Media bias is an enduring feature of news dissemination, reflecting the subjective perspectives of its creators across history. From archaic records like "The Victory Stele of Naram-Sin" to contemporary news channels, bias permeates media, influencing political, social, and public health narratives. This research aims to investigate the persistent phenomenon of media bias and the potential of large language models (LLMs)(Kojima et al., 2022) in its detection and classification, in order to deploy publicly available software tools aiming to enhance media literacy among news consumers.

Traditionally, media bias served the interests of ruling powers; even with the rise of modern journalism, objectivity is often compromised by commercial pressures and inherent human biases. (Rodrigo-Ginés et al., 2024). As media landscapes evolve, bias continues to shape public opinion, impacting democratic processes and public health perceptions—evident during the COVID-19 pandemic, where polarized media narratives swayed public health decisions and fueled misinformation. (Recio-Román et al., 2023)

Current research on the effects of labeling media bias or propaganda, whether automatically or with human involvement, highlights the complexity of the issue. Depending on different circumstances, labeling can lead to negative outcomes (such as reinforcing filter bubbles by providing means to avoid news with a different perspective), no change in news consumption behavior at all, or, in some cases, an actual improvement in media literacy as intended (Zavolokina et al., 2024).

This research aims to develop a technical solution for the automatic labeling of biased media content, emphasizing several proposals that we hope will lead to a positive effect on media literacy among those presented with the system’s assessments.

These proposals include using a fine-grained taxonomy of bias types rather than a simple binary left/right labeling, focusing on detailed explanations for each model decision in natural language, marking bias at the sentence level rather than at the article or publication level to provide more insights, fine-tuning autoregressive models like GPT-3.5 or Mistral with high-quality examples instead of using “simple” bidirectional models like BERT(Brown et al., 2020) or non-finetuned models, and focusing on the German language, which has not yet been properly explored for such systems.

Understanding readers' perceptions when exposed to bias-labeled content is another facet of this research. It will explore how bias labeling influences readers' views on credibility and neutrality and whether real-time bias indicators affect news consumption behaviors. As mentioned, practical applications serve as a cornerstone of this research. One aim is to implement bias detection systems in real-world settings, such as search engines and news aggregators, to promote balanced information consumption. The development of user tools, like browser extensions highlighting media bias, intends to address public need for transparent information evaluation.

In essence, this research contributes to media literacy enhancement by demystifying media bias through advanced computational methods. By refining detection mechanisms, classifying bias more effectively, and implementing practical tools, it aims to fortify democratic discourse and public understanding, thereby addressing the pervasive influence of media bias in today’s interconnected world.

References

Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., … Amodei, D. (2020). Language Models are Few-Shot Learners. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, & H. Lin (Eds.), Advances in Neural Information Processing Systems (Vol. 33, pp. 1877–1901). Curran.

Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., & Iwasawa, Y. (2022). Large language models are zero-shot reasoners. Advances in Neural Information Processing Systems, 35, 22199–22213.

Recio-Román, A., Recio-Menéndez, M., & Román-González, M. V. (2023). Influence of Media Information Sources on Vaccine Uptake: The Full and Inconsistent Mediating Role of Vaccine Hesitancy. Computation (Basel). https://doi.org/10.3390/computation11100208

Rodrigo-Ginés, F.-J., Carrillo-de-Albornoz, J., & Plaza, L. (2024). A systematic review on media bias detection: What is media bias, how it is expressed, and how to detect it. Expert Systems with Applications, 237, 121641.

Keywords: Media Bias, Large Language Models, Bias Detection, Natural Language Processing, Journalism, Public Opinion, Taxonomy

Learning is at the heart of every progress the human species makes. It is most effective when it considers who we are as individuals, what learning approach we prefer and what we already know to begin with. In the digital age, we strive to capture such information in the form of a digital representation -- the so-called learner model --, to tailor learning-related systems to this information and build upon it to create more personalised learning experiences. Over recent years, the proliferation of diverse models across various educational applications and disciplines has made it challenging to access targeted research.

In this survey, we aim to address this gap, reviewing the latest advances in learner modelling and conducting a comprehensive analysis of the existing approaches, focusing on developments from 2014 to 2023. With the help of a systematic literature review, we want to provide designers and developers of learner models with a structured overview and simplified entrance into the topic and the field of learner models. We investigate the question: What do learner models look like and how are they filled, kept up-to-date, and used?

To this end, we analyse and classify existing approaches. Our findings provide a comprehensive and structured overview of the field of learner modelling, allowing researchers to navigate and understand the diverse approaches more easily and providing developers of learner models or adaptive systems with a practical tool to access relevant information according to their needs.

There is a class of entropy-coding methods which do not substitute symbols by code words (such as Huffman coding), but operate on intervals or ranges and thus allow a better approximation of the data entropy. This class includes three prominent members: conventional arithmetic coding, range coding, and coding based on asymmetric numeral systems. To determine the correct symbol in the decoder, each of these methods requires the comparison of a state variable with subinterval boundaries.

In adaptive operation, considering varying symbol statistics, an array of interval boundaries must additionally be kept up to date. The larger the symbol alphabet, the more time-consuming both the search for the correct subinterval and the updating of interval borders become. These entropy coding methods play an important role in all data transmission and storage applications, and optimising speed can be crucial.

Based on detailed pseudo-code, different known and proposed approaches are discussed to speed up the symbol search in the decoder and the adaptation of the array of interval borders, both depending on the chosen alphabet size. It is shown that reducing the big O complexity in practical implementations does not necessarily lead to an acceleration, especially if the alphabet size is too small. For example, the symbol determination at the decoder shows an expected low cpu-clock ratio (O(logn) algorithm versus O(n) algorithm) of about 0.62 for an alphabet with 256 symbols. However, for an alphabet with only 4 symbols, this ratio is 1.05, that means the algorithm with lower theoretical complexity executes slightly faster here. In adaptive compression mode, the binary indexing (BI) method proves to be superior when considering the overall processing time. Although the symbol search (in the decoder) takes longer than using other algorithms (e.g. cpu-clock ratio BI/O(logn) is 1.57), the faster updating of the array of interval borders more than compensates for this disadvantage (total ratio BI/O(logn) is 0.72). A variant of the binary indexing method is proposed, which is more flexible and has a partially lower complexity than the original approach. Specifically, the rescaling of cumulative counts can be reduced in its complexity from O(4n+[log2(n)−2]·n/2) to O(3n).