Adversarial training for community question answer selection based on multi-scale matching