Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Weakly Supervised Medical Entity Extraction and Linking for Chief Complaints

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Chief complaints can be linked to a medical ontology with no human-annotated training data.

desk verdict A credible weak-supervision pipeline for chief complaint NER/linking with an honest evaluation, but the headline claim does not hold for the ~40% of records without punctuation separators. read the letter →

arxiv 2509.01899 v1 pith:X2ZBKXNX submitted 2025-09-02 cs.CL

Zhimeng Luo , Zhendong Wang , Rui Meng , Diyang Xue , Adam Frisch , Daqing He This is my paper
classification cs.CL
keywords weaksupervisionchiefcomplaintsnamedentityrecognitionlinkingmedicalontologyBERTlabelsmoothingemergencydepartmentrecords
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that chief complaint records—patients' own words for why they are seeking emergency care—can be automatically mined for medical entities and mapped to a standardized ontology with no human-annotated training data. The authors generate weak labels by splitting records on punctuation and matching the resulting chunks to ontology concepts through exact string, approximate dictionary, and embedding-based matching. A BERT-based extraction model and a BiLSTM linking model are then trained on these noisy labels, with label smoothing and a punctuation-deprivation trick to absorb the noise. On 1.2 million real emergency records with a physician-annotated test set, the resulting pipeline outperforms dictionary matching and a pretrained general medical linking system, although it trails fully supervised training. The significance is practical: if the claim holds, hospitals using different free-text entry styles could standardize chief complaints without paying for expert annotation.

What carries the argument

Split-and-match weak label generation: records are split on ten punctuation marks, chunks are matched to ontology concepts in three stages (exact string match against HaPPy synonyms; QuickUMLS approximate dictionary matching; fastText cosine-similarity matching), and matched chunks become weak entity spans with concept labels. The extraction model CCME-BERT applies confidence-weighted label smoothing to tolerate the noise in these labels; the linking model CCEL concatenates word and character BiLSTM embeddings of the mention with its two-token context to predict the concept. The denoising trick, training on records with punctuation removed, is the mechanism that extends the approach to the r

What would settle it

Take the model trained only on weak labels from punctuation-split records, without the punctuation-removal denoising, and evaluate it on a held-out set of chief complaints that contain no punctuation; if its exact-match F1 stays near the reported 9.91 rather than approaching the supervised level, then the no-annotation claim fails for a large fraction of real records. Alternatively, count on the released 1.2-million-record corpus how many records yield at least one weak label; a rate substantially below 100% would bound the pipeline's coverage.

Watch

Extended reading notes

Core claim

The central claim is that WESEEL (weakly supervised entity extraction and linking) achieves superior performance over previous methods without any human annotation. The pipeline first uses a split-and-match algorithm: each record is segmented by ten punctuation separators, each chunk is matched to the HaPPy ontology via exact string match, QuickUMLS approximate matching, and fastText embedding similarity; successful matches become weak span and concept labels. A BERT sequence-labeling model (CCME) is trained on these weak labels for mention extraction with confidence-weighted label smoothing, and a BiLSTM (CCEL) with word and character embeddings and left/right context encodes mentions for c

Load-bearing premise

The load-bearing premise is that the split-and-match algorithm's punctuation-dependent weak labels are of sufficient quality and coverage to train a model that generalizes to the roughly 40% of chief complaint records that contain no punctuation at all.

