Automated Scoring of Clinical Expressive Language Evaluation Tasks.

Area of Science:

• Computational Linguistics
• Developmental Psychology
• Speech-Language Pathology

Background:

• Clinical language assessment often involves sentence formulation tasks.
• Evaluating these tasks requires integrating syntactic and semantic knowledge.
• Current methods yield a single composite score for sentence appropriateness.

Purpose of the Study:

• To develop an automated approach for evaluating sentence appropriateness in children.
• To create a dataset of non-clinically elicited sentence formulation responses.
• To enhance training data using neural machine translation for synthetic sentence pairs.

Main Methods:

• Utilized neural machine translation to generate synthetic correct-incorrect sentence pairs.
• Developed a scoring model employing transfer learning for automatic evaluation.
• Compared custom word embeddings with pre-trained contextualized embeddings.

Main Results:

• Transfer learning significantly improved the accuracy of the scoring model.
• Pre-trained contextualized embeddings further enhanced scoring performance.
• The proposed approach offers a scalable method for evaluating sentence formulation.

Conclusions:

• Automated evaluation of sentence formulation is feasible and effective.
• Transfer learning and contextualized embeddings are key to improving accuracy.
• This method can support more efficient and objective language assessment in children.