priscilla-adenuga/low-resource-stress-tests-nlp

Berlin Buzzwords 2026 talk repo on low-resource languages as stress tests for NLP # Low-Resource Languages as Stress Tests for NLP > Research repository accompanying my Berlin Buzzwords 2026 talk on low-resource languages, linguistic structure, and NLP evaluation. --- ## Overview Low-resource languages are often treated primarily as a data scarcity problem. This project takes a different perspective: low-resource languages can also function as diagnostic tools for evaluating the assumptions built into NLP pipelines. By examining underrepresented languages and structurally complex linguistic phenomena, we can expose weaknesses that are often hidden in high-resource benchmark settings. These include issues related to annotation, ambiguity, representation, and evaluation design. --- ## Key Themes - Low-resource NLP - Linguistically informed evaluation - Negation and scope - Structural ambiguity - Multilingual AI evaluation - NLP benchmarking --- ## Main Idea The central argument of this project is that low-resource languages do not only present challenges for NLP systems. They also reveal where current models, datasets, and evaluation practices are too narrow. This perspective is especially relevant for: - multilingual NLP - evaluation of language models - linguistically informed benchmarking - low-resource and underrepresented language settings --- ## Why This Matters Many NLP systems are developed and evaluated on a relatively small set of high-resource languages. As a result, assumptions about segmentation, annotation, grammatical categories, and interpretation can appear natural or universal when they are not. Low-resource languages can help make these assumptions visible. They can reveal: - hidden ambiguity in annotation - mismatches between linguistic structure and model representations - evaluation gaps that surface-level benchmarks do not capture - limits of transferring methods from high-resource to low-resource settings --- ## Examples of Stress-Test Phenomena This repository is especially interested in phenomena such …