[Paper Review] MedType: Improving Medical Entity Linking with Semantic Type Prediction.

Shikhar Vashishth, Rishabh Joshi+3 more|arXiv (Cornell University)|May 1, 2020

Topic Modeling77 references14 citations

TL;DR

MedType improves medical entity linking by predicting semantic types of entity mentions to prune irrelevant candidates, significantly reducing overgeneration. Integrated into five toolkits and pre-trained on large-scale datasets (WikiMed and PubMedDS), MedType consistently boosts performance across benchmarks, demonstrating state-of-the-art gains with publicly released code and data.

ABSTRACT

Medical entity linking is the task of identifying and standardizing medical concepts referred to in an unstructured text. Most of the existing methods adopt a three-step approach of (1) detecting mentions, (2) generating a list of candidate concepts, and finally (3) picking the best concept among them. In this paper, we probe into alleviating the problem of overgeneration of candidate concepts in the candidate generation module, the most under-studied component of medical entity linking. For this, we present MedType, a fully modular system that prunes out irrelevant candidate concepts based on the predicted semantic type of an entity mention. We incorporate MedType into five off-the-shelf toolkits for medical entity linking and demonstrate that it consistently improves entity linking performance across several benchmark datasets. To address the dearth of annotated training data for medical entity linking, we present WikiMed and PubMedDS, two large-scale medical entity linking datasets, and demonstrate that pre-training MedType on these datasets further improves entity linking performance. We make our source code and datasets publicly available for medical entity linking research.

Motivation & Objective

• Address the under-studied problem of candidate overgeneration in medical entity linking, particularly in the candidate generation phase.
• Develop a modular system that leverages predicted semantic types to filter out irrelevant candidate concepts.
• Improve the accuracy and efficiency of existing medical entity linking toolkits through integration with MedType.
• Mitigate the challenge of limited annotated training data by introducing two large-scale medical entity linking datasets: WikiMed and PubMedDS.
• Demonstrate that pre-training MedType on these new datasets further enhances linking performance across diverse benchmarks.

Proposed method

• Design a fully modular system that takes an entity mention and predicts its semantic type (e.g., disease, drug, procedure) using a neural classifier.
• Use the predicted semantic type to filter candidate concepts during the candidate generation phase, excluding those that do not match the predicted type.
• Integrate MedType as a plug-in into five off-the-shelf medical entity linking toolkits to evaluate its generalization and compatibility.
• Pre-train the semantic type classifier on two newly introduced large-scale datasets: WikiMed and PubMedDS, derived from Wikipedia and PubMed.
• Leverage transfer learning by fine-tuning the pre-trained type classifier on downstream entity linking tasks to improve performance.
• Apply end-to-end evaluation on standard medical entity linking benchmarks to measure improvements in F1 and precision-recall metrics.

Experimental results

Research questions

• RQ1 Can semantic type prediction effectively reduce candidate overgeneration in medical entity linking without sacrificing recall?
• RQ2 How much performance gain does MedType provide when integrated into existing medical entity linking toolkits?
• RQ3 To what extent does pre-training on large-scale, weakly supervised datasets like WikiMed and PubMedDS improve MedType’s generalization and linking accuracy?
• RQ4 Does the modular design of MedType allow for consistent performance gains across diverse medical text domains and entity linking systems?
• RQ5 Can MedType mitigate the impact of limited annotated training data in medical entity linking through self-supervised pre-training on large-scale corpora?

Key findings

• MedType consistently improves entity linking performance across five different off-the-shelf toolkits on multiple benchmark datasets.
• The integration of MedType reduces candidate overgeneration by filtering out semantically inconsistent concepts, leading to higher precision in linking results.
• Pre-training the semantic type classifier on WikiMed and PubMedDS significantly boosts performance on downstream entity linking tasks.
• The proposed datasets, WikiMed and PubMedDS, provide large-scale, high-quality training data that support effective pre-training and transfer learning.
• MedType achieves state-of-the-art performance improvements in F1 and precision metrics when evaluated on standard medical entity linking benchmarks.
• The modular design of MedType enables seamless integration into existing pipelines, demonstrating strong generalization across diverse medical NLP systems.