Generative Biomedical Entity Linking via Knowledge Base-Guided Pre-training and Synonyms-Aware Fine-tuning

Hongyi YuanZheng YuanSheng Yu

article2022NAACL71 citations

Proposes a generative biomedical entity linking framework that injects synonym and definition knowledge through synthetic pre-training and constrained prefix-tree decoding, achieving state-of-the-art accuracy across multiple benchmarks without candidate selection.

Listen

Biomedical entity linking maps biomedical mentions found in unstructured text to standardized medical concepts. This process is essential for downstream clinical informatics, including disease phenotyping, relation extraction, and automated diagnosis. Traditional approaches rely on dense embedding similarity search, which imposes massive memory requirements to store vector representations for millions of concepts. While generative sequence-to-sequence models reduce memory footprints by generating concept names directly, adapting them to the biomedical domain has been difficult due to the severe scarcity of annotated training data and the prevalence of diverse entity synonyms.

The article demonstrates a generative sequence-to-sequence framework specifically designed for biomedical entity linking. The primary objective is to evaluate whether injecting structured medical knowledge into generative pre-training and fine-tuning can improve linking accuracy without relying on large memory stores or separate candidate retrieval pipelines.

The researchers developed a 406-million-parameter encoder-decoder model initialized from BART-large. They addressed data scarcity by pre-training the model on synthetic natural language sentences built from 2.37 million concepts, definitions, and synonyms in the Unified Medical Language System. For fine-tuning, the approach pairs mentions with their most textually similar synonym using a character 3-gram similarity score and introduces decoder prompt prefixes. During inference, the system constrains output generation across a multi-synonym prefix tree covering all allowable target names, mapping predicted synonyms back to concept identifiers. The framework was evaluated on four standard biomedical and clinical datasets: BC5CDR, NCBI-disease, COMETA, and AskAPatient.

The experiments show that the proposed framework sets new performance benchmarks. First, the fully pre-trained and fine-tuned model achieved state-of-the-art recall@1 accuracy on BC5CDR (93.3%), COMETA (81.4%), and AskAPatient (89.3%), outperforming prior approaches by up to 1.4 percentage points while remaining highly competitive on NCBI-disease (91.9%). Second, knowledge-guided pre-training delivered consistent improvements across all datasets, boosting accuracy by 0.3 to 0.7 percentage points. Third, the method required substantially fewer computational resources for pre-training than existing general-domain generative baselines (6 GPU-days versus 32 GPU-days), while attaining superior transfer capability on medical tasks. Fourth, in sub-population analyses on BC5CDR, the model demonstrated strong zero-shot generalization, surpassing prior data-augmented baselines by nearly 10 percentage points on unseen concepts (86.9% versus 77.5%).

These results demonstrate that generative language models can effectively replace memory-intensive retrieval pipelines for medical entity normalization. By relying on synthetic templates constructed from existing ontologies rather than costly human annotations, the framework significantly lowers the computational and financial barriers required to train and deploy medical text-processing systems. Furthermore, using multi-synonym prefix trees eliminates the operational complexity and failure points of multi-stage candidate selection architectures.

Organizations developing clinical natural language processing pipelines should consider transitioning from dual-encoder similarity models to generative sequence-to-sequence architectures, especially when deploying in memory-constrained environments. For implementation, engineering teams should incorporate synonym-aware fine-tuning and natural language decoder prompts, which proved critical for performance. Future research should evaluate integrating refined candidate generation mechanisms to further boost decoding accuracy and explore scaling to larger foundation model backbones.

Confidence in these findings is high across standard disease, chemical, and colloquial social media benchmarks. However, stakeholders should note two key operational limitations: the generative model showed comparatively lower accuracy on multi-word mentions and mentions that did not directly match knowledge base entries. Additionally, performance remains susceptible to ambiguous or inconsistent entity annotations within the underlying medical knowledge bases.

Tsinghua University knowledge base tuning method biomedical entity linking knowledge base-guided pre-training synonyms-aware fine-tuning

arXiv: 2204.05164

Paper: BioBERT: a pre-trained biomedical language representation model for biomedical text mining, Jinhyuk Lee et al. (2019). Introduces domain-adapted language modeling on PubMed corpora, establishing the foundational biomedical contextual representations utilized and adapted for generative entity linking.
Paper: Domain-Specific Language Model Pretraining for Biomedical Natural Language Processing, Yu Gu et al. (2020). Demonstrates the necessity of domain-specific biomedical pre-training from scratch and provides the standard BLURB benchmarks for biomedical entity extraction.
Paper: ERNIE: Enhanced Language Representation with Informative Entities, Zhengyan Zhang et al. (2019). Pioneers the injection of structured knowledge base facts and entity auto-encoding objectives into pre-trained transformer architectures.
Paper: GLM: General Language Model Pretraining with Autoregressive Blank Infilling, Zhengxiao Du et al. (2021). Establishes autoregressive blank infilling pre-training frameworks that formulate entity generation and name recovery directly within generative language models.

