Pretrained language models, particularly large language models (LLMs) like ChatGPT,
have achieved significant success across various NLP tasks. However, there is considerable
evidence that these models tend to adopt the cultural biases inherent in the datasets used for
training, thereby unintentionally reinforcing biased patterns and potentially causing harm.
This thesis aims to investigate these biases across different categories, such as racial
and gender biases, in both Estonian and English language models.
[1]
E. Kaukonen, A. Sabir, R. Sharma, How Aunt-Like Are You? Exploring Gender Bias in the Genderless Estonian Language: A Case Study, (2025 Nodalida)
Human values, as described in Schwartz’s theory, strongly shape how people interact with digital technologies.
For example, concerns about privacy influence how users manage their online data, the desire for freedom shapes
their engagement with open platforms, and the need for accessibility determines whether technologies can be used
by people with diverse abilities. These values are often reflected in online discussions, with platforms such as
Reddit providing a rich source of conversations where they are openly articulated. Prior research has shown that
mining Reddit can yield valuable insights for software requirements, demonstrating its potential as a source of
user feedback. Yet, despite their importance for attracting and retaining users, such values remain underrepresented
in software artefacts.
[1]
S. H. Schwartz, An Overview of the Schwartz Theory of Basic Values. Online Readings in Psychology and Culture (2012)
[2]
T. Iqbal et al., Mining Reddit as a New Source for Software Requirements. IEEE International Requirements Engineering Conference (2021)
[3]
A. Nurwidyantoro et al., Human values in software development artefacts: A case study on issue discussions in three Android applications. Information and Software Technology (2021)
Pretrained vision-language models (VLMs) like CLIP exhibit impressive performance in aligning images with language.
However, their behavior often reflects biases present in training data, affecting how they ground language in
visual inputs. This thesis examines how VLMs interpret objects and images through textual prompts, focusing
on the dynamics of language grounding and the emergence of bias in multimodal predictions.
[1]
Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried, Grounding Language Models to Images for Multimodal Inputs and Outputs (2023 ICML)