Editorial extensions

If this is right

  • Chief complaint processing can be framed as extraction plus linking rather than multi-label classification, giving interpretable entity spans and ontology mappings.
  • A dictionary plus a small set of separator heuristics can serve as the supervision signal for a strong neural model, reducing annotation cost to near zero for institutions that adopt an ontology.
  • Confidence-based label smoothing is an effective counter to noisy weak labels; it improves F1 over plain training on the same weak data.
  • Training with deliberately removed punctuation improves performance on no-punctuation records, suggesting that teaching the model to infer separators is a viable route to robustness.
  • Domain-specific general-purpose medical linking systems (MedType) underperform on chief complaints, confirming that chief complaint NLP needs either fine-tuning or task-specific weak supervision.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same split-and-match recipe should transfer to other short clinical free-text fields (triage notes, patient portal messages) where an ontology exists but curated annotations do not; the necessary condition would be the presence of delimiter conventions.
  • A testable extension: rather than relying on punctuation presence, a learned or rule-based segmenter that uses ontology terms as a lexicon could recover weak labels for the 40% no-punctuation records more effectively than the denoising trick, potentially pushing the no-punctuation F1 from 54.18 toward the supervised 83.29.
  • The fastText matching stage trades precision for recall and is the main source of false positives; replacing it with a more precise fuzzy matcher or a learned mention-classifier could improve weak label quality without annotations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes WESEEL, a weakly supervised pipeline for entity extraction and linking in emergency-department chief complaints. Weak labels are generated by splitting records on manually chosen punctuation separators and matching chunks to the HaPPy ontology via exact string match, QuickUMLS, and fastText embedding similarity. A BERT-based sequence tagger (CCME-BERT) is trained on these weak labels with label smoothing, and a BiLSTM linking model (CCEL) maps mention spans to ontology concepts. Experiments use 1.2 million real UPMC chief complaint records and a clinician-annotated test set of 1,013 records. The main reported results are extraction F1 of 67.51 (partial match) for CCME-BERT (soft) and linking F1 of 58.88 for CCEL. An ablation shows that weak supervision plus supervised fine-tuning is beneficial. However, Section 5.4 states that nearly 40% of records contain no punctuation separators, and Table 4 shows that the model performs poorly on this subset unless a separate denoising variant is used; the main tables do not clearly indicate whether that denoising variant is part of the reported pipeline.

Significance. If the central claim holds, WESEEL would be a practical means of building chief-complaint NER and linking systems without clinician annotation, with clear value for record standardization and syndromic surveillance. The paper's strengths include a large real-world corpus, an independent clinician-annotated test set with high inter-annotator agreement, multiple matching and neural baselines, and a careful analysis of label smoothing under weak supervision. The evaluation is not circular, since the test set was annotated by clinicians independently of the weak-label generator. The main fragility is the dependence of the weak-label generator on punctuation separators: for the large no-punctuation portion of the data, the method's performance is far below the headline numbers unless a denoising trick is added, and even then the gap remains substantial. The claimed superiority also lacks statistical support, and the reported anonymous repository makes replication currently impossible.

major comments (3)
  1. [§5.4, Tables 1 and 4] The central claim that WESEEL achieves superior performance 'without any human annotation' is not established for the ~40% of records without punctuation separators. As Section 5.4 states, the split-and-match algorithm cannot generate weak labels for these records. Table 4 shows that a model trained on punctuation-derived labels obtains F1 = 25.98 (Partial) on the no-punctuation test subset, rising to 54.18 only when punctuation is deliberately removed from training instances as a denoising strategy. This remains far below the overall extraction F1 of 67.51 in Table 1. Critically, the main experiments in Tables 1 and 2 do not state whether the denoising variant is included. If it is part of WESEEL, all comparisons and ablations must be rerun with it; if not, the reported 'superior performance' applies only to the punctuation-rich majority. Please report results for the no-punctuation sub
  2. [§4.3 and §5.1/5.2] No statistical significance tests or confidence intervals are reported. This is load-bearing because several headline differences are small. For example, in Table 1 CCME-BERT (soft) achieves F1 67.51 vs 67.32 for CCME-ClinicalBERT (soft) and 66.23 for CCME-PubMedBERT (soft); in Table 2, CCEL (58.88) is close to BERT-based Ranking (59.55). Without paired significance tests (e.g., bootstrap or McNemar) on the 1,013 test instances, the claim of superiority over the strongest baselines is not statistically grounded.
  3. [§4.1 (data availability footnote)] The paper promises that 'All code/data for reproducing the results will be released at https://github.com/anonymous_repo, under IRB restrictions.' At review time no code, data, or detailed hyperparameter configuration is available. Since the central empirical claim depends on a proprietary corpus, the authors should provide a detailed data-processing specification, pseudocode for the split-and-match algorithm, the exact hyperparameters for all models, and at least a sample of de-identified records to enable replication and to allow reviewers to assess the weak-label generation process.
minor comments (5)
  1. [§5.4, Table 4] The row labels 'w/ punct', 'w/o punct', and 'w/ punct + denoising' are ambiguous. Please specify that these refer to training-data conditions and that evaluation is always on the no-punctuation test subset.
  2. [§3.2] The acronym CCME is introduced for the mention-extraction model, but all experiments use CCME-BERT and variants. Define the relationship between CCME and CCME-BERT (soft) explicitly.
  3. [§5.2] The sentence 'The performance of CCEL drops from 58.88 to 51.04 (F1) after removing the context embedding part or the character embedding part' is ambiguous. Please report the two ablations separately, since removing context vs. character embeddings are different modifications.
  4. [§4.2] Please list the exact hyperparameter values (learning rate, batch size, epochs, label-smoothing coefficient, context window) rather than referring to 'default settings of HuggingFace' and 'the setting of (Yang et al., 2018)', which is not sufficiently reproducible.
  5. [§1 and §3] There are minor typos: 'chief complains' should be 'chief complaints', and the acronym HaPPy should be expanded at first use. Also, Figure 2 is not referenced in the main text where the split-and-match pipeline is described.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: WESEEL is a standard weak-supervision pipeline evaluated on an independent human-annotated test set.