Paper: BioGPT: Generative Pre-trained Transformer for Biomedical Text Generation and Mining, Renqian Luo et al. (2022). Expands on generative biomedical modeling by training a full-scale generative pre-trained transformer specifically targeting end-to-end biomedical information extraction.
Paper: Unifying Large Language Models and Knowledge Graphs: A Roadmap, Shirui Pan et al. (2023). Provides a comprehensive architectural roadmap for integrating structured knowledge bases with generative language models, contextualizing synonym- and KB-guided approaches.
Paper: Generative Knowledge Graph Construction: A Review, Hongbin Ye et al. (2022). Surveys the broader sequence-to-sequence paradigm for generative knowledge graph construction and end-to-end entity generation methodologies.
Paper: Knowledge-Augmented Reasoning Distillation for Small Language Models in Knowledge-Intensive Tasks, Minki Kang et al. (2023). Extends knowledge-intensive generative modeling by distilling knowledge-augmented reasoning into compact language models.

Cover for Generative Biomedical Entity Linking via Knowledge Base-Guided Pre-training and Synonyms-Aware Fine-tuning

Abstract

Entities lie in the heart of biomedical natural language understanding, and the biomedical entity linking (EL) task remains challenging due to the fine-grained and diversiform concept names. Generative methods achieve remarkable performances in general domain EL with less memory usage while requiring expensive pre-training. Previous biomedical EL methods leverage synonyms from knowledge bases (KB) which is not trivial to inject into a generative method. In this work, we use a generative approach to model biomedical EL and propose to inject synonyms knowledge in it. We propose KB-guided pre-training by constructing synthetic samples with synonyms and definitions from KB and require the model to recover concept names. We also propose synonyms-aware fine-tuning to select concept names for training, and propose decoder prompt and multi-synonyms constrained prefix tree for inference. Our method achieves state-of-the-art results on several biomedical EL tasks without candidate selection which displays the effectiveness of proposed pre-training and fine-tuning strategies. The source code is available at Github.com/Yuanhy1997/GenBioEL.

Topics

knowledge base tuning method

Keywords

biomedical entity linking knowledge base-guided pre-training synonyms-aware fine-tuning seq2seq entity linking candidate selection

1 Introduction
2 Approach
2.1 Seq2seq EL
2.2 KB-Guided Pre-training
2.3 Synonyms-Aware Fine-tuning
3 Experiments
3.1 Datasets and KBs
3.2 Implementation Details
3.3 Main Results
4 Discussion
5 Related Work
6 Conclusion
Acknowledgements
References
A Dataset Summary and Statistics
B License and Availability of Resources
C Additional Experiment Results
C.1 Recall@5 Results
C.2 Sub-population Analysis
D Case Study
E Implementation Details
E.1 Knowledge Base Pre-processing
E.2 Pre-training Clause Templates
E.3 Experiment Parameters
E.4 Computational Resource

