Children custody disputes are complicated for many judges owing to so many dynamic factors should to be considered. This paper proposes a novel model for legal expert system Based on LVQ Neural Network. Firstly, all clauses and discretionary factors, extracted from theories and judicial practice, involved with custody disputes in divorce were quantified to obtain related set elements. Then the LVQ neural network was applied to construct a model for children custody disputes weight vector analysis on the set elements. Ultimately, the weight vector close to decision boundaries between classes could promote the classification performance. Accordingly, the performance of the LVQ by generating weight vectors close to decision boundaries is higher. By testing, custody disputes legal expert system could consider all relevant circumstances from a good all-around point of view and effectively.