AxelNoun/ohsumed-23-single-label

Dataset Card for OHSUMED (23 classes, single-label TextGCN split)

Dataset Summary

This dataset is the canonical TextGCN Ohsumed protocol (Yao et al., AAAI 2019): 7,400 single-label documents, 3,357 train / 4,043 test, 23 MeSH disease chapters. Here the paper’s 3,357 training documents are split 3,021 / 336 so recipes that need a validation set can use it, or concatenate train+validation to recover the paper’s train count.

• 7,400 MEDLINE records (year 1991), each with exactly one label.
• Labels are the 23 MeSH disease chapters C01 to C23 (C01–C23), not the child terms of C14. C14 (Cardiovascular Diseases) is class 13 — one chapter in that set of 23. The document pool is Joachims/Moschitti’s 1991 cardiovascular-oriented subset; the label set remains those 23 top-level disease chapters.
• This repository is a documented Parquet republication of dxgp/ohsumed (same 7,400 rows, previously without a card). It is not a replacement for the full ~348k IR collection (community-datasets/ohsumed), and it is not a byte-identical filter of that dump.

Single-label derivation. The Joachims/Moschitti 1991 pool is multi-label. Yao et al. (TextGCN, 2019) dropped every document with more than one of the 23 MeSH chapter labels and kept the 7,400 remaining single-label abstracts. No “first category wins” rule was applied here: this file is a documented copy of that filtered split (via dxgp/ohsumed), not a new filter of the 348k dump. An abstract that is medically both C04 and C14 is therefore absent unless it was already single-label in the source.

Why this version? The Hub hosts several incompatible “OHSUMED” objects (full multi-label IR dump, pickle CV folds, undocumented 7.4k CSV). If you compare accuracy to TextGCN, you need this split.

Recommended headline metric: macro-F1, not accuracy.

Provenance and Lineage

1. OHSUMED (Hersh et al., 1994) — 348,566 MEDLINE references, 1987–1991.
2. Joachims (1998) — first 20,000 records of 1991, 23-way categorization over MeSH chapters C01 to C23.
3. Moschitti — packaged corpus ohsumed-all-docs.tar.gz / first-20k (13,929 unique after category selection, including multi-label).
4. TextGCN (Yao et al., 2019) — keep single-label only → 7,400 docs; paper split 3,357 train / 4,043 test.
5. dxgp/ohsumed — same 7,400 rows with a 336-row validation hold-out taken from the 3,357 train documents (3,021 + 336 = 3,357).
6. This repo — Parquet + ClassLabel names C01–C23 + dataset card + NLM other license.

Bytes of text match dxgp (title + newline + space + abstract). No extra normalization.

Dataset Structure

Data Fields

• text (string): title and abstract concatenated with newline then space (\n ). That separator is preserved on purpose so token-level baselines stay comparable to dxgp / common TextGCN reimplementations.
• label (ClassLabel): integer 0–22 with names C01 to C23 (0→C01, …, 22→C23). C14 is name index 13, not the definition of the whole label set. Ordering matches MeSH codes in ascending order (verified: train[0] is duodenitis, label 5 = C06).

Label mapping

id	MeSH code	Category
0	C01	Bacterial Infections and Mycoses
1	C02	Virus Diseases
2	C03	Parasitic Diseases
3	C04	Neoplasms
4	C05	Musculoskeletal Diseases
5	C06	Digestive System Diseases
6	C07	Stomatognathic Diseases
7	C08	Respiratory Tract Diseases
8	C09	Otorhinolaryngologic Diseases
9	C10	Nervous System Diseases
10	C11	Eye Diseases
11	C12	Urologic and Male Genital Diseases
12	C13	Female Genital Diseases and Pregnancy Complications
13	C14	Cardiovascular Diseases
14	C15	Hemic and Lymphatic Diseases
15	C16	Neonatal Diseases and Abnormalities
16	C17	Skin and Connective Tissue Diseases
17	C18	Nutritional and Metabolic Diseases
18	C19	Endocrine Diseases
19	C20	Immunologic Diseases
20	C21	Disorders of Environmental Origin
21	C22	Animal Diseases
22	C23	Pathological Conditions, Signs and Symptoms

Splits

Split	Rows	Role
train	3,021	Paper train minus hold-out
validation	336	Hold-out from the paper’s 3,357 train docs (seed not published by dxgp)
test	4,043	Same count as the TextGCN test set; assignment is dxgp’s, not a re-draw of the 7,400 pool
Total	7,400	3,021 + 336 = 3,357 paper train

Comparability: this is the canonical TextGCN protocol (Yao et al., 2019): 7,400 single-label documents, 3,357 / 4,043 on 23 MeSH disease chapters. Here the paper’s 3,357 train docs are split 3,021 / 336 so that training recipes without a val set can concatenate train+validation. Test-set accuracy is therefore comparable to that literature on the same split counts.

