Welsh, Yoruba, and Welsh: Why Low-Resource Language NLP Is the EU's Next Frontier

Arabic presents computational linguists with a genuinely difficult problem. The language is morphologically rich in a way that Indo-European languages are not. A single three-letter root such as k-t-b can generate dozens of derived words: kataba (he wrote), kaatib (writer), kutib (it was written), maktaba (library). Each variation carries a distinct meaning derived from the same consonant root combined with different vowel patterns. NLP systems trained predominantly on English, where word forms are comparatively stable, struggle to handle this kind of productive morphology.

Arabic also exhibits significant diglossia. Modern Standard Arabic, used in formal writing, broadcasting, and literature, coexists with regional dialects, including Egyptian, Levantine, and Moroccan, that differ from one another substantially enough to cause comprehension difficulties. Any NLP system deployed across Arabic-speaking populations must accommodate these variations, complicating both model training and real-world deployment.

Then there is the diacritical mark problem. Written Arabic frequently omits the short vowel markers known as tashkeel, forcing NLP systems to infer pronunciation and meaning from context alone. The ambiguity this creates is significant, and it has no direct equivalent in English-language NLP.

Underlying all of these technical challenges is a data scarcity problem. English NLP benefits from billions of words of training data derived from web crawls, digitised books, legal corpora, and academic archives. Comparable Arabic resources remain substantially more limited, and this gap directly restricts the sophistication of Arabic language models. The parallel in European terms is clear: Welsh has roughly 870,000 speakers and a comparatively thin digital footprint; Irish Gaelic has an even smaller corpus of high-quality training data. The mechanism of underservice is the same, even if the scale differs.

Machine Translation: Progress That Exposes Persistent Gaps

Neural machine translation has delivered genuine improvements in Arabic-English translation over the past decade. Modern transformer-based systems can convey meaning across these two structurally very different languages with a fidelity that rule-based and statistical approaches could not approach. Sentence structure, grammatical concepts, and expression conventions differ fundamentally between Arabic and English, and idiomatic or culturally specific content remains difficult. But for routine translation tasks, the quality is acceptable.

Transfer learning has been particularly useful here. Systems trained on well-resourced language pairs such as English-French or English-German can transfer linguistic principles to less-resourced pairs including Arabic-English, partially compensating for limited training data. This is directly relevant to EU language policy. The European Language Grid, maintained under the European Language Resource Association, applies similar logic to lower-resource European languages, and the results demonstrate both the promise and the ceiling of the transfer learning approach when data remains thin.

Anna Rogers, a researcher at the IT University of Copenhagen with extensive work on low-resource NLP and model evaluation, has argued publicly that benchmark-driven development systematically disadvantages languages that lack large standardised test sets. This observation applies with equal force to Arabic and to minority European languages: if you cannot measure performance, you cannot direct investment, and if you cannot direct investment, the gap widens.

Sentiment Analysis, Named Entity Recognition, and the Real-World Stakes

Sentiment analysis in Arabic must accommodate sarcasm, irony, regional dialect variation, and morphological complexity simultaneously. Transformer-based models fine-tuned on Arabic text, including variants of BERT adapted for Arabic such as AraBERT, have improved accuracy substantially. But performance on colloquial dialect content and sarcastic expression remains materially lower than equivalent English-language systems. For businesses operating across Arabic-speaking markets, this means automated customer feedback analysis, social listening tools, and content moderation systems are working with systematically degraded inputs.

Named entity recognition presents a related challenge. Arabic script allows multiple orthographic representations of the same name, and the rapid emergence of new organisations means many entities lack historical precedent in training data. Information extraction from Arabic documents, whether for news analysis, legal review, or regulatory compliance, carries a higher error rate than comparable English-language pipelines as a direct result.

The stakes here are not merely commercial. Content moderation on social platforms that serves Arabic-speaking users relies on NLP systems that are measurably less accurate than those serving English speakers. Hate speech, misinformation, and coordinated inauthentic behaviour are therefore harder to detect in Arabic than in English. This is an equity issue as much as a technical one, and it maps directly onto the EU AI Act's requirements for non-discrimination and transparency in high-risk AI systems.

Speech Recognition and Conversational AI: The Voice Interface Gap

Automatic speech recognition for Arabic faces the phonetic diversity of regional dialects, the challenge of inferring diacritical marks from audio alone, and limited quantities of labelled speech training data. Text-to-speech systems have improved and can now generate Arabic speech with prosody that sounds increasingly natural, but generating pronunciation quality that matches native speaker norms remains an active research problem.

Virtual assistants from major technology companies now support Arabic to varying degrees, but their capabilities lag comparable English-language deployments. Question answering systems are similarly less developed, primarily because Arabic-language question-answering datasets are scarce. Crowdsourcing efforts engaging native speakers in labelling answers to thousands of questions are underway, but progress is slow relative to the need.

Yoshua Bengio, scientific director of Mila and one of the most influential figures in contemporary deep learning, has repeatedly emphasised the importance of building AI systems that serve linguistic minorities equitably, noting in public statements that the concentration of AI research on a handful of high-resource languages creates structural exclusion that compounds over time. His argument applies directly to both Arabic and to Europe's lower-resource languages.

Code-Switching and the Multilingual Reality

Modern Arabic communication, particularly on social media, frequently intermixes Modern Standard Arabic, regional dialects, English, French, and occasionally other languages within a single message. This phenomenon, code-switching, is not unique to Arabic speakers. In Brussels, in Barcelona, in Cardiff, and across migrant communities throughout the EU and UK, everyday digital communication routinely crosses linguistic boundaries in ways that single-language NLP pipelines cannot handle. Developing systems capable of processing code-switched text is an emerging research priority, and the Arabic NLP community's work on this problem has direct methodological relevance for European researchers.

Infrastructure, Investment, and the Path Forward

The Arabic NLP community has built annotated corpora, lexical databases, and evaluation benchmarks that have materially accelerated progress. Specialised corpora for medical, legal, and technical Arabic remain limited, but the infrastructure trajectory is positive. Large language models trained on Arabic text, or multilingually on corpora that include Arabic, are demonstrating impressive capabilities. Distillation techniques, where knowledge from large models transfers to smaller and more efficient counterparts, are opening the possibility of sophisticated Arabic NLP on resource-constrained devices, which matters enormously in regions where high-bandwidth connectivity is not universal.

The EU has begun to take the infrastructure question seriously. The European High Performance Computing Joint Undertaking has supported multilingual model development, and the AI Office established under the EU AI Act has language diversity among its stated concerns. Whether stated concern translates into sustained funding at the scale required remains to be seen. The UK's AI Safety Institute, now operating as the AI Security Institute, has similarly acknowledged low-resource language capability as a dimension of AI safety that deserves attention, though concrete programme commitments remain limited.

The lesson from Arabic NLP is not simply that low-resource languages are hard. It is that the gap between high-resource and low-resource languages widens when the research community optimises for benchmark performance on well-resourced tasks, and narrows only when deliberate investment in data infrastructure, model development, and community building is sustained over years. Europe has the institutional capacity to make that investment. The question is whether the political will exists to prioritise it.