International Conference on Linguistic Research and Applications

Objectives
Language is dynamic, meanings converge, diverge, and form evolving semantic fields. In clinical
neuropsychology, however, this variability is typically reduced to fixed categories. Linguistic ability and
impairment are commonly assessed using standard neuropsychiatric instruments such as the Semantic Verbal
Fluency test (SVF), the Phonological Verbal Fluency test (FAS), and the Boston Naming Test (BNT). Most
often, responses on these measures are scored dichotomously as correct or incorrect. This binary scoring obscures
semantically related, approximate, or deviant responses. The objective of this study is to develop and evaluate
a reproducible computational method for continuous, semantically informed scoring of these tests in Swedish.
The primary research question is whether modern vector-based language models can generate stable and
interpretable continuous semantic scores that capture graded variation beyond binary classification.

Methodology
The study applies a computational linguistic framework grounded in distributional semantics, where word
meaning is represented as position in a high-dimensional semantic space. Anonymised, synthetically generated
lexical responses were used to enable controlled methodological development without sensitive data. Text
preprocessing, including normalisation and lemmatisation, was performed using tools from Språkbanken’s text
infrastructure. Responses and target words were represented using Swedish-adapted BERT- based vector
embeddings. BERT (“Bidirectional Encoder Representations from Transformers”) is a transformer-based
language model that learns contextual word representations by analysing large corpora of text and modelling
how words relate to surrounding words in both left and right contexts. In this framework, lexical meaning is
encoded as numerical vectors in a high-dimensional semantic space, where semantically similar words are
positioned closer to one another. This representation enables graded measurement of semantic proximity
rather than categorical judgments of correctness. For the verbal fluency tests (SVF and FAS), semantic
dispersion was also computed to quantify how responses are distributed within the semantic space. In this
context, semantic dispersion denotes the quantitative distribution of response vectors within a highdimensional embedding space, operationalised as the extent to which lexical items diverge from one another in semantic representation.

Results
Vector-based representations generated stable and interpretable continuous scores. The method captured
fine-grained variation among semantically related responses that is lost under binary scoring. Systematic
differences in response structure were observed across the Boston Naming Test (BNT), Semantic Verbal
Fluency (SVF), and Phonological Verbal Fluency (FAS). Linguistic performance could thus be modelled as
movement within a semantic space rather than as a series of discrete outcomes.

Discussion
The study demonstrates the feasibility of continuous semantic scoring for Swedish language assessment. The
proposed method provides a methodological foundation for future clinical validation and contributes to
research on how meaning is structured and dynamically organised in cognitive processes. By reconceptualising
test performance as graded semantic movement, the study advances computational approaches to linguistic
assessment in neuropsychology. Importantly, this framework enables the quantification of latent semantic
structure in a manner that is theoretically grounded, statistically scalable, and reproducible across datasets.
Such an approach may facilitate more sensitive detection of subtle linguistic deviations, potentially improve
early identification of cognitive decline and supporting longitudinal monitoring of semantic change over time.