Improving the robustness and accuracy of biomedical language models through adversarial training