full rationale

The paper's derivation chain is empirical rather than definitional. Split-and-match generates weak labels using the HaPPy ontology, a BERT-based extractor and a BiLSTM linker are trained on those labels, and all headline results are measured against 1,013 chief complaint records independently annotated by two clinicians. The paper explicitly states: 'There were no weak labels used in the process of human annotation.' Thus the test set is not produced by the label generator, and the reported F1 scores are not identities forced by construction. The S&M matching baselines in Tables 1 and 2 are components of the weak-label generator, but the neural models are trained to denoise and generalize beyond those matching outputs; for example, CCME-BERT (soft) achieves partial-match extraction F1 67.51 versus the best S&M pipeline's 62.91. These are empirical comparisons, not circular reductions. There are no load-bearing self-citations: the references cite external prior work, and no uniqueness theorem or ansatz is imported from the authors' own previous publications. The §5.4 limitation is real but not circular: nearly 40% of records lack punctuation separators, and Table 4 shows weak generalization (F1 25.98 without denoising, 54.18 with denoising) on that subset. This threatens the breadth of the central claim of universal weak supervision, but it does not make the evaluation circular because the evaluation still uses independent human annotations. The central claim is therefore not equivalent to its inputs by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the weak label generation pipeline and the HaPPy ontology. The main free parameters are the manually chosen separators, window sizes, embedding sizes, and matching thresholds. The key domain assumptions are the validity of HaPPy and the representativeness of the test set. No new entities are invented.

free parameters (5)
  • separator punctuation set = 10 manually selected marks
    Section 3.0; choice of which punctuation marks split records affects the chunking and thus all weak labels.
  • context window size = 2 tokens
    Section 4.2; window of tokens before and after mention for linking model.
  • word/character embedding sizes = 100/30
    Section 4.2; sizes of fastText word embeddings and character embeddings.
  • label smoothing parameter = not specified
    Section 3.2; label smoothing is used but the exact value is not reported, affecting training.
  • matching similarity thresholds = not specified
    Section 3.1; thresholds for QuickUMLS and fastText cosine similarity are not given, directly controlling weak label precision/recall.
assumptions (4)
  • domain assumption HaPPy ontology is the correct target ontology for chief complaints
    The linked concepts are defined by HaPPy; the paper reduces it to 501 concepts assuming no clinically significant information loss.
  • domain assumption The human-annotated test set is reliable and representative
    Two domain experts annotated 1,013 records; Kappa is high, but the sample may not cover all concept distributions.
  • ad hoc to paper Weak labels from split-and-match are sufficiently accurate for training
    The method relies on these labels; the paper shows that without them, performance drops, but the noise in the labels is not fully quantified.
  • domain assumption Punctuation-based chunking captures most multi-entity records
    Section 3.0; 55.7% of labeled records contain multiple entities, but 40% of all records lack punctuation, so chunking fails on a large subset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Weakly Supervised Medical Entity Extraction and Linking for Chief Complaints." pith.science (2026). https://pith.science/paper/X2ZBKXNX

