uznlp-uz/patient-comments

Uzbek Patient Comments for Doctor Specialty Classification

An Uzbek text classification dataset for mapping patient comments to the appropriate medical doctor specialty. The dataset contains raw and normalized patient comments together with specialty labels and metadata.

Dataset Summary

• Dataset ID: uznlp-uz/patient-comments
• Language: Uzbek (uz)
• Rows: 205,350 comment rows
• Columns: 15
• Split: train
• Format: UTF-8 TSV
• Data file: patient_comments_doctor.tsv
• License: CC BY 4.0
• Main input fields: text_raw, text_normalized
• Main target field: final_doctor_label
• Human-readable target field: final_doctor_name

Data Fields

Field	Description
master_record_id	Unique master record identifier.
text_raw	Original patient comment text.
text_normalized	Normalized version of the patient comment.
final_doctor_label	Final doctor specialty label. This is the primary classification target.
final_doctor_name	Human-readable doctor specialty name corresponding to final_doctor_label.
resolution_status	How the final label was resolved for the master record.
occurrence_count	Number of source occurrences represented by the master record.
candidate_labels	Candidate doctor labels observed before final resolution.
candidate_doctor_names	Candidate doctor specialty names observed before final resolution.
confidence_mean	Mean confidence score across candidate/source assignments.
confidence_max	Maximum confidence score across candidate/source assignments.
manual_review_final	Final manual review indicator: No, Medium, or Yes.
source_manual_review_any	Whether any source record had a manual review indicator.
assignment_method_sample	Sample method or rule family used during assignment.
matched_signals_sample	Sample matched signals or evidence used during assignment.

Tagset

The primary tagset is the doctor specialty label set in final_doctor_label. NONE is used when the doctor specialty cannot be determined reliably from the comment.

Label	Doctor specialty	Rows	Share
NONE	Aniqlanmadi / Noma'lum	106,770	51.99%
PED	Pediatr	38,780	18.88%
CAR	Kardiolog	8,076	3.93%
ENT	LOR / Otorinolaringolog	6,792	3.31%
GAS	Gastroenterolog	5,454	2.66%
END	Endokrinolog	5,450	2.65%
PUL	Pulmonolog	4,280	2.08%
PSY	Psixiatr / Psixoterapevt / Psixolog	4,084	1.99%
GYN	Ginekolog	3,853	1.88%
NEU	Nevrolog / Nevropatolog	3,099	1.51%
DEN	Stomatolog	3,083	1.50%
DER	Dermatolog	2,668	1.30%
INF	Infeksionist	2,330	1.13%
ALL	Allergolog / Immunolog	2,038	0.99%
URO	Urolog	1,696	0.83%
SUR	Umumiy jarroh	1,591	0.77%
OPH	Oftalmolog	1,366	0.67%
TRM	Travmatolog / Ortoped	1,314	0.64%
TER	Terapevt / Umumiy amaliyot shifokori	602	0.29%
RHE	Revmatolog	511	0.25%
HEM	Gematolog	489	0.24%
EMR	Tez yordam / Shoshilinch tibbiy yordam	450	0.22%
REH	Reabilitolog / Fizioterapevt	258	0.13%
NPH	Nefrolog	146	0.07%
ONC	Onkolog	146	0.07%
PSUR	Plastik va rekonstruktiv jarroh	24	0.01%

Label Values

Doctor specialty labels: ALL, CAR, DEN, DER, EMR, END, ENT, GAS, GYN, HEM, INF, NEU, NONE, NPH, ONC, OPH, PED, PSUR, PSY, PUL, REH, RHE, SUR, TER, TRM, URO

Resolution status labels: single, duplicate_consensus, duplicate_conflict_to_NONE

