AIX Lab at LREC 2026: Two Papers Accepted on Low-Resource Knowledge Graphs and Narrative Graphs Annotation
2026-03-10
We are pleased to share that two papers from the AIX Lab have been accepted to LREC 2026, to be held in Palma de Mallorca, Spain.
The first paper, “Amharic DBpedia Chapter: A Knowledge Graph for a Low-Resource Language”, presents the first steps toward an Amharic DBpedia chapter by extending the DBpedia Extraction Framework with Amharic-specific capabilities. Our technical contributions include a new parser for Ge’ez, a Gregorian–Ethiopian calendar converter, Ge’ez numeral conversion tools, and Amharic-specific template mappings. Since manual mapping is time-consuming, we also explored automation in two directions. For template mapping, we used zero-shot translation with the NLLB-200 model to translate Amharic infobox property names. For ontology alignment, we benchmarked AfroXLM-R Base, XLM-R Base, and mBERT (fine-tuned for Amharic) across 58 DBpedia classes. While zero-shot performance was limited, reformulating the task as natural language inference and fine-tuning the models led to substantial improvements, indicating that automated ontology alignment for Amharic is both feasible and effective even under low-resource conditions. This work is co-authored with colleagues at Leuphana University (Tilahun Abedissa Taffa & Ricardo Usbeck), Google Summer of Code contributors (Meti Bayissa & Andargachew Asfaw), and collaborators from DICE Research (Hizkiel Mitiku Alemayehu, Hamada Zahera, & Axel Ngonga).
Our second accepted paper is titled “From Variance to Invariance: Qualitative Content Analysis for Narrative Graph Annotation”. This work juxtaposes interdisciplinary research methodologies from the social sciences and computer science to annotate news narratives as Directed Acyclic Graphs (DAGs). Additionally, it provides a rigorous annotation evaluation methodology to provide various reliability measures. It provided a Python package for computing Krippendorff’s alpha for graphs. It asks: given the subjective nature of different readings of the same news narrative, how do we create and evaluate narrative graph annotations that capture the text's structural and semantic information and our multiple plausible perceived meanings?
preprint: http://arxiv.org/abs/2603.01930