Abstract: Since several decades, the research on the Natural Language Processing (NLP) has been dominated by the symbolic approach.. However, during these last years, an increasing interest is appeared as the connectionist technical applicability for the NLP to converge thus toward the Connectionist Natural Language processing (CNLP). In this paper, we propose a connectionist model for the generation of an internal representation of the sense of a sentence in Arabic language, based on the semantic cases. We use the back propagation algorithm in a Simple Recurrent Network (SRN). The sentence is analyzed word by word. Every word is introduced to the network under semantic features. The task of the network consists of reading the sentence, and deciding the suitable semantic role for each word.. The network successfully learned the case role assignment task. It has been experimented on several corpora.. The network has also been tested on a composed corpus of different sizes of sentences (2, 3, 4 and more) and the rate of generalization approaches of 92% was obtained.