Abstract:Zero-shot entity linking (EL) aims at aligning entity mentions to unseen entities to challenge the generalization ability. Previous methods largely focus on the candidate retrieval stage and ignore the essential candidate ranking stage, which disambiguates among entities and makes the final linking prediction. In this paper, we propose a read-and-select (ReS) framework by modeling the main components of entity disambiguation, i.e., mention-entity matching and cross-entity comparison. First, for each candidate, the reading module leverages mention context to output mention-aware entity representations, enabling mention-entity matching. Then, in the selecting module, we frame the choice of candidates as a sequence labeling problem, and all candidate representations are fused together to enable cross-entity comparison. Our method achieves the state-of-the-art performance on the established zero-shot EL dataset ZESHEL with a 2.55% micro-average accuracy gain, with no need for laborious multi-phase pre-training used in most of the previous work, showing the effectiveness of both mention-entity and cross-entity interaction.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the candidate entity ranking problem in the zero - shot entity linking (ZEL) task. Specifically, ZEL aims to align entity mentions in a document with unseen entities, thus challenging the generalization ability of the model. Most previous methods have focused on the candidate entity retrieval stage and ignored the candidate entity ranking stage, which is crucial for disambiguation and making final linking predictions.
### Main problems:
1. **Candidate entity ranking**: In the zero - shot entity linking task, how to effectively rank candidate entities to ensure the correct identification of the target entity.
2. **Cross - entity comparison**: How to conduct fine - grained comparisons between candidate entities, especially in cases where there is lexical similarity between candidate entities, to avoid misjudgment.
### Solutions proposed in the paper:
The authors propose a Read - and - Select (ReS) framework to solve the above problems by modeling the main components of entity disambiguation (i.e., mention - entity matching and cross - entity comparison). The ReS framework consists of two modules:
1. **Reading Module**:
- This module uses mention contexts to generate mention - aware entity representations, thereby achieving mention - entity matching.
- For each candidate entity, the reading module obtains a mention - aware entity representation \( P_e \) by adding prefix tokens before the concatenation of the entity description and the mention context and passing it through an encoder.
2. **Selecting Module**:
- This module redefines the candidate entity selection problem as a sequence - labeling problem and fuses the representations of all candidate entities together, thereby achieving cross - entity comparison.
- The model takes the mention, its context, and all mention - aware entity representations as inputs and conducts interactions between candidate entities through the self - attention mechanism, and finally outputs the prediction score for each candidate entity through a classification head.
### Experimental results:
- The ReS framework achieves state - of - the - art performance on the ZESHEL dataset, with a 2.55% improvement in micro - average accuracy.
- ReS does not require cumbersome multi - stage pre - training, demonstrating its strong generalization ability.
- Ablation experiments show that cross - entity interactions help to disambiguate lexically similar entities, especially in the Multiple Categories subset.
### Summary:
This paper proposes a new Read - and - Select (ReS) framework specifically for the candidate entity ranking stage in the zero - shot entity linking task. By explicitly modeling mention - entity matching and cross - entity comparison, ReS significantly improves the accuracy of entity linking and does not require complex multi - stage pre - training, showing its wide applicability and effectiveness in different fields.