ltg/norbert3-fine-absa

NorBERT3-fine-ABSA

This model is a fine-tuned version of NorBERT3-large, applied on the sentence-level NorPaC_absa dataset. The model is trained on a total of 66 unique aspect+sentiment labels. Details with code and guidelines can be found in our GitHub repository.

Example Usage

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("ltg/norbert3-fine-absa")
model = AutoModelForSequenceClassification.from_pretrained("ltg/norbert3-fine-absa", trust_remote_code=True)

model.eval()

text = "fastlegen lytter til meg, men jeg synes ventetiden er for lang."

# tokenize input
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)

# Run inference
with torch.no_grad():
    outputs = model(**inputs)

# Get predictions
threshold = 0.5
probs = torch.sigmoid(outputs.logits).squeeze()
predictions = [model.config.id2label[i] for i, prob in enumerate(probs) if prob > threshold]
print(predictions)
# -> ['ppr_pos', 'wtp_neg'] (patient-provider/staff relationships:positive, waiting time for appointment:negative)

Class labels

Aspect		NorPaCabsa		GP		SMH
Full name	Short-name	#	%	#	%	#	%
Healthcare providers and staff
Competence of providers	cp	401	7.2	247	8.9	154	5.5
Information sharing with patients	isp	88	1.6	44	1.6	44	1.6
Language	lang	13	0.2	12	0.4	1	0.1
Patient–provider/staff relationships	ppr	917	16.5	412	14.9	505	18.0
Time Spent with healthcare Professionals	tshp	265	4.8	102	3.7	163	5.8
Organization of health services
External cooperation with other services	excos	128	2.3	100	3.6	28	1.0
Internal cooperation and communication	incc	30	0.5	10	0.4	20	0.7
Structure and routines	sr	167	3.0	58	2.1	109	3.9
System-level organization of health services	slohs	193	3.5	101	3.7	92	3.3
Duration of treatment and stays	dur	77	1.4	–	–	77	2.8
Access and availability
Geographical distance to GP office	gd	19	0.3	19	0.7	–	–
Telephone and digital communication	td	181	3.3	181	6.6	–	–
Waiting times in clinic	wtc	54	1.0	54	2.0	–	–
Waiting time for appointment	wtp	121	2.2	121	4.4	–	–
Workload	wol	66	1.2	66	2.4	–	–
Environment and facilities
Physical and psychosocial environment	ppe	110	2.0	16	0.6	94	3.4
Activities	act	113	2.0	–	–	113	4.0
Interaction with other patients	iop	51	0.9	–	–	51	1.8
Quality of food and meal routines	qfm	90	1.6	–	–	90	3.2
Treatment
Medication	med	106	1.9	35	1.3	71	2.5
Stability and continuity in treatment	sct	482	8.7	392	14.2	90	3.2
Forced treatment / coercion	ftc	36	0.7	–	–	36	1.3
Uncategorized / Top-level aspects
Outcome and impact of treatment / stay	oits	319	5.7	–	–	319	11.4
Patient involvement and participation	pip	68	1.2	–	–	68	2.4
General	gen	1376	24.7	727	26.3	649	23.2
No aspect / Neutral	no-asp	92	1.7	64	2.3	28	1.0
Total		5563	100.0	2761	100.0	2802	100.0

Evaluation

For evaluation metrics, please refer to our paper (link coming).

Citation

@inproceedings{storset-etal-2026-pain,
    title = "From Pain to Praise: Aspect-Based Sentiment Analysis for {N}orwegian Patient Feedback",
    author = "Storset, Lilja Charlotte  and
      Jelin, Elma  and
      Norman, Rebecka Maria  and
      Bjertnaes, Oyvind  and
      {\O}vrelid, Lilja  and
      Velldal, Erik",
    editor = {Danilova, Vera  and
      Kurfal{\i}, Murathan  and
      S{\"o}derfeldt, Ylva  and
      Reed, Julia  and
      Burchell, Andrew},
    booktitle = "Proceedings of the 1st Workshop on Linguistic Analysis for Health ({H}ea{L}ing 2026)",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.healing-1.16/",
    doi = "10.18653/v1/2026.healing-1.16",
    pages = "191--202",
    ISBN = "979-8-89176-367-8",
    abstract = "This paper describes a new dataset for aspect-based sentiment analysis (ABSA) for analyzing patient feedback about healthcare services. In an interdisciplinary collaboration spanning the fields of natural language processing and healthcare research, we manually annotate a dataset of 2382 free-text comments collected from national patient experience surveys in Norway, covering two sub-fields of services {--} special mental healthcare and general practitioners. Annotations are provided on both the sentence- and comment-level, covering a fine-grained set of 25 unique healthcare-related aspects and their polarities. We also report results for fine-tuning both encoder- and decoder models on the resulting dataset, comparing different modeling strategies, like joint and sequential prediction of aspects and polarity. The resources developed in this work can assist healthcare researchers in the analysis of patient feedback, bringing a much more efficient approach compared to today{'}s manual analysis, potentially leading to improved patient satisfaction and clinical outcomes."
}