Improving long-tail relation extraction via adaptive adjustment and causal inference

Extracting long-tail relations poses a significant challenge. Traditional models struggle with weak generalization on tail classes due to the limited sample size. To overcome the limitation, we propose a novel long-tail relation extraction model based on Adaptive Adjustment and Causal Inference (AACI). Specifically, AACI leverages class -adaptive adjustment terms to increase the relative margins between head and tail classes, improving the dis-criminability of tail classes and further enhancing their generalization. Moreover, the learning of our model may encounter multiple spurious correlations due to confounding variables. Therefore, we construct a Structural Causal Model (SCM) for AACI to formalize all spurious correlations and apply causal inference methods to eliminate negative effects of these correlations, thus improving the robustness of AACI. We evaluate our model on the NYT24 and NYT datasets. Our experiments demonstrate that AACI effectively modulates the class margins, eliminates the spurious correlations, and outperforms existing state-of-the-art methods.