Knowls

Knowl 1 — Generative biomedical entity linking with synonym-constrained decoding
model/method
The paper formulates biomedical entity linking as sequence generation. Let $E$ be the target-concept set. Each concept $e\in E$ has synonym names $f(e)=\{s_e^i\}_{i=1}^{n_e}$ , and the complete name set is $S=\bigcup_{e\in E}f(e)$ . A name-to-concept mapping $\sigma:S\rightarrow E$ maps every synonym to its concept. For a mention $m$ with left and right contexts $c_l$ and $c_r$ , the encoder receives
$[\mathrm{BOS}]\ c_l\ [\mathrm{ST}]\ m\ [\mathrm{ET}]\ c_r\ [\mathrm{EOS}],$
where $[\mathrm{ST}]$ and $[\mathrm{ET}]$ mark the mention boundaries. The decoder is given the prompt $P_m=\langle m\ \mathrm{is}\rangle$ and generates a target synonym $s=(y_1,\ldots,y_{N_s})$ belonging to the gold concept. The BART-style encoder-decoder is trained to maximize
$p_\theta(s\mid P_m,c_l,c_r,m)=\prod_{i=1}^{N_s}p_\theta(y_i\mid y_{<i},P_m,c_l,c_r,m),$
where $\theta$ denotes model parameters, $N_s$ is the number of tokens in $s$ , and $y_{<i}$ denotes the preceding output tokens. At inference, beam search is constrained by a prefix tree containing all names in $S$ , rather than a candidate set supplied by a separate retriever. The generated name $\hat{s}_m$ is converted to a concept with $\hat{e}_m=\sigma(\hat{s}_m)$ , allowing multiple synonyms to map to one concept and avoiding per-concept representation storage.
Knowl 2 — Knowledge-base-guided synthetic pre-training
model/method
The paper pre-trains the generative linker directly from knowledge-base content to compensate for the small amount of labeled biomedical entity-linking data. For a concept $e$ , two synonym names $s_e^a$ and $s_e^b$ and contextual information $c_e$ are selected; $c_e$ is normally the concept definition and is replaced by another synonym when no definition is available. A manually designed clause template combines $s_e^a$ and $c_e$ into a synthetic encoder discourse, while the decoder is trained to generate $s_e^b$ with the natural-language form $s_e^a$ is $s_e^b$ . Examples of encoder constructions are $s_e^a$ is defined as $c_e$ and $c_e$ describes $s_e^a$ .

The synthetic task therefore asks the model to recover one name of a concept from another name plus a definition or synonym-based context. For concepts with exactly two synonyms, the second synonym also supplies $c_e$ when a definition is absent; for concepts with one synonym, the input synonym, output synonym, and context are the same name. The task has the same encoder-decoder format as entity linking while exposing the model to synonym equivalence and concept definitions. The pre-training source is a UMLS st21pv subset containing 2.37 million concepts, including 160,000 concepts with definitions and 1.11 million concepts with multiple synonyms.
Knowl 3 — Textual-similarity-aware synonym selection for fine-tuning
model/method
For supervised fine-tuning, the paper selects the target synonym that is most textually similar to the input mention instead of fixing one canonical name or sampling a synonym randomly. Given a mention $m$ and the synonym set $f(e_m)$ of its gold concept $e_m$ , the selected target is
$s^*=\arg\max_{s\in f(e_m)}\cos\!\left(\operatorname{TFIDF}_3(m),\operatorname{TFIDF}_3(s)\right),$
where $\operatorname{TFIDF}_3(\cdot)$ is a character 3-gram TF-IDF vector and $\cos(\cdot,\cdot)$ is cosine similarity. The resulting target name is used with the decoder prompt $\langle m\ \mathrm{is}\rangle$ . This choice is motivated by the observed tendency of the generative model to produce names that are textually similar to the mention. The similarity criterion is used only for fine-tuning, not for knowledge-base-guided pre-training, so that pre-training continues to expose the model to diverse synonym pairs.
Knowl 4 — Multi-synonym knowledge-base construction and name normalization
model/method
For each benchmark, the model constructs the inference name set from all synonyms in the original target knowledge base and an expanded synonym set from the 2017 AA Active Release of UMLS. Synonyms are lower-cased and punctuation such as dashes and commas is removed. If the same name is associated with multiple concepts, the name is removed from the concept that has more alternative synonyms, reducing imbalance in synonym counts while retaining a single name-to-concept assignment. The NCBI target knowledge base is the exception: overlapping synonyms are retained to match the evaluation convention used by earlier NCBI systems.

