@article{yan-etal-2026-masa,
    title = "{MASA}: A Novel Multimodal Foundation Model for {L}2 Speaking Assessment in Picture-description Scenarios",
    author = "Yan, Bi-Cheng  and
      Chao, Fu-An  and
      Lin, Hong-Yun H.Y.  and
      Chen, Berlin",
    editor = "Piperidis, Stelios  and
      Bel, N{\'u}ria  and
      van den Heuvel, Henk  and
      Ide, Nancy  and
      Krek, Simon  and
      Toral, Antonio",
    journal = "International Conference on Language Resources and Evaluation",
    volume = "main",
    month = may,
    year = "2026",
    address = "Palma de Mallorca, Spain",
    publisher = "ELRA Language Resource Association",
    url = "https://preview.aclanthology.org/ingest-lrec/2026.lrec-main.433/",
    pages = "5545--5554",
    abstract = "Automatic speaking assessment (ASA) manages to quantify the language competence of second language (L2) learners by providing a proficiency score based on their spoken responses. Existing efforts typically employ a neural grader coupled with a set of handcrafted features to gauge the competence of language in L2 learners from multiple facets. Despite their decent efficacy, these methods are limited by a laborious feature engineering process and largely overlook the utilization of scoring rubrics that are presented to human raters in speaking assessment. In light of this, we put forward a novel Multimodal foundation model for ASA, termed MASA, for use in picture-description scenarios. Our approach effectively streamlines the feature engineering process by leveraging the pre-trained encoders of a multimodal foundation model, and emulates the nuanced scoring behaviors of human raters by incorporating scoring rubrics directly into the modeling process. Furthermore, a simple, training-free method is introduced to alleviate the scoring bias in MASA by contrasting the output distributions derived from the multimodal and single-modal inputs. A series of experiments conducted on a picture-description task of the General English Proficiency Test (GEPT) dataset validates the feasibility and superiority of our method in comparison to several cutting-edge baselines."
}