[Paper Review] Bipartite Flat-Graph Network for Nested Named Entity Recognition

Ying Luo, Hai Zhao|arXiv (Cornell University)|May 1, 2020

Topic Modeling37 references8 citations

TL;DR

This paper proposes BiFlaG, a bipartite flat-graph network for nested named entity recognition that jointly models outermost entities via a flat NER module and inner nested entities via a graph convolutional network (GCN) module. By enabling bidirectional interaction—where graph-learned inner entity representations improve outermost entity predictions—the model outperforms prior state-of-the-art methods on ACE2005, GENIA, and KBP2017 datasets.

ABSTRACT

In this paper, we propose a novel bipartite flat-graph network (BiFlaG) for nested named entity recognition (NER), which contains two subgraph modules: a flat NER module for outermost entities and a graph module for all the entities located in inner layers. Bidirectional LSTM (BiLSTM) and graph convolutional network (GCN) are adopted to jointly learn flat entities and their inner dependencies. Different from previous models, which only consider the unidirectional delivery of information from innermost layers to outer ones (or outside-to-inside), our model effectively captures the bidirectional interaction between them. We first use the entities recognized by the flat NER module to construct an entity graph, which is fed to the next graph module. The richer representation learned from graph module carries the dependencies of inner entities and can be exploited to improve outermost entity predictions. Experimental results on three standard nested NER datasets demonstrate that our BiFlaG outperforms previous state-of-the-art models.

Motivation & Objective

• Address the limitations of unidirectional error propagation in layered nested NER models.
• Overcome the sparsity of high-level nested entities in deep nesting structures.
• Improve entity boundary and type prediction by modeling bidirectional interactions between outermost and inner nested entities.
• Leverage rich sequence-level representations of outermost entities while capturing hierarchical dependencies through graph-based modeling.
• Develop a generalizable framework that can handle both nested and flat NER tasks without architectural overhaul.

Proposed method

• Use a BiLSTM-CRF model as the flat NER module to identify outermost named entities.
• Construct an entity graph $G^1$ from the outermost entities identified by the flat module to represent potential nested spans.
• Apply a graph convolutional network (GCN) to the entity graph to learn representations of inner nested entities through iterative message passing between span start and end nodes.
• Feed the enriched graph representations back into the flat module to refine outermost entity predictions via bidirectional interaction.
• Use a two-stage framework: first detect outermost entities, then use their structural context to guide inner entity detection through graph propagation.
• Integrate both modules end-to-end with shared token representations and joint optimization to enable joint learning of flat and nested structures.

Experimental results

Research questions

• RQ1 Can bidirectional interaction between outermost and inner nested entities improve nested NER performance compared to unidirectional layered models?
• RQ2 How does incorporating graph-based modeling of inner entities alleviate the sparsity issue in high-level nesting layers?
• RQ3 To what extent does leveraging sequence-level context from outermost entities enhance the detection of nested entities?
• RQ4 Can a bipartite architecture that separates flat and nested entity learning improve generalization and reduce error propagation?
• RQ5 Does the proposed interaction mechanism outperform region-based or layered sequence labeling approaches in terms of precision and recall on nested NER benchmarks?

Key findings

• BiFlaG achieves state-of-the-art performance on three standard nested NER benchmarks: ACE2005, GENIA, and KBP2017 under the same evaluation settings.
• The model significantly reduces error propagation by enabling bidirectional feedback between the flat and graph modules, improving outermost entity detection accuracy.
• The graph module effectively captures dependencies among inner nested entities, especially in deep nesting scenarios where entity sparsity is a challenge.
• Ablation studies show that removing the graph module or disabling inter-module interaction leads to a notable drop in performance, confirming the importance of bidirectional interaction.
• The model successfully identifies complex nested structures such as 'the location in my town that was recently abandoned by Krispy Kreme' with correct boundaries and types, as shown in case studies.
• The framework generalizes well to non-nested cases by simply removing the graph module, maintaining strong performance on flat NER tasks.