On-Line Error Detection of Annotated Corpus Using Modular Neural Networks

This paper proposes an on-line error detecting method for a manually annotated corpus using min-max modular (M3) neural networks. The basic idea of the method is to use guaranteed convergence of the M3 network to detect errors in learning data. To confirm the effectiveness of the method, a preliminary computer experiment was performed on a small Japanese corpus containing 217 sentences. The results show that the method can not only detect errors within a corpus, but may also discover some kinds of knowledge or rules useful for natural language processing.