Exact-text overlap between train∪validation and test is 0 (7,400 unique strings). The files are a documented copy of dxgp/ohsumed; we did not redraw 3,357/4,043 from the 7,400 pool. We do not claim byte identity with Yao’s original archive: only that the Hub split preserves dxgp’s assignment, and that those counts match the paper.

The GNN literature often glosses this task as “23 cardiovascular disease categories”. That is incorrect. The 23 labels are MeSH disease chapters C01 to C23. C14 (Cardiovascular Diseases) is one class. Spot-check: duodenitis is C06 (Digestive).

Published accuracy / macro-F1 figures and the input-matching caveat are in Baseline results below.

Class counts (all splits, 7,400 rows)

MeSH	Rows	Share
C14	1,175	15.88%
C04	1,030	13.92%
C23	796	10.76%
C10	557	7.53%
C21	553	7.47%
C20	410	5.54%
C06	354	4.78%
C12	342	4.62%
C18	323	4.36%
C08	250	3.38%
C17	233	3.15%
C05	223	3.01%
C01	216	2.92%
C13	195	2.64%
C15	133	1.80%
C16	129	1.74%
C11	125	1.69%
C19	86	1.16%
C02	75	1.01%
C07	63	0.85%
C09	63	0.85%
C03	50	0.68%
C22	19	0.26%

Head/tail ratio (C14 vs C22) is about 62:1. Accuracy is misleading: a classifier that ignores the tail can still look respectable. Report macro-F1 and per-class F1. Categories are not mutually exclusive in reality (e.g. cardiac neoplasms are C04 and C14); forcing one label caps achievable top-1 accuracy. Top-3 is a more honest secondary metric.

Token length on BiomedBERT (train sample n=2,000): p50=237, p90=377, p95=422, p99=517, max=650. Training default --max_length is 448.

Uses

• Research on single-label medical text classification and GNN/text-graph models, with numbers comparable to TextGCN on Ohsumed.

Out of scope

• Clinical care, diagnosis, or any non-experimental library/clinical setting (OHSUMED / NLM conditions).
• The full OHSUMED IR collection or multi-label MeSH assignment on ~348k records.
• Claiming this 7,400-row split was reconstructed by filtering community-datasets/ohsumed (whitespace-normalized overlap is high but incomplete).

Limitations

C22 is never predicted

C22 - Animal Diseases scores F1 0.000 on all three seeds. With roughly a dozen test examples and similar training support, the model never learned to emit it: no input will receive this label. Any application that needs C22 detection must use another mechanism.

Metric consequence: a dead class costs 1/23 mechanically, i.e. 4.3 macro-F1 points. Hence dual reporting — 0.699 on 23 classes, 0.731 on the 22 classes with usable support.

Scores are not calibrated

Weighted cross-entropy (α = 0.5) used against class imbalance shifts calibration. Displayed scores are relative confidence scores, not probabilities. Temperature scaling on validation would be needed before any threshold-based use.

Intrinsic label ambiguity

Two distinct mechanisms coexist; they are not fixed the same way.

1. Granularity (C23). An abstract about a general clinical sign (hemorrhage, shock, pain, syncope…) will be routed to an organ chapter. Predictable behavior: C23 is a MeSH catch-all, not a domain; the model under-emits it (precision 0.592, recall 0.370; 264 outbound vs 107 inbound leaks). Demo users should know this.

2. Polyhierarchy. MeSH is not a tree. Illustration (app.py examples):

• Lung Neoplasms: C04 and C08. The model predicts C08 (0.90) — legitimate; the corpus mostly labels “lung carcinoma / cancer” as C08.
• Multiple Sclerosis: C10 and C20. The model predicts C20 (0.86); the corpus says C20 (train 10/12, test 15/15).

The top-1 → top-3 gap (0.774 → 0.933) mostly captures these two effects — not pure error — hence Top-3 in the demo.

Future work (formulation, not capacity)

The principled fix is not a better encoder but a better task formulation: multi-label or hierarchical labeling that does not force a choice between the sign (C23) and the organ, or between two cross-listed chapters. That is the conclusion the matrix supports.

Temporal drift

Abstracts from 1987–1991. Dated terminology, drugs, and protocols. No guarantee of generalization to contemporary biomedical literature.

Not a medical device

Bibliographic categorization only. No diagnostic or clinical decision-support use.