@misc{pith2026250901899,
  author       = {Pith},
  title        = {Pith review of: Weakly Supervised Medical Entity Extraction and Linking for Chief Complaints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X2ZBKXNX}},
  note         = {Machine review of arXiv:2509.01899}
}
read the original abstract

A Chief complaint (CC) is the reason for the medical visit as stated in the patient's own words. It helps medical professionals to quickly understand a patient's situation, and also serves as a short summary for medical text mining. However, chief complaint records often take a variety of entering methods, resulting in a wide variation of medical notations, which makes it difficult to standardize across different medical institutions for record keeping or text mining. In this study, we propose a weakly supervised method to automatically extract and link entities in chief complaints in the absence of human annotation. We first adopt a split-and-match algorithm to produce weak annotations, including entity mention spans and class labels, on 1.2 million real-world de-identified and IRB approved chief complaint records. Then we train a BERT-based model with generated weak labels to locate entity mentions in chief complaint text and link them to a pre-defined ontology. We conducted extensive experiments, and the results showed that our Weakly Supervised Entity Extraction and Linking (\ours) method produced superior performance over previous methods without any human annotation.

Figures

Figures reproduced from arXiv: 2509.01899 by the authors.

Figure 1
Figure 1. Examples of chief complaint records and corresponding concepts in HaPPy ontology. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our proposed method WESEEL. 200,000 patient-generated reasons-for-visit entries and mapped them to 795 unique chief complaint concepts. Teng et al. (2020) formulated automatic diagnose code assignment as multi-label classifica￾tion to predict ICD codes from free-text medical records including chief complaints. Most datasets in these studies are not publicly available. 2.2 Medical Entity Extraction and Li… view at source ↗
Figure 3
Figure 3. Architecture of the model for entity linking in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 37 canonical work pages

  1. [1]

    Alan R Aronson. 2001. Effective mapping of biomedical text to the umls metathesaurus: the metamap program. In Proceedings of the AMIA Symposium, page 17. American Medical Informatics Association

    work page 2001

  2. [2]

    Parminder Bhatia, Busra Celikkaya, Mohammed Khalilia, and Selvan Senthivel. 2019. Comprehend medical: a named entity recognition and relationship extraction web service. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), pages 1844--1851. IEEE

    work page 2019

  3. [3]

    Lidong Bing, Sneha Chaudhari, Richard C Wang, and William Cohen. 2015. Improving distant supervision for information extraction using label propagation through lists. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 524--529

    work page 2015

  4. [4]

    David Chang, Woo Suk Hong, and Richard Andrew Taylor. 2020. Generating contextual embeddings for emergency department chief complaints. JAMIA open, 3(2):160--166

    work page 2020

  5. [5]

    Suchanek

    Lihu Chen, Gaël Varoquaux, and Fabian M. Suchanek. 2021. https://ojs.aaai.org/index.php/AAAI/article/view/17499 A lightweight neural model for biomedical entity linking . Proceedings of the AAAI Conference on Artificial Intelligence, 35(14):12657--12665

    work page 2021

  6. [6]

    Mike Conway, John N Dowling, and Wendy W Chapman. 2013. Using chief complaints for syndromic surveillance: a review of chief complaint based classifiers in north america. Journal of biomedical informatics, 46(4):734--743

    work page 2013

  7. [7]

    Marko C uljak, Andreas Spitz, Robert West, and Akhil Arora. 2022. Strong heuristics for named entity linking. arXiv preprint arXiv:2207.02824

    work page Pith review arXiv 2022

  8. [8]

    Dowling, Debbie Travers, Gregory F

    Jagan Dara, John N. Dowling, Debbie Travers, Gregory F. Cooper, and Wendy W. Chapman. 2008. https://doi.org/https://doi.org/10.1016/j.jbi.2007.11.004 Evaluation of preprocessing techniques for chief complaint classification . Journal of Biomedical Informatics, 41(4):613--623

    work page doi:10.1016/j.jbi.2007.11.004 2008

Show all 44 references
  1. [9]

    Mostafa Dehghani, Hamed Zamani, Aliaksei Severyn, Jaap Kamps, and W Bruce Croft. 2017. Neural ranking models with weak supervision. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 65--74

