A self-organizing two-stream model of language comprehension

A self-organizing model of language comprehension at syntactic and semantic levels is presented. It has a two-stream architecture aimed at facilitating language acquisition. A head stream forms phrases that denote entities or events. Such phrases are called extended maximal projections or EMPs. An argument stream attaches an EMP as an argument of another EMP. The learning algorithm within each stream is unsupervised, consistent with the assumption that children can learn language without being corrected. The algorithm is based on distribution analysis of words and phrases, and feature extraction from the obtained distribution. The model was trained using 2000 sentences from a corpus that consists of parents’ inputs to children, and was tested on 48 novel sentences that were randomly selected from the next 500 sentences. The model yielded correct syntactic structures for 44% of the test set and correct semantic memory structures for 38% of the test set. A basic assumption behind the network design was also tested using statistical analysis.