[Paper Review] Knowledge Guided Named Entity Recognition for BioMedical Text

Pratyay Banerjee, Kuntal Kumar Pal+2 more|arXiv (Cornell University)|Nov 10, 2019

Topic Modeling36 references4 citations

TL;DR

This paper reformulates biomedical Named Entity Recognition (NER) as a knowledge-guided, multi-answer question answering (KGQA) task, using entity types, definitions, and examples as context to improve entity detection. By training on a large, unified dataset of 18 biomedical corpora and employing a BERT-CNN model with re-contextualization, the approach achieves state-of-the-art F1 scores on 12 of 18 public datasets, outperforming strong baselines including BERT-large fine-tuned on traditional NER.

ABSTRACT

In this work, we formulate the NER task as a multi-answer knowledge guided QA task (KGQA) which helps to predict entities only by assigning B, I and O tags without associating entity types with the tags. We provide different knowledge contexts, such as, entity types, questions, definitions and examples along with the text and train on a combined dataset of 18 biomedical corpora. This formulation (a) enables systems to jointly learn NER specific features from varied NER datasets, (b) can use knowledge-text attention to identify words having higher similarity to provided knowledge, improving performance, (c) reduces system confusion by reducing the prediction classes to B, I, O only, and (d) makes detection of nested entities easier. We perform extensive experiments of this KGQA formulation on 18 biomedical NER datasets, and through experiments we note that knowledge helps in achieving better performance. Our problem formulation is able to achieve state-of-the-art results in 12 datasets.

Motivation & Objective

• To address the challenges of labeling errors, class confusion, and data scarcity in biomedical NER by reformulating the task beyond traditional multi-class classification.
• To reduce system confusion by simplifying prediction to only B, I, O tagging, eliminating the need to jointly predict multiple entity types simultaneously.
• To improve detection of nested and rare entities by integrating external knowledge such as definitions, examples, and entity types as contextual prompts.
• To enable joint learning across 18 diverse biomedical NER datasets through a unified, knowledge-augmented training framework.
• To evaluate transfer learning performance and model generalization across different biomedical domains and entity types.

Proposed method

• Formulates NER as a multi-answer knowledge-guided question-answering task, where each input is a text context paired with knowledge prompts (e.g., 'What is a disease?', 'Define disease: a medical condition that affects health.').
• Introduces a re-contextualization layer called BERT-CNN that refines token representations using local convolutional features to better identify entity boundaries.
• Uses a unified dataset of 398,495 training, 148,166 validation, and 502,306 test samples from 18 biomedical corpora, each enriched with knowledge contexts.
• Employs BERT-base as the backbone model, fine-tuned end-to-end on the KGQA-formulated NER task with BIO tagging.
• Applies knowledge-text attention mechanisms to dynamically weigh the relevance of knowledge context to each token in the input sequence.
• Performs transfer learning by training on source domains and evaluating on target domains to assess cross-domain generalization.

Experimental results

Research questions

• RQ1 Can reformulating NER as a knowledge-guided QA task reduce labeling errors and confusion in biomedical entity recognition?
• RQ2 How does the inclusion of diverse knowledge types—such as definitions, examples, and entity type questions—affect NER performance?
• RQ3 To what extent can a single, jointly trained model generalize across multiple biomedical NER datasets with varying entity types and domain characteristics?
• RQ4 Does the proposed BERT-CNN model with re-contextualization outperform standard BERT-based NER models in terms of exact match F1 and robustness to rare or nested entities?
• RQ5 How effective is the model in zero-shot or few-shot transfer learning across different biomedical domains?

Key findings

• The knowledge-guided QA formulation significantly improves NER performance, achieving state-of-the-art exact match F1 scores on 12 out of 18 publicly available biomedical NER datasets.
• The model achieves a 1.5% to 11% improvement in exact match F1 over strong baselines, including a BERT-large model fine-tuned on a medical corpus using traditional NER.
• Transfer learning experiments show that the model generalizes well across domains, with only a 0.5% to 20% drop in F1 depending on the entity type, and maintains 75% F1 on chemical entities despite high out-of-vocabulary challenges.
• The model achieves 94.43 F1 for 'Problem' entities and 94.85 F1 for 'Test' entities in transfer learning, indicating strong alignment between similar entity types across datasets.
• The use of knowledge-text attention enables better interpretability and improved attention to relevant knowledge, enhancing detection of ambiguous or rare entities.
• The BERT-CNN re-contextualization layer effectively refines token representations, leading to more accurate boundary detection for entities, especially in nested or complex cases.