| Identifying language impairment at an early stage of childhood has always drawn attention among researchers in communication disorders. Researchers have proposed several measures for identifying language impairment. This thesis explores the application of natural language processing (NLP) and machine learning techniques for identification of language impairment. To classify children into typically developing or language impaired category, we use language models and various machine learning algorithms. We use features derived from previous work in communication disorder and NLP communities. We evaluate the feasibility of our approach on different speech samples of adolescents kids for story telling and personal narrative task and young children for play sessions.;Our experiment results show that both language models and machine learning algorithms perform better than the baseline method that we use. We identify that morphosyntactic skills, perplexity values from language models, and sentence complexity measures are important features. Support Vector Machines with polynomial kernel along with Bayesian networks are consistently found to provide better results across different tasks of story telling, personal narrative, and play sessions. |
