CausalAPM: NLU 去偏见的通用文字解开（英文版）

英文标题：CausalAPM: Generalizable Literal Disentanglement for NLU Debiasing中文摘要：通过因果推断的角度分析数据集偏差的成因，提出了一种通用的语义与文本面貌解耦的框架 CausalAPM，能够极大地提高模型的泛化性能并同时保持其 ID（in-domain）性能。英文摘要：Dataset bias, i.e., the over-reliance on dataset-specific literal heuristics,is getting increasing attention for its detrimental effect on thegeneralization ability of NLU models. Existing works focus on eliminatingdataset bias by down-weighting problematic data in the training process, whichinduce the omission of valid feature i