On the Effectiveness of Hybrid Pooling in Mixup-Based Graph Learning for Language Processing