[Paper Review] Integrating Dictionary Feature into A Deep Learning Model for Disease Named Entity Recognition

Hamada Nayel, Shashrekha H. L|arXiv (Cornell University)|Nov 4, 2019

Topic Modeling5 citations

TL;DR

This paper proposes a deep learning model for disease named entity recognition (Disease-NER) that integrates pre-trained word embeddings, character-level embeddings, dictionary information, and CRF with global scoring. The model achieves state-of-the-art performance on NCBI and BC5CDR datasets, with an F1-score of 85.40% and 79.62%, respectively, outperforming prior methods by leveraging structured linguistic features and sequence labeling optimization.

ABSTRACT

In recent years, Deep Learning (DL) models are becoming important due to their demonstrated success at overcoming complex learning problems. DL models have been applied effectively for different Natural Language Processing (NLP) tasks such as part-of-Speech (PoS) tagging and Machine Translation (MT). Disease Named Entity Recognition (Disease-NER) is a crucial task which aims at extracting disease Named Entities (NEs) from text. In this paper, a DL model for Disease-NER using dictionary information is proposed and evaluated on National Center for Biotechnology Information (NCBI) disease corpus and BC5CDR dataset. Word embeddings trained over general domain texts as well as biomedical texts have been used to represent input to the proposed model. This study also compares two different Segment Representation (SR) schemes, namely IOB2 and IOBES for Disease-NER. The results illustrate that using dictionary information, pre-trained word embeddings, character embeddings and CRF with global score improves the performance of Disease-NER system.

Motivation & Objective

• To improve disease named entity recognition (Disease-NER) performance by integrating external dictionary knowledge into a deep learning framework.
• To evaluate the impact of character-level embeddings and pre-trained word embeddings on Disease-NER performance.
• To compare the effectiveness of IOB2 and IOBES sequence labeling schemes in the context of biomedical named entity recognition.
• To assess the contribution of CRF with global scoring versus local scoring in sequence decoding for Disease-NER.
• To demonstrate that combining multiple signal sources—lexical, orthographic, and dictionary-based—enhances recognition accuracy in complex biomedical text.

Proposed method

• The model employs a bidirectional LSTM architecture to process both word-level and character-level embeddings, capturing morphological and syntactic patterns.
• Pre-trained word embeddings trained on general and biomedical corpora are used as input representations, improving semantic generalization over randomly initialized embeddings.
• Character-level embeddings are learned via a separate bidirectional LSTM, capturing subword information crucial for rare or OOV disease terms.
• Dictionary information is integrated as a feature vector to guide the model in recognizing known disease entities, reducing ambiguity from synonyms and abbreviations.
• Conditional Random Fields (CRF) with global score decoding are applied to model label dependencies and improve sequence-level prediction accuracy.
• Two sequence labeling schemes—IOB2 and IOBES—are evaluated to determine their impact on entity boundary detection and overall F1-score.

Experimental results

Research questions

• RQ1 How does integrating dictionary information affect the performance of a deep learning model in Disease-NER?
• RQ2 What is the relative contribution of character-level embeddings compared to word-level embeddings in recognizing multi-word and rare disease entities?
• RQ3 Does using pre-trained word embeddings lead to better performance than randomly initialized embeddings in the biomedical domain?
• RQ4 Is CRF with global scoring superior to local scoring in sequence labeling for Disease-NER tasks?
• RQ5 Which sequence labeling scheme—IOB2 or IOBES—yields better performance on NCBI and BC5CDR datasets?

Key findings

• The integration of dictionary information significantly improves Disease-NER performance, with the best F1-score of 85.40% on the NCBI dataset.
• Character embeddings contribute substantially to performance, increasing the F1-score by over 10% compared to models without them.
• Pre-trained word embeddings outperform randomly initialized embeddings, demonstrating the value of pre-training on large-scale biomedical and general text.
• CRF with global scoring decoding yields better results than local scoring, especially when combined with other features.
• The IOBES scheme outperforms IOB2 on the BC5CDR dataset by 1.13 percentage points in F1-measure, indicating better handling of entity boundaries.
• The proposed model achieves state-of-the-art performance on both NCBI (85.40% F1) and BC5CDR (79.62% F1) datasets, surpassing previous methods including CNN-based and ensemble models.