Introduction
Entity Linking (EL) is a fundamental task in natural language processing that involves identifying text segments (mentions) and mapping them to corresponding entries in a Knowledge Graph (entities). Traditionally, EL is decomposed into two distinct phases: Mention Detection (MD) and Entity Disambiguation (ED).
However, conventional approaches suffer from two primary limitations. First, inter-dependency issues: pipeline methods treat MD and ED as isolated steps, failing to leverage the fact that a successful entity link can provide critical feedback for refining mention boundaries. Second, intra-dependency issues: most models treat mentions within a document as independent tasks, ignoring the contextual relationships between entities (e.g., linking "Jobs" to Steve Jobs significantly increases the likelihood that "Apple" refers to the technology company).
Hierarchical Reinforcement Learning Framework
To address these challenges, this paper proposes a novel Hierarchical Reinforcement Learning (HRL) framework. This approach models EL as a sequential decision-making process, mimicking the cognitive workflow of a human annotator who identifies and links mentions iteratively.
The architecture consists of two levels:
- High-level Policy (The Strategist): Responsible for sequential MD. It selects the next text span to be linked using a Longformer encoder to maintain global document context.
- Low-level Policy (The Executor): Performs ED for the selected mention. It uses a generative approach to autoregressively produce the entity name token-by-token via an LSTM.
The high-level policy is optimized to maximize the total EL reward. If the low-level policy fails to link a mention correctly, the high-level policy receives a penalty, effectively teaching it to identify more "linkable" and accurate mention boundaries. To stabilize training, the authors introduced a discriminator-based value function that estimates the potential success of remaining text spans.
Key Results and Evaluation
The model was evaluated on the AIDA-CoNLL benchmark and several out-of-domain datasets. The HRL approach demonstrated significant improvements over existing pipeline and end-to-end models.
As shown in the results table, the model achieved a new state-of-the-art 87.6 Micro-F1 score on the AIDA-CoNLL test set. Beyond in-domain performance, the HRL framework exhibited superior generalization, achieving a +1.6 improvement in macro-averaged Micro-F1 across four out-of-domain datasets. Ablation studies further confirmed that "Hierarchical Inference" (HI)—making decisions step-by-step—outperforms "Pipeline Inference" (PI), validating the importance of capturing intra-dependencies.
Conclusion and Future Work
By framing Entity Linking as a hierarchical, sequential task, this work successfully captures the complex dependencies between mentions and entities. While the model achieves state-of-the-art accuracy, its sequential nature results in slower inference speeds compared to parallel models. Future research may focus on optimizing this latency or replacing the low-level LSTM with larger pre-trained transformers like BART to further enhance disambiguation capabilities.
Summary updated