Forgotten Words: Benchmarking NeoBERT for Dementia Detection in Low-Resource Conversational Filipino and English Speech

The real novelty is the parallel bilingual dataset construction itself. Rather than simply translating existing English clinical speech data, the researchers manually aligned 4,000 transcripts to isolate which performance drops come from language scarcity versus which come from genuine cognitive signal loss. This lets them measure model degradation predictably, not just observe it.

This connects directly to the WhoSaidIt framework from late May, which treated annotation disagreement as a feature and used iterative expert feedback to stabilize labels across multilingual contexts. Both papers share a core insight: quality datasets for underrepresented populations require deliberate design choices, not just scaling existing English pipelines. The dementia work goes further by making the language-domain interaction explicit through controlled parallel data, whereas WhoSaidIt focused on handling inherent ambiguity in demographic inference. Together they signal the field is moving past treating non-English clinical NLP as a translation problem and toward treating it as a distinct validation requirement.

If the same five transformer architectures are benchmarked on a held-out Filipino clinical cohort (not translated, but native speakers) within the next six months, and performance gaps match the predictions from the bilingual dataset, that confirms the method generalizes. If performance diverges significantly, it suggests the translation process itself introduced artifacts that don't reflect real-world code-switching behavior.