Fig. 3.

Binding of words to semantic roles. (A) After each input sentence, the network is queried with a semantic role label. The readout maps the network state onto a probability distribution of word responses for the queried role. A correct response occurs if the noun is identified that fills the query slot. (B) Feature binding accuracy for sentences with two occurrences of the target word as a function of the amount of language input. One readout identifies the lexical target, and the other readout returns the ordinal position of the target word. Error bars show 95% confidence intervals. (C) Example sentence and its trajectory through state space. Multiple occurrences of the same lexical noun (boy) in different semantic roles (agent, recipient) are separated by history-dependent neuronal processing.