This study examines how Transformer language models utilise lexico-phrasal information to interpret the polysemy of the Korean adverbial postposition -(u)lo. We analysed the attention weights of both a Korean pre-trained BERT model and a fine-tuned version. Results show a general reduction in attention weights following fine-tuning, alongside changes in the lexico-phrasal information used, depending on the specific function of -(u)lo. These findings suggest that, while fine-tuning broadly affects a model’s syntactic sensitivity, it may also alter its capacity to leverage lexico-phrasal features according to the function of the target word.

The present study investigates a computational model's ability to capture monolingual children's language behaviour during comprehension in Korean, an understudied language in the field. Specifically, we test whether and how two neural network architectures (LSTM, GPT-2) cope with a suffixal passive construction involving verbal morphology and required interpretive procedures (i.e., revising the mapping between thematic roles and case markers) driven by that morphology. To this end, we fine-tune our models via patching (i.e., pre-trained model + caregiver input) and hyperparameter adjustments, and measure their binary classification performance on the test sentences used in a behavioural study manifesting scrambling and omission of sentential components to varying degrees. We find that, while these models’ performance converges with the children's response patterns found in the behavioural study to some extent, the models do not faithfully simulate the children's comprehension behaviour pertaining to the suffixal passive, yielding by-model, by-condition, and by-hyperparameter asymmetries. This points to the limits of the neural networks’ capacity to address child language features. The implications of this study invite subsequent inquiries on the extent to which computational models reveal developmental trajectories of child language that have been unveiled through corpus-based or experimental research.

Despite the growing use of NLP in second language (L2) research, model accuracy in L2 settings remains underexplored. This study addresses this gap by evaluating and fine-tuning a Korean language model to extract morphosyntactic features (i.e., morpheme tokenization/tagging and dependency parsing) from L2-Korean texts. We begin by evaluating a domain-general Korean language model on a gold-annotated L2-Korean treebank. We then fine-tune the model on L2-Korean data and quantify the resulting gains across diverse L1- and L2- datasets. Finally, we examine how model reliability varies with learner proficiency scores. Three key findings emerge: while the domain-general model excels at morpheme tokenization, it underperforms on morpheme tagging and dependency parsing; fine-tuning substantially improves adaptability to L2 morphosyntax; and proficiency has minimal effect on morpheme-level tasks but significantly affects dependency-parsing reliability. These results highlight the importance of incorporating L2 training data to improve morphosyntactic analysis in L2 settings and caution against uncritical reliance on automated dependency annotations, especially when performance varies across proficiency levels.

We introduce a manually annotated syntactic treebank based on Universal Dependencies, derived from the written data of second language (L2) Korean learners. In developing this new dataset, we critically evaluated previous works and revised the annotation guidelines to better reflect the linguistic properties of Korean and the characteristics of L2 learners. The L2 Korean treebank encompasses 7,530 sentences (66,982 words; 129,333 morphemes) and is publicly available at: https://github.com/NLPxL2Korean/L2KW-corpus.