This paper uses the IF-IDF and word co-occurrence model to extract and process high-frequency words from the legal compliance text of cross-border dispute arbitration. The LDA theme model then combines with it to extract the theme of legal text compliance, which is then used to construct a cross-border dispute arbitration legal compliance evaluation index system. After that, the genetic algorithm is employed to optimize the BP neural network, construct the GA-BP legal compliance evaluation model, and conduct training simulation. The results show that the word frequency of regional, arbitration, cross-border, fair, consultation, maintenance, reasonable, system, perfect, and conformity is up to more than 3,000 times, which is a high-frequency keyword in the legal text of cross-border dispute arbitration. The output values of the three legal sample cases based on GA-BP are very compliant, compliant, and non-compliant, indicating that the regional arbitration path’s legal compliance in cross-border dispute resolution needs to be improved.