Abstract
Collective entity linking (EL) aims to map mentions in text to a knowledge base by considering global entity relationships. This paper presents a Markov Random Field (MRF) model that jointly learns local mention-context and global entity-relatedness features. By utilizing submodular potentials, the authors enable exact inference via min-cut/max-flow algorithms. Additionally, an interactive active learning framework is proposed to reduce human effort in data curation while maintaining high performance.
Introduction
Entity Linking is the task of disambiguating mentions in unstructured text to specific entries in a catalog. While local models treat mentions independently, collective entity linking leverages the fact that entities within a document are typically semantically related. However, joint modeling often faces computational intractability and difficulties in training weights for both local and global features. This work addresses these challenges by framing the problem as an MRF with submodular potentials, allowing for efficient joint training and inference.
Methods
The proposed model represents candidate entities as nodes in an MRF, where edges represent relationships. Each node is assigned a binary label indicating whether it is the correct link.
The authors assume submodular potentials, transforming MAP inference into a min-cut/max-flow problem on an augmented graph, solvable in polynomial time. Weights for local (node) and global (edge) features are learned jointly using a max-margin framework with subgradient descent. To minimize labeling effort, an active learning strategy queries the "least certain" documents for human curation.
Evaluation and Results
The model was evaluated on benchmarks including IITB, AQUAINT, and MSNBC. A significant finding was the discovery of over 8,000 missing or incorrect annotations in standard datasets like IITB_part, emphasizing the need for robust curation tools. Among global features, Inlink and Outlink relatedness proved to be the most effective signals.
On the IITB_part dataset, the proposed model achieved an F1-score of 0.67, significantly outperforming baselines like AIDA (0.07) and TagMe (0.44).
As shown in the evaluation, the active learning framework (AL) showed superior performance over passive learning (PL), reaching higher F1-scores with fewer training documents. Inference was highly efficient, typically taking less than 0.5 seconds per document.
Conclusion
The study demonstrates that joint training of local and global features in a submodular MRF framework provides state-of-the-art results for collective entity linking. While inference is fast, graph construction remains a bottleneck for large documents, sometimes taking up to 1.5 minutes. Future work will focus on optimizing this through pre-computation and parallelization.
Summary updated