Manual review labels: No, Medium, Yes

Statistics

Overview

Metric	Value
Comment rows	205,350
Columns	15
Doctor specialty labels	26
Unique text_raw values	205,350
Unique text_normalized values	196,870
Source occurrences represented by occurrence_count	241,021
Duplicate TSV rows	0
Empty values in released TSV	0
Minimum raw text length, characters	1
Maximum raw text length, characters	3,907
Mean raw text length, characters	100.81
Median raw text length, characters	78
Minimum normalized text length, characters	1
Maximum normalized text length, characters	2,257
Mean normalized text length, characters	99.25
Median normalized text length, characters	76
Minimum normalized text length, tokens	1
Maximum normalized text length, tokens	395
Mean normalized text length, tokens	13.79
Median normalized text length, tokens	11

Resolution Status Distribution

Resolution status	Rows	Share
single	169,764	82.67%
duplicate_consensus	33,342	16.24%
duplicate_conflict_to_NONE	2,244	1.09%

Manual Review Distribution

manual_review_final	Rows	Share
No	112,241	54.66%
Medium	58,047	28.27%
Yes	35,062	17.07%

source_manual_review_any	Rows	Share
No	112,683	54.87%
Medium	59,741	29.09%
Yes	32,926	16.03%

Occurrence Count

Metric	Value
Minimum occurrence_count	1
Maximum occurrence_count	12
Mean occurrence_count	1.17
Median occurrence_count	1
Sum of occurrence_count	241,021

Confidence Scores

Field	Min	Max	Mean	Median
confidence_mean	0.1085	0.9987	0.8416	0.9300
confidence_max	0.1085	0.9987	0.8549	0.9500

Assignment Method Distribution

Top assignment method values by row count:

Assignment method	Rows	Share
rule	26,215	12.77%
rule_medium	22,271	10.85%
rule_strong	17,975	8.75%
model	16,795	8.18%
none_service_rule	14,176	6.90%
model_none	13,667	6.66%
none_model	13,173	6.41%
none_admin	11,263	5.48%
none_no_medical_signal	9,698	4.72%
none_unclear	7,023	3.42%
none_reply_rule	5,636	2.74%
none_very_short_nonmedical	4,795	2.34%
none_lowconf	4,592	2.24%
none_short	4,050	1.97%
none_low_signal	3,271	1.59%
context_none	3,073	1.50%
context_override	2,905	1.41%
exact_match	2,720	1.32%
model_lowconf	2,378	1.16%
none_advice_reply_like	2,099	1.02%

Normalization and Cleanup

• The released file is UTF-8 TSV.
• Rows with empty text_raw or empty text_normalized were removed.
• The released TSV has no empty field values.
• text_raw stores the original patient comment text.
• text_normalized stores a normalized form used for grouping and matching.
• Whitespace-normalized exact duplicate text groups were merged into master records.
• Ambiguous duplicate label groups were resolved conservatively to NONE.

Loading

from datasets import load_dataset

dataset = load_dataset("uznlp-uz/patient-comments", split="train")
print(dataset[0])

The TSV file can also be loaded directly:

from datasets import load_dataset

dataset = load_dataset(
    "csv",
    data_files="patient_comments_doctor.tsv",
    delimiter="\t",
    split="train",
)

Intended Use

This dataset can be used for Uzbek medical NLP research and development, including patient comment classification, doctor specialty routing, label distribution analysis, weak supervision research, and benchmarking Uzbek text classification models.

Limitations

• The dataset is intended for NLP research and model development, not for clinical decision-making.
• NONE is a valid label and represents comments where the doctor specialty could not be determined reliably.
• Patient comments are user-generated and may contain spelling variants, informal language, short replies, or non-medical content.
• Labels include rule-based, model-based, context-based, and duplicate-resolution signals. Users should consider these provenance fields during evaluation.
• The class distribution is imbalanced: NONE and PED are the largest classes.

Citation

If you use Uzbek Patient Comments Doctor Specialty, cite the dataset repository:

@misc{uzbek_patient_comments_doctor_specialty,
  title = {Uzbek Patient Comments Doctor Specialty},
  author = {{Elov B.B., Alaev R.H.}},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/uznlp-uz/patient-comments}},
  license = {CC BY 4.0}
}