Optimizing LLM-based Discourse Analysis in Low-Resource Languages: A Case Study of Bulgarian Euro Adoption Narratives

The article by Dobromir Tsolyov, Keith Kiely, Georgi Goranov, Iglika Ivanova, and Stanimir Dimitrov, titled “Optimizing LLM-based Discourse Analysis in Low-Resource Languages: A Case Study of Bulgarian Euro Adoption Narratives,” investigates how Large Language Models can be configured to perform stable, multidimensional discourse analysis of politically and culturally specific content in Bulgarian. The study focuses on narratives surrounding Bulgaria’s euro adoption debate and asks a methodological question: how much analytical complexity can an LLM handle before its outputs become unstable across repeated runs?

The authors introduce the concept of an LLM “complexity ceiling”—the highest-complexity configuration within a predefined search space that still produces sufficiently stable results. Rather than increasing complexity until performance breaks, they propose a “descent to feasibility” approach: starting with a demanding configuration and progressively reducing internal or external complexity until a predefined stability threshold is reached. The objective is therefore not simply to maximize reasoning depth, but to find a configuration that combines higher-order reasoning with reproducible scoring.

The empirical case study uses Bulgarian-language Facebook posts published between 2 May and 22 June 2025, during the debate surrounding then-President Rumen Radev’s proposal for a referendum related to euro adoption. The original collection contained 445 posts retrieved through the Meta Content Library using the keywords euro and referendum. After duplicate removal, the dataset contained 302 posts. For the pipeline optimization experiment, the authors selected a diversity-oriented sample using stance categories (pro, anti, and neutral), multilingual embeddings with BAAI/bge-m3, and KMeans clustering.

The analytical framework combines two approaches: Indicators of Weak Hegemony (IWH) and FABLE, a framework for assessing potential misinformation harm. Together they produce 11 analytical dimensions, including fragmented consensus, counter-hegemonic forces, crisis of authority, social fragmentation, actionability, believability, likelihood of spread, and exploitativeness. These abstract dimensions are operationalized through 62 questions, with Yes, No, and Unclear answers. Weighted responses are subsequently transformed into ordinal scores of 0, 1, or 2 for each dimension.

A particularly important methodological contribution is the proposed “80/80 rule” for stability. A configuration is considered stable when at least 80% of the sampled texts receive the same modal score in at least 8 out of 10 independent runs, together with an average modal consistency above 80% across the dimensions. Crucially, the authors emphasize that stability is not the same as accuracy: because there is no straightforward ground-truth answer for complex discourse interpretation, repeated consistency is treated as a proxy for reliability rather than proof that the model’s interpretation is correct.

The experiments use Gemini 3 Flash and manipulate several sources of complexity: the model’s thinking level, temperature, prompt size, the placement of analytical guidelines, and whether questions are grouped into large methodology-focused prompts or smaller dimension-focused prompts. In total, 14 configurations were systematically tested, with two additional configurations used for reference.

The results suggest that more reasoning is not automatically associated with greater analytical stability. The most stable methodology-focused configuration used large methodology-specific batches, a low thinking level, temperature 1, and guidelines placed in the system prompt. At the same time, smaller dimension-focused prompts could support a higher reasoning budget and remained comparatively stable. The authors therefore identify two different strategies: greater contextual complexity combined with constrained reasoning, or reduced contextual complexity combined with greater reasoning depth.

Some analytical dimensions proved substantially harder to stabilize than others. In particular, Fragmented Consensus (IWH-1) showed instability across many configurations, while Believability (FABLE-3) was problematic under dimension-focused prompting. The latter contained 15 questions, compared with roughly 4–6 for most other dimensions, suggesting that instability can result not only from semantic ambiguity but also from the structural complexity of the analytical task itself.

Overall, the paper argues that reliable LLM-based discourse analysis in a low-resource language such as Bulgarian requires deliberate calibration of reasoning depth, prompt complexity, and contextual guidance. Its main contribution is therefore less about determining the content or validity of particular euro-adoption narratives and more about establishing a methodology for making automated analysis of such narratives reproducible. The authors also stress the study’s limitations: it uses one model, one political case, one language, and a constrained search space, and evaluates stability separately from ground-truth correctness. Future work is intended to introduce human-annotated validation, test additional languages and models, expand the configuration search space, and extend the approach to multimodal content.

Suggested citation

Tsolyov, Dobromir, Keith Kiely, Georgi Goranov, Iglika Ivanova, and Stanimir Dimitrov. 2026. “Optimizing LLM-based Discourse Analysis in Low-Resource Languages: A Case Study of Bulgarian Euro Adoption Narratives.” Proceedings of the Seventh International Conference on Computational Linguistics in Bulgaria (CLIB 2026), 319–332.

Article: Full-text PDF

Proceedings: CLIB Conference Proceedings