[Paper Review] Neural Entity Reasoner for Global Consistency in NER

Xiaoxiao Yin, Daqi Zheng+2 more|arXiv (Cornell University)|Sep 30, 2018

Topic Modeling22 references8 citations

TL;DR

This paper proposes Neural Entity Reasoner (NE-Reasoner), a multi-layered neural network framework that enhances global consistency in Named Entity Recognition (NER) by enabling cross-sentence reasoning through a symbolic candidate pool and a reasoning mechanism that refines predictions across layers. It achieves state-of-the-art performance on the CoNLL-2003 English NER benchmark and improves recall on Chinese court judgment NER by leveraging inter-sentence entity consistency.

ABSTRACT

We propose Neural Entity Reasoner (NE-Reasoner), a framework to introduce global consistency of recognized entities into Neural Reasoner over Named Entity Recognition (NER) task. Given an input sentence, the NE-Reasoner layer can infer over multiple entities to increase the global consistency of output labels, which then be transfered into entities for the input of next layer. NE-Reasoner inherits and develops some features from Neural Reasoner 1) a symbolic memory, allowing it to exchange entities between layers. 2) the specific interaction-pooling mechanism, allowing it to connect each local word to multiple global entities, and 3) the deep architecture, allowing it to bootstrap the recognized entity set from coarse to fine. Like human beings, NE-Reasoner is able to accommodate ambiguous words and Name Entities that rarely or never met before. Despite the symbolic information the model introduced, NE-Reasoner can still be trained effectively in an end-to-end manner via parameter sharing strategy. NE-Reasoner can outperform conventional NER models in most cases on both English and Chinese NER datasets. For example, it achieves state-of-art on CoNLL-2003 English NER dataset.

Motivation & Objective

• Address the limitation of local decision-making in sequence labeling models that fail to maintain global consistency across entities in a document.
• Enable neural models to reason over multiple entities across sentences, mimicking human reading behavior that integrates context from the entire document.
• Introduce a symbolic memory (candidate pool) to store and retrieve entity-level information across layers for improved inference.
• Develop an end-to-end trainable framework that combines neural learning with symbolic reasoning to improve generalization on rare or ambiguous entities.
• Achieve better performance on both English and Chinese NER benchmarks by refining predictions through inter-sentence consistency.

Proposed method

• Employ a multi-layered architecture where each layer performs NER independently, with results passed to a shared candidate pool.
• Use a symbolic memory (candidate pool) to store recognized entities from previous layers, enabling cross-layer reference.
• Implement an interaction-pooling mechanism that connects local word representations to global entities via similarity scoring.
• Apply a reasoning module that computes similarity scores between local predictions and candidate entities to refine decisions in deeper layers.
• Use parameter sharing to enable end-to-end training while maintaining symbolic reasoning capabilities.
• Leverage a deep architecture to progressively refine entity sets from coarse to fine-grained recognition.

Experimental results

Research questions

• RQ1 Can a neural model effectively maintain global consistency in NER by reasoning over multiple entities across sentences?
• RQ2 How does incorporating a symbolic candidate pool improve the recognition of ambiguous or rare named entities?
• RQ3 To what extent can a multi-layered reasoning mechanism outperform single-pass sequence labeling models in NER tasks?
• RQ4 Does the proposed reasoning mechanism enhance performance on low-resource or out-of-vocabulary entities?
• RQ5 Can the framework generalize across languages and domains, such as in Chinese legal text with complex entity patterns?

Key findings

• NE-Reasoner achieves state-of-the-art F1 score on the CoNLL-2003 English NER dataset, demonstrating superior global consistency.
• On the Chinese court judgment NER dataset, the model improves recall from 96.13% to 97.26% while maintaining high precision, showing enhanced detection of missed entities.
• The model successfully recognizes rare or previously unseen names (e.g., 'Guang' as a surname) in later layers by leveraging context from previously identified entities.
• The reasoning mechanism corrects misclassifications—such as changing 'Salem Bitar' from organization to person—by comparing similarity scores to known entities.
• Ablation studies confirm that the reasoning module is responsible for improvements, as results remain unchanged when the name or similarity score is manually altered.
• The model reduces false negatives by using inter-sentence context, especially for entities mentioned only once but with strong contextual similarity to known entities.