With the in-depth use of in-depth learning in the industry , The previous algorithm model based on correlation shows defects , Such as robustness 、 Interpretability 、 Fairness, etc . Recently from Israel Institute of Technology 、 Stanford 、Google Wait for the release of 《 Causal reasoning in natural language processing 》 The paper of , Describe the estimation of causal reasoning in natural language processing 、 forecast 、 Interpretable, etc .
A basic goal of scientific research is to understand causality . However , Although causality plays a key role in life and Social Sciences , But in natural language processing (NLP) There is no equal importance in , The latter traditionally pays more attention to prediction tasks . With the rise of interdisciplinary research on the integration of causal reasoning and language processing , This difference is beginning to disappear . However , About NLP The study of causality in is still scattered in various fields , There is no uniform definition 、 Benchmark data sets and clear statements of remaining challenges .
In this review , We have consolidated cross academic research , And put it in a broader NLP In the landscape . We introduce the statistical challenge of estimating causal effects , Include text as a result 、 The setting of means to treat or solve confusion . Besides , We also explore the potential use of causal reasoning , To improve NLP The performance of the model 、 Robustness 、 Fairness and interpretability . therefore , We provide a unified overview of causal reasoning for the field of computational linguistics .
Address of thesis ：https://arxiv.org/abs/2109.00725