple individuals characterized by different gender, race, and religious attributes.

To measure the spatial depth, the transformers library provides off-the-shelf pipelines for depth estimation:

transformers.pipeline(task="depth-estimation", model="Intel/dpt-large")

References

Arnheim (1974). On the Nature of Photography. Critical Inquiry, 1(1), 149-161.

Leder et al. (2004). A Model of Aesthetic Appreciation and Aesthetic Judgments. British Journal of Psychology, 95(4), 489-508.

Rombach et al. (2022). High-resolution Image Synthesis with Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 10684-10695).

Conversational agents powered by large language models have become essential tools for assisting users in personal, professional, and educational tasks. Previous studies have examined various aspects of dialogic interaction, including linguistic alignment and stylistic variation (Koulouri et al., 2016, Wang et al., 2023).

By conducting a lexico-grammatical discourse anayslsis, this thesis aims to investigate selected aspects of linguistic pragmatics and their role in the discourse, such as:

• Discourse Markers
• Intensifier Usage
• Modality Usage
• Types of Questions
• Types of Feedback
• ..

LMSYS-Chat-1M provides a large corpus of dyadic dialogues between users and agents, which can serve as a valuable dataset for discourse analysis.

References

Koulouri, T., Lauria, S., & Macredie, R. D. (2016). Do (and say) as I say: Linguistic adaptation in human–computer dialogs. Human–Computer Interaction, 31(1), 59-95.

Schmid, H. J., Würschinger, Q., Fischer, S., & Küchenhoff, H. (2021). That’s cool. Computational sociolinguistic methods for investigating individual lexico-grammatical variation. Frontiers in Artificial Intelligence, 3, 547531.

Wang, B., Theune, M., & Srivastava, S. (2023, November). Examining Lexical Alignment in Human-Agent Conversations with GPT-3.5 and GPT-4 Models. In International Workshop on Chatbot Research and Design (pp. 94-114). Cham: Springer Nature Switzerland.