Applying machine learning to language problem analysis

During childhood, language development plays a pivotal role, serving as the cornerstone for early learning and social integration while significantly influencing future academic and social accomplishments. Therefore, effective assessment and support of children’s language development are paramount. Among the various assessment tools, language sample analysis stands out as a key method. This involves experts collecting, transcribing, and analyzing children’s language samples to evaluate their language development. The present study endeavors to merge language sample analysis with artificial intelligence (AI) machine learning technology to establish a classification model capable of discerning between children with normal development and those potentially facing language-related challenges. To achieve this goal, the study leverages the ASDBank and CHILDES as data sources, representing children with potential language issues and those with typical development, respectively. Through a series of processes, the corpus undergoes language sample analysis, yielding eight characteristic indicators (MLU-w, MLU-c, MLU5-w, MLU5-c, CTTR-w, CTTR-c, VOCD-D, VOCD-c). Utilizing these indicators as model features, synthetic data technology is employed to address data insufficiency, resulting in the generation of six combined datasets, totaling 1,000 data points, with 854 entries of 91.50% quality. Subsequently, a supervised machine learning classification model is constructed utilizing Bayesian classification, random forest, and gradient boosting machine methods. Considering the appropriate quantity of synthetic data generated and overall model evaluation performance, the random forest model constructed emerges as the optimal classification model, yielding highest accuracy (0.76), sensitivity (0.94), and F1-score (0.85). The study’s finding is that the appropriate quantity of synthetic data positively impacts model performance, with its effect stabilizing after reaching a certain threshold.