The resulting mapping is generally N-to-1: several generated names can identify one concept. This differs from using only the shortest or a randomly selected name for each concept, which creates a 1-to-1 mapping and discards synonym information during constrained decoding.
Knowl 5 — Biomedical entity-linking evaluation setup
experimental setup
The generative linker is evaluated on NCBI-disease, BC5CDR, COMETA, and AskAPatient (AAP), covering disease, chemical, and colloquial medical mentions. The target vocabularies are MELIC for NCBI-disease, MeSH for BC5CDR, and SNOMED CT for COMETA; AAP uses SNOMED CT and the Australian Medicines Terminology. Abbreviations are expanded with AB3P, text is lower-cased, mention boundaries are marked with $[\mathrm{ST}]$ and $[\mathrm{ET}]$ , and overlapping mentions or mentions absent from the target knowledge base are discarded. Performance is measured with Recall@1 and Recall@5; reported means and standard deviations are computed over three runs for the proposed systems.
Could not parse LaTeX table
The evaluation uses the complete target name sets rather than retrieving a small candidate list before generation.
Knowl 6 — State-of-the-art Recall@1 across four biomedical benchmarks
empirical result
Using all synonym names for constrained decoding, the proposed generative biomedical linker obtains the following Recall@1 results. FT denotes synonyms-aware fine-tuning from the BART-large checkpoint, while PT + FT adds knowledge-base-guided pre-training before fine-tuning. The proposed system does not use candidate selection.
Could not parse LaTeX table
Knowledge-base-guided pre-training improves FT by 0.7 Recall@1 points on BC5CDR, 0.3 on NCBI, 0.7 on COMETA, and 0.5 on AAP. PT + FT exceeds the strongest listed earlier result by 1.4 points on BC5CDR, 1.3 on COMETA, and 0.3 on AAP, while remaining competitive on NCBI.
Knowl 7 — Effect of knowledge-base-guided pre-training
empirical result
An initialization ablation on BC5CDR and COMETA separates the effect of the proposed pre-training task from the effect of fine-tuning. Without fine-tuning, ordinary BART performs poorly because its denoising pre-training objective does not match entity linking; GENRE pre-trained on general-domain entity-linking data performs better; and the proposed KB-guided checkpoint performs best among the three. After synonyms-aware fine-tuning, the KB-guided checkpoint remains strongest.
Could not parse LaTeX table
After fine-tuning, the KB-guided initialization exceeds BART by 0.8 points on BC5CDR and 0.5 on COMETA, and exceeds GENRE by 0.4 and 0.6 points, respectively. The paper reports that the KB-guided pre-training used 6 GPU-days, compared with 32 GPU-days for the GENRE pre-training used in the comparison.
Knowl 8 — Ablation of synonym selection, synonym coverage, and decoder prompts
empirical result
The paper evaluates Recall@1 on BC5CDR and COMETA while varying the fine-tuning target name $s$ , the inference name set $S$ , and the decoder prompt. TF-IDF selects the most similar synonym for fine-tuning; Shortest selects the shortest synonym; Sample selects a randomly sampled synonym. All uses every retained synonym for inference, whereas Shortest and Sample use one name per concept. Prompt Yes uses the decoder prefix $\langle m\ \mathrm{is}\rangle$ .
Could not parse LaTeX table
The results show that selecting a textually similar target is substantially better than selecting a shortest or random target during training. Keeping all synonyms for inference is also better than enforcing a one-name-per-concept mapping. Removing the decoder prompt lowers Recall@1 by 0.3 points on BC5CDR and 0.5 points on COMETA. Within COMETA, accuracy increases as the character 3-gram TF-IDF similarity between a mention and its selected target name increases.
Knowl 9 — Performance across BC5CDR mention and concept sub-populations
empirical result
The proposed system is evaluated on BC5CDR subsets defined by mention form and training exposure: single-word versus multi-word mentions, unseen mentions, unseen concepts, mentions that are not direct synonyms of the target concept, and mentions mapped to the 100 most frequent concepts. FT is fine-tuning without KB-guided pre-training, and PT + FT includes both proposed stages.
Could not parse LaTeX table
KB-guided pre-training improves most subsets, with particularly large gains for unseen concepts: PT + FT reaches 86.9 compared with 77.5 for the full Varma et al. system. Single-word and frequent concepts are easier to resolve than multi-word mentions and unseen concepts. Multi-word performance changes little with pre-training, and the proposed systems underperform the Varma full system on the Not Direct Match subset.
Knowl 10 — Model configuration and training procedure
experimental setup
The implementation uses BART-large with 406 million parameters, a 12-layer Transformer encoder, and a 12-layer Transformer decoder. Both knowledge-base-guided pre-training and synonyms-aware fine-tuning use teacher forcing and label smoothing. The optimizer is Adam with $\epsilon=10^{-8}$ , $\beta=(0.9,0.999)$ , weight decay $0.01$ , attention dropout $0.1$ , gradient clipping at $0.1$ , and label smoothing $0.1$ . Each benchmark is trained three times and evaluated only at the end of training.
Could not parse LaTeX table
Pre-training runs on six A100 GPUs with 40 GB memory using DeepSpeed ZeRO-2 and takes one day of wall-clock time; fine-tuning on each benchmark uses one A100 GPU.