Schema and provenance

• Schema is text + label only (TextGCN / dxgp comparability). There is no medline_ui, and title/abstract are not separate columns.
• The 336-row validation hold-out seed was not published by dxgp. With 23 classes, C22 has at most one validation example. One class is 4.3 points of macro-F1, so checkpoint selection on this split is noisy; the 4,043-row test set is the reliable measure.
• Bytes of text match dxgp; they are not a guaranteed exact subset of the 348k Hub dump. We do not claim byte identity with Yao’s original archive — only matching split counts and a documented copy of dxgp’s assignment.

Licensing and Usage

License tag: other (medline-research-use). Not CC-BY-NC or MIT (those Hub tags on related repos are unsourced).

MEDLINE-derived records follow NLM Terms and Conditions and NLM copyright guidance:

• Courtesy: "Courtesy of the U.S. National Library of Medicine."
• No endorsement: do not imply that NLM endorses this repository or any model trained on it.
• Not current MEDLINE: this snapshot is 1987–1991 (this split: 1991 abstracts) and is incomplete relative to PubMed today.
• Non-clinical: do not use for patient care or clinical decision-making.
• Abstracts: may remain under publisher/author copyright. NLM does not grant a Creative Commons license on those texts.

Redistribution here is for NLP/IR research documentation of a long-standing public benchmark.

Baseline results

Fine-tune with train.py (checkpoint selection on eval macro-F1, weighted CE α=0.5, 5 epochs, max_length 448, effective batch 16). Hardware: RTX 3060 Laptop 6 GB. Test: 4,043 documents. Dataset revision pinned: c2e11d41da4d537366e75e02bc3b21a29a4978e0.

Model	Size	Acc	Macro-F1	Macro-F1 excl. C22	Weighted-F1	Top-3
DistilBERT	66M	0.666	0.497	0.520	0.643	0.854
bert-base-uncased	110M	0.686	0.550	0.575	0.668	0.865
BiomedBERT	110M	0.774	0.699	0.731	0.766	0.933
BiomedBERT (TextGCN-preprocessed input)	110M	0.741	0.662	0.692	0.734	0.919

Decomposition: DistilBERT → bert-base = +2.0 acc / +5.3 macro-F1 (capacity); bert-base → BiomedBERT = +8.8 / +14.9 (domain).

Uncertainty

Two distinct dispersions are reported; they answer different questions:

• Initialization variability — seeds 42 / 43 / 44, same split: accuracy 0.775 ± 0.003 · macro-F1 0.699 ± 0.005 · top-3 0.934 ± 0.002. This speaks to training stability.
• Sampling variability — 95% bootstrap on the 4,043 test documents, 1,000 resamples: macro-F1 in [0.680, 0.715]. This is the one to use when comparing to a figure published elsewhere.

Released weights are seed 42 (accuracy 0.774, macro-F1 0.699). The table above reports the three-seed mean; a gap on the order of 0.001 between the two is expected.

Positioning on the TextGCN protocol

Method	Input	Accuracy	Macro-F1
TF-IDF + logistic regression	filtered	54.66	—
CNN-non-static	filtered	58.44	—
RB-GAT (RoBERTa + BiGRU + GAT)	filtered	71.48	67.90
BiomedBERT — paired comparison	filtered	74.1	66.2
BiomedBERT — released model	raw	77.4	69.9

Reading. Only the “paired” row is directly comparable to the cited work: it uses the same preprocessing. On that footing, BiomedBERT is slightly ahead in accuracy (+2.6) and slightly behind in macro-F1 (−1.7): the two approaches are at the same level. We claim no superiority.

The released model consumes raw text and posts higher absolute scores, but that number is not comparable to anything in the literature — it only shows what you get without degrading the input.

What TextGCN preprocessing costs a domain encoder

Absent from the cited literature: same BiomedBERT, same seed 42, same 3,357 / 4,043 split; only the input changes (preprocess_textgcn.py, NLTK stopwords + tokens with frequency < 5, vocabulary computed on train alone). From 0.774 / 0.699 (raw) to 0.741 / 0.662 (filtered): −3.3 accuracy points, −3.7 macro-F1.

Mechanism. The freq < 5 filter on ~3,357 training documents shrinks the kept vocabulary to 8,290 tokens (after stopword drop). Those are precisely the rare technical terms — where a WordPiece trained on PubMed beats a generalist vocabulary. The TextGCN protocol therefore systematically handicaps subword transformers relative to graph methods, which rely on frequent-word co-occurrences and lose less from that filtering. Confirming signature: macro-F1 drops more than accuracy (−3.7 vs −3.3), consistent with rare terms carrying mostly rare classes.

Scope of the comparison

Three models were compared, chosen to isolate domain effect at fixed architecture. This protocol shows that pretraining domain drives most of the gain — not that BiomedBERT is the best possible choice. Not evaluated: BioLinkBERT, recent long-context encoders, and a zero-shot LLM baseline that would set the floor without training. Contributions welcome.

