The presented model of language evolution is based on simulating discussions between agents on Barabási-Albert network. As a result of these discussions existing words are propagated in the network, weights representing how well words are known are increased and new words are created. Vocabulary expands due to mutation, crossover and shortening but words can also be forgotten. Through computer simulations several properties of the proposed model are shown, inter alia alteration of average number of known words, power-law distribution of weights along with correlation between dictionaries similarity and agents distance.