Few-Shot Document-Level Relation Extraction with Representation Enhance and Counterfactual Analiysis

Few-shot document-level relation extraction poses a more complex and challenging task compared to sentence-level relation extraction. It also features a more realistic schema, especially in its NOTA (none-of-the-above) distribution. This paper proposes two improvements to address these challenges: entity pair enhancement and bias distillation. We enhance entity pair representation by incorporating the semantic information of relation labels. An attention mechanism is employed to strengthen the relation representation, resulting in more robust feature embedding. To mitigate biases, we introduce bias distillation using counterfactual analysis. Causal counterfactual intervention is leveraged to alleviate entity bias, label bias, and model bias. By adopting this approach, we aim to reduce the influence of biases on the relation extraction task. Our proposed approaches outperform previous benchmarks for the few-shot document-level relation extraction task, as demonstrated on the DocRED dataset.