Bi-granularity Adversarial Training for Non-factoid Answer Retrieval

Answer Retrieval is a task of automatically retrieving relevant answers towards a specific question. The recent studies, in this field, have witnessed the vast success of non-factoid QA methods which leverage the large pre-trained models. The findings in intensive experiments have shown that the existing large and deep models can be enhanced by the utilization of adversarial examples, the ones which effectively challenge encoders during training and enable them to generalize well during test. However, the majority of adversarial training methods still suffer from two limitations: 1) they separately take into consideration single-granularity adversarial examples (e.g. character, token or sentence-level examples), resulting in a monotonous mode that easily make encoders get accustomed to such examples, and 2) they fail to actively detect and apply the truly challenging adversarial examples for training. In this paper, we propose a Bi-granularity Adversarial Training (BAT) approach. It not only involves multiple perturbation into the generation of adversarial examples, but selectively utilizes them in terms of perturbation strength. A self-adaptive adversarial training method is developed for recognizing perturbative examples. We conduct experiments on the WikiPassageQA and TREC-QA benchmarks. Experimental results show that BAT substantially improve the answer retrieval performance, reaching the MAP score of about 80.05 % and MRR of 86.27 % for WikiPassageQA, MAP of 93.99 % and MRR of 97.55 % for TREC-QA.