Coverage note — The complete inventory of clause templates, Recall@5 appendix results, and the NCBI inconsistent-label case study are omitted because they provide implementation detail or supplemental analysis beyond the ten load-bearing contribution units.

References

1.Dhruv Agarwal, Rico Angell, Nicholas Monath, and Andrew McCallum. 2021. Entity linking and discovery via arborescence-based supervised clustering.
2.Rico Angell, Nicholas Monath, Sunil Mohan, Nishant Yadav, and Andrew McCallum. 2021. Clustering-based inference for biomedical entity linking. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2598–2608, Online. Association for Computational Linguistics.
3.Marco Basaldella, Fangyu Liu, Ehsan Shareghi, and Nigel Collier. 2020. COMETA: A corpus for medical entity linking in the social media. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3122–3137, Online. Association for Computational Linguistics.
4.Rajarshi Bhowmik, Karl Stratos, and Gerard de Melo. 2021. Fast and effective biomedical entity linking using a dual encoder. arXiv preprint arXiv:2103.05028.
5.Olivier Bodenreider. 2004. The unified medical language system (umls): integrating biomedical terminology. Nucleic acids research, 32:D267–D270.
6.Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2021a. Autoregressive entity retrieval. In International Conference on Learning Representations.
7.Nicola De Cao, Ledell Wu, Kashyap Popat, Mikel Artetxe, Naman Goyal, Mikhail Plekhanov, Luke Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and Fabio Petroni. 2021b. Multilingual autoregressive entity linking. In arXiv pre-print 2103.12528.
8.Allan Peter Davis, Thomas C Wiegers, Michael C Rosenstein, and Carolyn J Mattingly. 2012. Medic: a practical disease vocabulary used at the comparative toxicogenomics database. Database, 2012.
9.Rezarta Islamaj Dogan, Robert Leaman, and Zhiyong Lu. 2014. Ncbi disease corpus: a resource for disease name recognition and concept normalization. Journal of biomedical informatics, 47:1–10.
10.Tuan Lai, Heng Ji, and ChengXiang Zhai. 2021. BERT might be overkill: A tiny but effective biomedical entity linker based on residual convolutional neural networks. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1631–1639, Punta Cana, Dominican Republic. Association for Computational Linguistics.
11.Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online. Association for Computational Linguistics.
12.Jiao Li, Yueping Sun, Robin J Johnson, Daniela Sciaky, Chih-Hsuan Wei, Robert Leaman, Allan Peter Davis, Carolyn J Mattingly, Thomas C Wiegers, and Zhiyong Lu. 2016. Biocreative v cdr task corpus: a resource for chemical disease relation extraction. Database, 2016.
13.Nut Limsopatham and Nigel Collier. 2016. Normalising medical concepts in social media texts by learning semantic representation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1014–1023, Berlin, Germany. Association for Computational Linguistics.
14.Fangyu Liu, Ehsan Shareghi, Zaiqiao Meng, Marco Basaldella, and Nigel Collier. 2021a. Self-alignment pretraining for biomedical entity representations. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4228–4238.
15.Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2021b. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586.
16.Zulfat Miftahutdinov and Elena Tutubalina. 2019. Deep neural models for medical concept normalization in user-generated texts. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pages 393–399, Florence, Italy.
17.Sunil Mohan and Donghui Li. 2019. Medmentions: A large biomedical corpus annotated with {umls} concepts. In Automated Knowledge Base Construction (AKBC).
18.Mark Neumann, Daniel King, Iz Beltagy, and Waleed Ammar. 2019. SciSpaCy: Fast and robust models for biomedical natural language processing. In Proceedings of the 18th BioNLP Workshop and Shared Task, pages 319–327, Florence, Italy. Association for Computational Linguistics.
19.Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. 2020. Zero: Memory optimizations toward training trillion parameter models. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC ’20. IEEE Press.
20.Sunghwan Sohn, Donald C Comeau, Won Kim, and W John Wilbur. 2008. Abbreviation definition identification based on automatic precision estimates. BMC bioinformatics, 9(1):1–10.
21.Mujeen Sung, Hwisang Jeon, Jinhyuk Lee, and Jaewoo Kang. 2020. Biomedical entity representations with synonym marginalization. In ACL.
22.Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pages 3104–3112.
23.Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Shlens, and Zbigniew Wojna. 2016. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826.
24.Shogo Ujiie, Hayate Iso, and Eiji Aramaki. 2021. Biomedical entity linking via contrastive context matching. arXiv preprint arXiv:2106.07583.
25.Maya Varma, Laurel Orr, Sen Wu, Megan Leszczynski, Xiao Ling, and Christopher Ré. 2021. Cross-domain data integration for named entity disambiguation in biomedical text. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4566–4575, Punta Cana, Dominican Republic. Association for Computational Linguistics.
26.Ronald J Williams and David Zipser. 1989. A learning algorithm for continually running fully recurrent neural networks. Neural computation, 1(2):270–280.
27.Ledell Wu, Fabio Petroni, Martin Josifoski, Sebastian Riedel, and Luke Zettlemoyer. 2020. Zero-shot entity linking with dense entity retrieval. In EMNLP.
28.Sheng Yu, Katherine P Liao, Stanley Y Shaw, Vivian S Gainer, Susanne E Churchill, Peter Szolovits, Shawn N Murphy, Isaac S Kohane, and Tianxi Cai. 2015. Toward high-throughput phenotyping: unbiased automated feature extraction and selection from knowledge sources. Journal of the American Medical Informatics Association, 22(5):993–1000.
29.Hongyi Yuan and Sheng Yu. 2021. Efficient symptom inquiring and diagnosis via adaptive alignment of reinforcement learning and classification. arXiv preprint arXiv:2112.00733.
30.Zheng Yuan, Zhengyun Zhao, Haixia Sun, Jiao Li, Fei Wang, and Sheng Yu. 2022. Coder: Knowledge-infused cross-lingual medical term embedding for term normalization. Journal of Biomedical Informatics, page 103983.

