Automatic Detection of Parental Interference Behaviors during Bilingual Child Language Assessment
Abstract: Recent clinical research has developed novel protocols that enable children to participate in bilingual language assessment remotely with parents to assist in this process. However, since parents are not trained clinicians, they often perform interference behaviors---actions that could compromise the validity of the assessment (e.g., providing hints). In this paper, we study whether language models can help automate the detection and categorization of parental interference behaviors during bilingual English-Mandarin child language assessment. Such a system would reduce the burden on clinicians, who must otherwise rely on transcribing video recordings and checking them manually for signs of interference. We release a new, expert-annotated dataset for this task, and evaluate multiple state-of-the-art large language models. While these models achieve non-trivial accuracy, they currently lag far behind human annotators. We find that understanding Mandarin and code-mixed text are key challenges these models need to overcome. We hope that our new dataset inspires modeling advances that could improve the practice of bilingual child language assessment.
Paper Type: Short
Research Area: Resources and Evaluation
Research Area Keywords: NLP Applications, Resources and Evaluation
Contribution Types: Data resources, Data analysis
Languages Studied: English, Mandarin
Submission Number: 4466
Loading