Selected work

Representative Papers

Robustness and User-Perceived Value of Popularity Calibration in Music Recommendation: A User Study

Aug 05, 2026

This study investigates whether popularity calibration genuinely enhances user experience in music recommendation and examines its reliability across varying levels of user listening history and item familiarity. The authors construct three types of playlists—high-popularity, low-popularity, and calibrated—and employ a controlled naive recommender to generate personalized lists. Calibration is quantified using Jensen–Shannon divergence (JSD), and subjective user feedback is collected through controlled experiments. This work presents the first systematic validation of JSD’s stability with respect to real users’ perceived calibration. Results indicate that while users can discern differences in popularity, they do not exhibit a significant preference for calibrated recommendations. Moreover, computed popularity labels show only weak alignment with users’ subjective judgments, and the relationship between JSD and perceived calibration is significantly moderated by item familiarity, playlist composition, and the availability of historical interaction data.

0 citationsRead paper

Mapping Political-Elite Networks in Europe with a Multilingual Joint Entity-Relation Extraction Pipeline

Jun 25, 2026

Existing approaches struggle to capture at scale the complex, informal, and adversarial relationship networks among European political elites, constrained by manual annotation or insufficient cross-lingual capabilities. This work proposes the first end-to-end, scalable multilingual joint entity–relation extraction pipeline, integrating span-based named entity recognition, a three-stage Wikidata entity linking module, ontology-constrained relation extraction, and guided decoding to automatically construct signed, timestamped knowledge graphs from massive news corpora. The framework enables language-agnostic entity alignment and directed relation extraction, achieving 68.2% (strict) to 93.7% (lenient) accuracy on a gold-standard set of 3,491 relations. It successfully reconstructs Austrian party system evolution and uncovers the structure of Polish political–business networks.

0 citationsRead paper

Comparing BERT Sentence-Pair Classification and Few-Shot LLM Prompting for Detecting Threat and Solution Framing in German Climate News

Jun 24, 2026

This study addresses the automated identification of threat and solution framing at the sentence level in German-language climate news, aiming to facilitate large-scale media content analysis. For the first time in this context, it systematically compares a fine-tuned, context-aware BERT model (deepset/gbert-large) against a few-shot prompted open-source large language model (Llama 4 Maverick), the latter augmented with chain-of-thought reasoning, structured output formatting, and confidence scoring. Experimental results show that the BERT-based dual binary classifier achieves an F1 score of 0.83 on both tasks, significantly outperforming the LLM’s 0.78. Ablation studies further confirm that incorporating preceding sentence context is crucial for enhancing BERT’s performance. This work provides an effective approach for frame analysis in German and highlights the impact of contextual modeling and model paradigm choice on frame detection accuracy.

0 citationsRead paper

Deriving the Variance-Minimizing Design for Standard Addition via c-Optimality

Jun 05, 2026

This study addresses the lack of optimal experimental design guidance for the standard addition method under non-decreasing measurement error structures. Building on c-optimality theory and integrating linear response modeling with analysis of variance, the authors systematically derive an optimal two-concentration-point design that minimizes estimation variance under constant, linear, or quadratic error growth. This work represents the first application of optimal experimental design theory to the standard addition method and demonstrates that the proposed two-point design achieves universal optimality across all considered non-decreasing error scenarios. Notably, the optimal allocation of replicate measurements deviates from the conventional 50:50 ratio and yields minimum-variance unbiased estimates without requiring weighted regression.

0 citationsRead paper

Infini-News: Efficiently Queryable Access to 1.3 Billion Processed Common Crawl News Articles

May 18, 2026

This study addresses the challenges of acquiring large-scale news corpora, which are often hindered by the high cost of commercial archives and the substantial storage and processing demands of open-source datasets such as CC-News. The authors comprehensively clean and parse metadata from the entire CC-News archive dating back to August 2016, performing language identification on 1.35 billion news articles using GlotLID, lingua, and CommonLingua, and introducing multi-source geolocation annotations that cover 83.4% of the articles across 222 countries. Leveraging suffix arrays, they construct an Infini-gram index enabling sub-second full-text pattern matching for arbitrary queries, thereby substantially lowering the barrier to cross-national, longitudinal media research.

