VLSP 2025 challenge: Vietnamese Semantic Parsing
Main Article Content
Abstract
In 2025, the Eleventh Workshop on Vietnamese Language and Speech Processing (VLSP
2025) introduced its first shared task on Vietnamese Semantic Parsing, known as viSemParse. This
task aims to assess how effectively participating systems can represent the deep semantic structure of
Vietnamese sentences. To support model development and evaluation, the organizers created highquality, task-specific annotated datasets.
The viSemParse 2025 corpus comprises 2,500 Vietnamese sentences, carefully partitioned into
training, public test, and private test splits to support fair and reproducible evaluation. The shared
task was conducted on the AIHub platform, where teams were required to submit predictions on
the public test set before receiving their final ranking based on the hidden private test set, ensuring
robustness against overfitting.
The best-performing system in the viSemParse track achieved a Smatch score of 58%, a result
that highlights not only the inherent complexity of semantic parsing in Vietnamese but also the substantial opportunities for methodological advances and future research in this area.
Keywords: Vietnamese semantic parsing, viSemParse, VLSP 2025