From short-term memory to semantics-a computational model

Clinical disorders of language, known as aphasia, cause impaired comprehension of speech in written and spoken forms. This impairment is due to the patient's inability to process semantics that arise from sequence independent co-occurrence of words with content in a short-term memory (STM) of preceding words. If W-i is the immediately forthcoming word in input to the patient, STM, in the context of this disorder, consists of a window, STMWin, that contains the k words that immediately precede W-i. We use a generative approach to model semantics that ensue from the co-occurrence of W-i and STMWin, and view these semantics as the output of a random process with parameters theta. The model uses supervised learning to maximize the likelihood of theta, given labeled content in STMWin. Experimental validation on standard text classification data sets gives an accuracy that is comparable to, or better than, that obtained using support vector machines (SVMs).