0 citationsRead paper

Recent publications

Latest Papers

Robustness and User-Perceived Value of Popularity Calibration in Music Recommendation: A User Study

Aug 05, 2026

This study investigates whether popularity calibration genuinely enhances user experience in music recommendation and examines its reliability across varying levels of user listening history and item familiarity. The authors construct three types of playlists—high-popularity, low-popularity, and calibrated—and employ a controlled naive recommender to generate personalized lists. Calibration is quantified using Jensen–Shannon divergence (JSD), and subjective user feedback is collected through controlled experiments. This work presents the first systematic validation of JSD’s stability with respect to real users’ perceived calibration. Results indicate that while users can discern differences in popularity, they do not exhibit a significant preference for calibrated recommendations. Moreover, computed popularity labels show only weak alignment with users’ subjective judgments, and the relationship between JSD and perceived calibration is significantly moderated by item familiarity, playlist composition, and the availability of historical interaction data.

0 citationsRead paper

Mapping Political-Elite Networks in Europe with a Multilingual Joint Entity-Relation Extraction Pipeline

Jun 25, 2026

Existing approaches struggle to capture at scale the complex, informal, and adversarial relationship networks among European political elites, constrained by manual annotation or insufficient cross-lingual capabilities. This work proposes the first end-to-end, scalable multilingual joint entity–relation extraction pipeline, integrating span-based named entity recognition, a three-stage Wikidata entity linking module, ontology-constrained relation extraction, and guided decoding to automatically construct signed, timestamped knowledge graphs from massive news corpora. The framework enables language-agnostic entity alignment and directed relation extraction, achieving 68.2% (strict) to 93.7% (lenient) accuracy on a gold-standard set of 3,491 relations. It successfully reconstructs Austrian party system evolution and uncovers the structure of Polish political–business networks.

0 citationsRead paper

Comparing BERT Sentence-Pair Classification and Few-Shot LLM Prompting for Detecting Threat and Solution Framing in German Climate News

Jun 24, 2026

This study addresses the automated identification of threat and solution framing at the sentence level in German-language climate news, aiming to facilitate large-scale media content analysis. For the first time in this context, it systematically compares a fine-tuned, context-aware BERT model (deepset/gbert-large) against a few-shot prompted open-source large language model (Llama 4 Maverick), the latter augmented with chain-of-thought reasoning, structured output formatting, and confidence scoring. Experimental results show that the BERT-based dual binary classifier achieves an F1 score of 0.83 on both tasks, significantly outperforming the LLM’s 0.78. Ablation studies further confirm that incorporating preceding sentence context is crucial for enhancing BERT’s performance. This work provides an effective approach for frame analysis in German and highlights the impact of contextual modeling and model paradigm choice on frame detection accuracy.

0 citationsRead paper

Deriving the Variance-Minimizing Design for Standard Addition via c-Optimality

Jun 05, 2026

This study addresses the lack of optimal experimental design guidance for the standard addition method under non-decreasing measurement error structures. Building on c-optimality theory and integrating linear response modeling with analysis of variance, the authors systematically derive an optimal two-concentration-point design that minimizes estimation variance under constant, linear, or quadratic error growth. This work represents the first application of optimal experimental design theory to the standard addition method and demonstrates that the proposed two-point design achieves universal optimality across all considered non-decreasing error scenarios. Notably, the optimal allocation of replicate measurements deviates from the conventional 50:50 ratio and yields minimum-variance unbiased estimates without requiring weighted regression.

0 citationsRead paper

Infini-News: Efficiently Queryable Access to 1.3 Billion Processed Common Crawl News Articles

May 18, 2026

This study addresses the challenges of acquiring large-scale news corpora, which are often hindered by the high cost of commercial archives and the substantial storage and processing demands of open-source datasets such as CC-News. The authors comprehensively clean and parse metadata from the entire CC-News archive dating back to August 2016, performing language identification on 1.35 billion news articles using GlotLID, lingua, and CommonLingua, and introducing multi-source geolocation annotations that cover 83.4% of the articles across 222 countries. Leveraging suffix arrays, they construct an Infini-gram index enabling sub-second full-text pattern matching for arbitrary queries, thereby substantially lowering the barrier to cross-national, longitudinal media research.

0 citationsRead paper