Decoding Morphosyntax: Can LLMs Handle Inflection and Derivation in English and Greek?
Abstract
This study explores the extent to which Large Language Models (LLMs) can accurately analyze and compare
inflectional morphosyntactic features and derivational patterns in English and Greek — two typologically
divergent systems. While English is predominantly analytic, Greek exhibits a rich fusional inventory of
morphosyntactic features encoded inflectionally, including tense, aspect, number, case, gender, and person.
The central research question is whether LLMs can reliably identify, categorise, and disambiguate these
features across the two systems and how their performance on inflectional paradigms interacts with their
handling of derivational morphology. A secondary focus concerns the degree to which morphosyntactic
feature encoding constrains or assists LLMs in recognizing derivational processes and their productivity.
The study hypothesizes that the typological mismatch between English and Greek exposes systematic gaps
in LLM morphological competence, particularly in the processing of inflectionally dense paradigms and
derivationally complex lexemes.
Methodology
The study adopts a mixed-methods design integrating theoretical morphological analysis with
computational modeling and empirical evaluation. Two annotated corpora — one for English and one for
Greek — are constructed from diverse text sources, with words tagged for morphosyntactic feature values,
derivational patterns (prefixation, suffixation), and morphological complexity ranging from transparent to
opaque forms. Complex phenomena receiving special attention include syncretism, allomorphy, suppletion,
and morphosemantic ambiguity — all of which pose well- documented challenges for both human parsers
and computational models. State-of-the-art LLMs (GPT-5.1, Gemini 3, Claude 4.6, and Perplexity 4.5) are
evaluated on the annotated datasets using standard metrics (precision, recall, F1-score), complemented by
novel morphology-sensitive metrics developed specifically for LLM morphological evaluation: a
Morphosyntactic Context Sensitivity metric, a Morphological Complexity Score, and a Morpheme Accuracy
Metric, among others. Supervised fine-tuning and Reinforcement Learning from Human Feedback (RLHF)
via prompt engineering are further explored as strategies for improving model performance.
Results
Evaluation results reveal that LLMs perform inconsistently across morphosyntactic feature categories, with
the greatest difficulties emerging in Greek paradigms characterized by high degrees of syncretism and
morphophonological alternation (e.g., inactive phonological phenomena, allomorphy). In derivational
analysis, models tend to rely on surface analogical patterns rather than rule-governed morphological
operations, leading to systematic errors in disambiguating derivation from inflection and in correctly
identifying the base and affix structure of complex words. Cross- linguistic comparison further confirms
that morphosyntactic typology significantly affects LLM generalization, with Greek consistently yielding
lower accuracy scores than English across all evaluation metrics.
Conclusions
The findings demonstrate that current LLMs lack robust morphosyntactic feature representations and that
their handling of inflection and derivation falls short of linguistically informed analysis. Crucially, the study
shows that targeted fine-tuning on morphologically annotated data — particularly for feature-rich languages
like Greek — can meaningfully improve performance. These results have direct implications for the designof NLP tools in morphologically complex languages and call for evaluation frameworks that foreground
morphosyntactic adequacy rather than surface fluency alone.
