Mitigating Mention Surface Bias for Entity Disambiguation

Anonymous

Abstract: Entity disambiguation (ED) is a foundational task in NLP for question-answering and information extraction applications. One of the main challenges of ED in real-world settings is the handling of overshadowed entities, i.e., the entities that share mention surfaces with common entities. The current approach for handling overshadowed entities relies on the coherence of entities in a given text, which is not always available and requires additional computing resources. In this paper, we formulated a causal graph for ED and found that the mention surfaces can act as a shortcut, misleading the ED models to be biased towards common entities. We propose a simple yet effective debiasing method that mitigates the effect of the mention surfaces on model predictions. Experimental results demonstrate that our method yields the best results against overshadowed entities.

Paper Type: short

Research Area: Information Extraction

Contribution Types: NLP engineering experiment

Languages Studied: English