Abstract
Nested Named Entity Recognition (NER) presents a significant challenge for traditional sequence tagging models due to overlapping entity boundaries. This paper proposes a novel multi-agent communication framework that facilitates iterative information exchange between a boundary tagger and a type tagger. By allowing these specialized agents to refine their predictions collaboratively, the model achieves state-of-the-art performance on standard benchmarks, particularly excelling in entity recall.
Introduction
Standard NER tasks assume that each token belongs to at most one entity. However, in domains like biomedicine, entities are frequently nested. For instance, in the phrase "IL-6 gene," "IL-6" is a protein nested within the DNA entity "IL-6 gene." Traditional models often fail to capture these layers because they cannot assign multiple labels to the same token. While two-stage models (boundary detection followed by classification) exist, they typically treat these steps as independent. This work argues that boundary and type information are mutually beneficial and should be processed iteratively to refine predictions.
Methodology
The core of the proposed architecture is a Multi-agent Communication Module. The process begins with feature extraction using independent Bi-LSTMs to generate separate representations for boundaries and types.
Iterative Communication
Two agents—a Boundary Tagger and a Type Tagger—collaborate over $l$ iterations:
- Type Tagger: Employs an attention mechanism to "listen" to the boundary tagger's output, allowing it to better distinguish between entity types based on predicted spans.
- Boundary Tagger: Similarly uses attention to leverage type insights, refining boundaries that might be ambiguous without semantic context.
Following these iterations, a Candidate Entity Classifier uses a Bi-LSTM to encode identified regions, and a softmax layer assigns the final entity type. The system is optimized using a joint loss function encompassing initial detection, iterative communication, and final classification.
Experimental Setup
The model was evaluated on two key nested NER benchmarks:
- GENIA: A biomedical dataset characterized by complex nested structures such as proteins, DNA, and RNA.
- GermEval 2014: A German language dataset used for NER tasks.
Results and Discussion
The multi-agent approach consistently outperformed strong baselines, including Seq2Seq and BENSC models, across both datasets.
As shown in the table above, the model achieved an F1-score of 76.5 on GENIA and 81.6 on GermEval 2014. A key insight from the results is the model's superior recall, particularly in boundary detection. By iteratively refining predictions, the model is able to "find" entities that simpler models overlook due to boundary ambiguity.
In a direct comparison of boundary detection performance on the GENIA dataset, the proposed model reached a recall of 78.3, a substantial improvement over the 69.8 recall achieved by the BENSC model. This confirms that the communication mechanism effectively helps locate entities by leveraging type-aware boundary signals.
Conclusion
The proposed multi-agent communication framework demonstrates that iterative feedback between boundary and type tagging significantly enhances nested NER performance. While the iterative nature may introduce additional computational complexity, the gains in recall and overall F1-score on complex datasets like GENIA justify the approach. Future work may explore alternative communication architectures to further optimize this collaborative tagging process.
Summary updated