    2017

  2. [10]

    Jarunee Duangsuwan and Pawin Saeku. 2018. Semi-automatic classification based on icd code for thai text-based chief complaint by machine learning techniques. International Journal of Future Computer and Communication, 7(2)

    2018

  3. [11]

    Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. 2021. Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH), 3(1):1--23

    2021

  4. [12]

    Maximilian Hofer, Andrey Kormilitzin, Paul Goldberg, and Alejo Nevado-Holgado. 2018. Few-shot learning for named entity recognition in medical text. arXiv preprint arXiv:1811.05468

    2018 arXiv

  5. [13]

    Raphael Hoffmann, Congle Zhang, Xiao Ling, Luke Zettlemoyer, and Daniel S Weld. 2011. Knowledge-based weak supervision for information extraction of overlapping relations. In Proceedings of the 49th annual meeting of the association for computational linguistics: human languag...

    2011

  6. [14]

    Steven Horng, Nathaniel R Greenbaum, Larry A Nathanson, James C McClay, Foster R Goss, and Jeffrey A Nielson. 2019. Consensus development of a modern ontology of emergency department presenting problems—the hierarchical presenting problem ontology (happy). Applied clinical inf...

    2019

  7. [15]

    Jia-Hao Hsu, Ting-Chia Weng, Chung-Hsien Wu, and Tzong-Shiann Ho. 2020. Natural language processing methods for detection of influenza-like illness from chief complaints. In 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA AS...

    2020

  8. [16]

    William Hsu, Simon X Han, Corey W Arnold, Alex AT Bui, and Dieter R Enzmann. 2016. A data-driven approach for quality assessment of radiologic interpretations. Journal of the American Medical Informatics Association, 23(e1):e152--e156

    2016

  9. [17]

    Kexin Huang, Jaan Altosaar, and Rajesh Ranganath. 2019. Clinicalbert: Modeling clinical notes and predicting hospital readmission. arXiv preprint arXiv:1904.05342

    2019 arXiv

  10. [18]

    Zongcheng Ji, Qiang Wei, and Hua Xu. 2020. Bert-based ranking for biomedical entity normalization. AMIA Summits on Translational Science Proceedings, 2020:269

    2020

  11. [19]

    Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, H \'e rve J \'e gou, and Tomas Mikolov. 2016. Fasttext. zip: Compressing text classification models. arXiv preprint arXiv:1612.03651

    2016 arXiv

  12. [20]

    Sarkar, and Elizabeth S

    Sotiris Karagounis, Indra N. Sarkar, and Elizabeth S. Chen. 2020. Coding free-text chief complaints from a health information exchange: A preliminary study. In AMIA Annual Symposium Proceedings. American Medical Informatics Association

    2020

  13. [21]

    Zeljko Kraljevic, Thomas Searle, Anthony Shek, Lukasz Roguski, Kawsar Noor, Daniel Bean, Aurelie Mascio, Leilei Zhu, Amos A Folarin, Angus Roberts, Rebecca Bendayan, Mark P Richardson, Robert Stewart, Anoop D Shah, Wai Keong Wong, Zina Ibrahim, James T Teo, and Richard J B Dob...

    2021

  14. [22]

    Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. 2020. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234--1240

    2020

  15. [23]

    Scott H Lee, Drew Levin, Patrick D Finley, and Charles M Heilig. 2019. Chief complaint classification with recurrent neural networks. Journal of biomedical informatics, 93:103158

    2019

  16. [24]

    Dingcheng Li, Sijia Liu, Majid Rastegar-Mojarad, Yanshan Wang, Vipin Chaudhary, Terry Therneau, and Hongfang Liu. 2016. A topic-modeling based framework for drug-drug interaction classification from biomedical text. In AMIA Annual Symposium Proceedings, volume 2016, page 789. ...

    2016

  17. [25]

    Katherine P Liao, Tianxi Cai, Guergana K Savova, Shawn N Murphy, Elizabeth W Karlson, Ashwin N Ananthakrishnan, Vivian S Gainer, Stanley Y Shaw, Zongqi Xia, Peter Szolovits, et al. 2015. Development of phenotype algorithms using electronic medical records and incorporating nat...