Citation

MLA

Yuan, H., et al. “Generative Biomedical Entity Linking via Knowledge Base-Guided Pre-training and Synonyms-Aware Fine-tuning”. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2022, pp. 4038–48, https://doi.org/10.18653/v1/2022.naacl-main.296.

APA

Yuan, H., Yuan, Z., & Yu, S. (2022). Generative Biomedical Entity Linking via Knowledge Base-Guided Pre-training and Synonyms-Aware Fine-tuning. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 4038–4048. https://doi.org/10.18653/v1/2022.naacl-main.296

Chicago

Yuan, H., Z. Yuan, and S. Yu. 2022. “Generative Biomedical Entity Linking via Knowledge Base-Guided Pre-training and Synonyms-Aware Fine-tuning”. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 4038–48. https://doi.org/10.18653/v1/2022.naacl-main.296.

Harvard

Yuan, H., Yuan, Z. and Yu, S. (2022) “Generative Biomedical Entity Linking via Knowledge Base-Guided Pre-training and Synonyms-Aware Fine-tuning”, Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, pp. 4038–4048. Available at: https://doi.org/10.18653/v1/2022.naacl-main.296.

Vancouver

1. Yuan H, Yuan Z, Yu S (2022) Generative Biomedical Entity Linking via Knowledge Base-Guided Pre-training and Synonyms-Aware Fine-tuning. In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, pp 4038–4048

BibTeX

@inproceedings{yuan-etal-2022-generative,
    title = "Generative Biomedical Entity Linking via Knowledge Base-Guided Pre-training and Synonyms-Aware Fine-tuning",
    author = "Yuan, Hongyi  and
      Yuan, Zheng  and
      Yu, Sheng",
    editor = "Carpuat, Marine  and
      de Marneffe, Marie-Catherine  and
      Meza Ruiz, Ivan Vladimir",
    booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jul,
    year = "2022",
    address = "Seattle, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.naacl-main.296/",
    doi = "10.18653/v1/2022.naacl-main.296",
    pages = "4038--4048"
}

Metadata:ACL Anthology

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by/4.0/

Generative Biomedical Entity Linking via Knowledge Base-Guided Pre-training and Synonyms-Aware Fine-tuning

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Generative biomedical entity linking with synonym-constrained decoding

Knowl 2 — Knowledge-base-guided synthetic pre-training

Knowl 3 — Textual-similarity-aware synonym selection for fine-tuning

Knowl 4 — Multi-synonym knowledge-base construction and name normalization

Knowl 5 — Biomedical entity-linking evaluation setup

Knowl 6 — State-of-the-art Recall@1 across four biomedical benchmarks

Knowl 7 — Effect of knowledge-base-guided pre-training

Knowl 8 — Ablation of synonym selection, synonym coverage, and decoder prompts

Knowl 9 — Performance across BC5CDR mention and concept sub-populations

Knowl 10 — Model configuration and training procedure

References

Citation

BibTeX

Access the Paper