Salvato in:
Dettagli Bibliografici
Autori principali: Hoogland, Damar, Koloski, Boshko, Caporusso, Jaya, Kolenik, Tine, Vitez, Ana Zwitter, Pollak, Senja, Manouilidou, Christina, Purver, Matthew
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:https://arxiv.org/abs/2604.06758
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866917391411183616
author Hoogland, Damar
Koloski, Boshko
Caporusso, Jaya
Kolenik, Tine
Vitez, Ana Zwitter
Pollak, Senja
Manouilidou, Christina
Purver, Matthew
author_facet Hoogland, Damar
Koloski, Boshko
Caporusso, Jaya
Kolenik, Tine
Vitez, Ana Zwitter
Pollak, Senja
Manouilidou, Christina
Purver, Matthew
contents We evaluate cognitive impairment (CI) classification from transcripts of speech in English, Slovene, and Korean. We compare zero-shot large language models (LLMs) used as direct classifiers under three input settings -- transcript-only, linguistic-features-only, and combined -- with supervised tabular approaches trained under a leave-one-out protocol. The tabular models operate on engineered linguistic features, transcript embeddings, and early or late fusion of both modalities. Across languages, zero-shot LLMs provide competitive no-training baselines, but supervised tabular models generally perform better, particularly when engineered linguistic features are included and combined with embeddings. Few-shot experiments focusing on embeddings indicate that the value of limited supervision is language-dependent, with some languages benefiting substantially from additional labelled examples while others remain constrained without richer feature representations. Overall, the results suggest that, in small-data CI detection, structured linguistic signals and simple fusion-based classifiers remain strong and reliable signals.
format Preprint
id arxiv_https___arxiv_org_abs_2604_06758
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Multilingual Cognitive Impairment Detection in the Era of Foundation Models
Hoogland, Damar
Koloski, Boshko
Caporusso, Jaya
Kolenik, Tine
Vitez, Ana Zwitter
Pollak, Senja
Manouilidou, Christina
Purver, Matthew
Computation and Language
We evaluate cognitive impairment (CI) classification from transcripts of speech in English, Slovene, and Korean. We compare zero-shot large language models (LLMs) used as direct classifiers under three input settings -- transcript-only, linguistic-features-only, and combined -- with supervised tabular approaches trained under a leave-one-out protocol. The tabular models operate on engineered linguistic features, transcript embeddings, and early or late fusion of both modalities. Across languages, zero-shot LLMs provide competitive no-training baselines, but supervised tabular models generally perform better, particularly when engineered linguistic features are included and combined with embeddings. Few-shot experiments focusing on embeddings indicate that the value of limited supervision is language-dependent, with some languages benefiting substantially from additional labelled examples while others remain constrained without richer feature representations. Overall, the results suggest that, in small-data CI detection, structured linguistic signals and simple fusion-based classifiers remain strong and reliable signals.
title Multilingual Cognitive Impairment Detection in the Era of Foundation Models
topic Computation and Language
url https://arxiv.org/abs/2604.06758