    2015

  18. [26]

    Aur \'e lie N \'e v \'e ol, Pierre Zweigenbaum, et al. 2017. Making sense of big textual data for health care: findings from the section on clinical natural language processing. Yearbook of medical informatics, 26(1):228

    2017

  19. [27]

    Osborne, James S

    John D. Osborne, James S. Booth, Tobias O'Leary, Amy Mudano, Giovanna Rosas, Phillip Foster, Kenneth Saag, and Maria Danila. 2020. Identification of gout flares in chief complaint text using natural language processing. In AMIA Annual Symposium Proceedings. American Medical In...

    2020

  20. [28]

    C Rochefort, A Verma, T Eguale, and D Buckeridge. 2015. O-037: surveillance of adverse events in elderly patients: a study on the accuracy of applying natural language processing techniques to electronic health record data. European Geriatric Medicine, (6):S15

    2015

  21. [29]

    Guergana K Savova, James J Masanz, Philip V Ogren, Jiaping Zheng, Sunghwan Sohn, Karin C Kipper-Schuler, and Christopher G Chute. 2010. Mayo clinical text analysis and knowledge extraction system (ctakes): architecture, component evaluation and applications. Journal of the Ame...

    2010

  22. [30]

    Isabel Segura Bedmar, Paloma Mart \' nez, and Mar \' a Herrero Zazo. 2013. Semeval-2013 task 9: Extraction of drug-drug interactions from biomedical texts (ddiextraction 2013). Association for Computational Linguistics

    2013

  23. [31]

    Aliaksei Severyn and Alessandro Moschitti. 2015. Twitter sentiment analysis with deep convolutional neural networks. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, pages 959--962

    2015

  24. [32]

    Feichen Shen, Sijia Liu, Yanshan Wang, Liwei Wang, Naveed Afzal, and Hongfang Liu. 2017. Leveraging collaborative filtering to accelerate rare disease diagnosis. In AMIA Annual Symposium Proceedings, volume 2017, page 1554. American Medical Informatics Association

    2017

  25. [33]

    Luca Soldaini and Nazli Goharian. 2016. Quickumls: a fast, unsupervised approach for medical concept extraction. In MedIR workshop, sigir, pages 1--4

    2016

  26. [34]

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon 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

    2016

  27. [35]

    Fei Teng, Zheng Ma, Jie Chen, Ming Xiao, and Lufei Huang. 2020. Automatic medical code assignment via deep learning approach for intelligent healthcare. IEEE journal of biomedical and health informatics, 24(9):2506--2515

    2020

  28. [36]

    Ilya Valmianski, Caleb Goodwin, Ian M Finn, Naqi Khan, and Daniel S Zisook. 2019. Evaluating robustness of language models for chief complaint extraction from patient-generated text. arXiv preprint arXiv:1911.06915

    2019 arXiv

  29. [37]

    Shikhar Vashishth, Rishabh Joshi, Ritam Dutt, Denis Newman-Griffis, and Carolyn Rose. 2020. Medtype: Improving medical entity linking with semantic type prediction. arXiv preprint arXiv:2005.00460

    2020 arXiv

  30. [38]

    Michael M Wagner, William R Hogan, Wendy W Chapman, and Per H Gesteland. 2006. Chief complaints and icd codes. Handbook of biosurveillance, page 333

    2006

  31. [39]

    Byron C Wallace, Jo \"e l Kuiper, Aakash Sharma, Mingxi Zhu, and Iain J Marshall. 2016. Extracting pico sentences from clinical trial reports using supervised distant supervision. The Journal of Machine Learning Research, 17(1):4572--4596

    2016

  32. [40]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020

  33. [41]

    Jie Yang, Shuailong Liang, and Yue Zhang. 2018. Design challenges and misconceptions in neural sequence labeling. In Proceedings of the 27th International Conference on Computational Linguistics, pages 3879--3889

    2018

  34. [42]

    Sendong Zhao, Ting Liu, Sicheng Zhao, and Fei Wang. 2019. A neural multi-task learning framework to jointly model medical named entity recognition and normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 817--824

    2019

  35. [43]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before...

  36. [44]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.