Aiming at reducing the woik required for the language potting of spoken language system, a conversational second language acquisition system is proposed. This system need only small lexicon in the initial stage. It need neither hand-description of rules nor the collection/annotation of large corpus. It refer the corpus of semantic frames which is obtained through development/use of first language version of the system. Then, it make hypotheses which leed to reasonable semantic frames and parse the sentence with them. The system drive the back-end system with the interpretation and confirm if the result is suit for the user's will. With above process, the weakly supervised training of the spoken language system is realized.