My main research interest is developing cognitive computational models to simulate different aspects of child language acquisition. I am especially interested in verb argument structure acquisition, the development of thematic roles and the underlying semantic structure. Currently, I am working on a computational model of children's verb argument structure acquisition. The model starts by acquiring allowable argument structure frames for each verb individually. General constructions are captured gradually through a classification process. The model receives input from an artificially generated corpus in which an augmented predicate structure representation for each observed scene is paired with an utterance. The corpus is produced on the basis of a manually created lexicon for English which contains the most frequent verbs in mother-child conversations, gathered (along with their most frequent argument structure patterns) from the caretakers' utterance in CHILDES database. The generated corpus reflects the distributional characteristics of the real data. In processing each scene-utterance correspondence, an argument structure frame is extracted and learned for the main predicate term. Each frame contains the thematic roles of the arguments, the semantic primitives associated to the predicate, the semantic categories of the arguments in a WordNet-like semantic hierarchy, and the syntactic pattern used in the utterance. Different frames may be learned for a single verb over time, and some of them may merge together later to form a more general frame. A probabilistic model is proposed for learning abstract knowledge from the instances of scene-utterance correspondence through a classification process. Similar frames of different verbs are incrementally clustered into non-overlapping classes, resulting in the emergence of abstract constructions that form shared semantic components of the underlying frames with the common syntactic patterns used by a group of verbs. The membership degree of each verb in each class is a function of the frequency of the verb appearing in the frame represented by the class. A prediction model combines both lexical and class-level knowledge to predict the best value for a missing piece of syntactic or semantic information, that is, to predict part of the information in a frame based on other available parts. There are two sources of knowledge to count on. First, the lexical knowledge acquired for the main verb, which is embedded in each lexeme in the form of weighted links to a number of classes. Second, the general argument structure constructions learned for the language which are represented by the classes. The more established a class is, the more reliable is the construction it represents. The final decision is made based on a combination of these two sources, with a higher weight given to the lexical knowledge to emphasize what the model has learned about each individual verb. Nonetheless, the abstract knowledge is also considered as an option in case a construction exists in which the verb under investigation has not appeared yet, but matches the current frame the best. The latter is a typical generalization case. The prediction model is used as the core mechanism for a variety of language tasks, including sentence generation, learning fine-grained meaning of verbs in the form of both thematic role assignment and learning additional semantic primitives for each verb, and language comprehension.