Dominant confusions

Matrix (test, seed 42). Top pairs:

• C23 → C14 (50) · C23 → C10 (42) · C23 → C21 / C02 → C20 (24 each)

C23 asymmetry. Strongly outbound flow: 264 C23→X vs 107 X→C23 (ratio 2.47). Precision 0.592, recall 0.370, F1 0.455 (n=419). The asymmetry makes the mechanism visible: the model is not symmetrically confusing C23 with something else — it under-emits C23. When it predicts C23 it is right six times out of ten; it only catches one third of true C23. Signature of a catch-all class learned to fire only when no organ signal is present. Second-largest macro-F1 hole after C22 (F1 0.000) — C23 is one of the three most frequent classes.

Headline figure: macro-F1 0.699 (23 classes). Diagnostic on the 21 well-formed classes — structural criterion fixed a priori, not “the two worst scores”: C22 has no usable support, C23 is not a domain but a residual chapter — : 0.744. A weak class for any other reason would stay in the average.

Dominant confusions are not domain-to-domain ambiguity but the residual nature of chapter C23 (Pathological Conditions, Signs and Symptoms), which absorbs signs and symptoms independently of the organ system. A C23-labeled abstract talks about a sign; its text is saturated with organ vocabulary; the model predicts the organ (hemorrhage/shock → C14, pain/coma → C10, post-traumatic signs → C21). That is a labeling granularity problem, not MeSH polyhierarchy.

Only one top pair matches a documented MeSH cross-listing: C02→C20. Of the 24 cases, 23 mention HIV / AIDS / immunodeficiency / HTLV — the 1987–1991 AIDS literature, cross-listed under Virus Diseases and Immunologic Deficiency Syndromes. C04↔C08 and C10↔C20 are rare in the matrix (4 and 1) despite appearing in the demo.

python inspect_dataset.py
python train.py --dataset_revision c2e11d41da4d537366e75e02bc3b21a29a4978e0 --max_train_samples 200 --epochs 1 --max_length 128
python train.py --dataset_revision c2e11d41da4d537366e75e02bc3b21a29a4978e0 --train_batch_size 8 --gradient_accumulation_steps 2 --eval_batch_size 16 --max_length 448
python train.py --model_name distilbert-base-uncased --output_dir outputs/distilbert-ohsumed23 --dataset_revision c2e11d41da4d537366e75e02bc3b21a29a4978e0 --train_batch_size 8 --gradient_accumulation_steps 2 --eval_batch_size 16
python train.py --model_name bert-base-uncased --output_dir outputs/bertbase-ohsumed23 --dataset_revision c2e11d41da4d537366e75e02bc3b21a29a4978e0 --train_batch_size 8 --gradient_accumulation_steps 2 --eval_batch_size 16
python eval_checkpoint.py --model_id outputs/biomedbert-ohsumed23/final
python analyze_results.py --run_dir outputs/biomedbert-ohsumed23

Demo: MODEL_ID=outputs/biomedbert-ohsumed23/final python app.py (PowerShell: $env:MODEL_ID="AxelNoun/biomedbert-ohsumed23"; python app.py).

Citation

Please cite the split, the original collection, and (if you discuss the 23-way 1991 task) Joachims:

@inproceedings{yao2019graph,
  title={Graph Convolutional Networks for Text Classification},
  author={Yao, Liang and Mao, Chengsheng and Luo, Yuan},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={33},
  number={01},
  pages={7370--7377},
  year={2019}
}

@article{lv2024rbgat,
  title={{RB-GAT}: A Text Classification Model Based on {RoBERTa}-{BiGRU} with Graph {ATtention} Network},
  author={Lv, Shanshan and Dong, Jinming and Wang, Chao and Wang, Xianwei and Bao, Zhiyu},
  journal={Sensors},
  volume={24},
  number={11},
  pages={3365},
  year={2024}
}

@inproceedings{hersh1994ohsumed,
  title={{OHSUMED}: An Interactive Retrieval Evaluation and New Large Test Collection for Research},
  author={Hersh, William and Buckley, Chris and Leone, T. J. and Hickam, David},
  booktitle={Proceedings of the 17th Annual International {ACM} {SIGIR} Conference on Research and Development in Information Retrieval},
  pages={192--201},
  year={1994}
}

@inproceedings{joachims1998text,
  title={Text Categorization with Support Vector Machines: Learning with Many Relevant Features},
  author={Joachims, Thorsten},
  booktitle={Machine Learning: {ECML}-98},
  pages={137--142},
  year={1998},
  publisher={Springer}
}