Dataset Card for LSC-CARE

Dataset Summary

LSC-CARE (Catalan Annotated Resource for hEalthcare) is a specialized multimodal dataset designed to advance research in Catalan Sign Language (Llengua de Signes Catalana - LSC) within the healthcare and medical domains. The dataset provides high-quality video recordings of professional interpreters signing patient protocols along with synchronized text translations and human-annotated glosses.

Unlike general administrative corpora, all video clips in LSC-CARE were recorded specifically for this dataset based on real-world medical patient protocols obtained from Catalan Hospitals. Note that this dataset does not contain audio.

Supported Tasks and Leaderboards

This dataset enables research on medical sign language recognition (SLR), translation (SLT), tokenization/gloss generation, and video classification based on clinical specialties.

Languages

The dataset focuses on Catalan Sign Language (LSC) translations with textual support and human-annotated glosses in Catalan.

Dataset Splits & Statistics

The dataset is divided into three distinct splits (Train, Validation, and Test). Below is the distribution data for the 10 listed medical topics and anonymized interpreter codes across all splits.

1. Train Split (2,227 total rows)

The training split contains 2,227 examples, for a total of 3.5033 hours of video, excluding all fragments without active signing. Medical topics are distributed in the following way:

Medical Topic	Count	Percentage
General	523	23.5%
Trauma	334	15.0%
Triage	299	13.4%
Information	252	11.3%
Pneumo	198	8.9%
Dermatological	192	8.6%
Gynecology	137	6.2%
Ear, Nose and Throat	123	5.5%
Oftalmology	92	4.1%
Pediatrics	77	3.5%

Interpreter Frequencies (Train):

Interpreter Code	Count	Percentage
FB	441	19.8%
DT	395	17.7%
TT	377	16.9%
NU	209	9.4%
DC	124	5.6%
CW	122	5.5%
BF	122	5.5%
EB	116	5.2%
YB	110	4.9%
FW	107	4.8%
GW	104	4.7%

2. Validation Split (278 total rows)

We extracted 10% of the examples for validation, acounting for 278 examples or 0,4308 hours of active signing. Topics are distributed as follows:

Medical Topic	Count	Percentage
General	65	23.4%
Trauma	40	14.4%
Information	37	13.3%
Triage	36	12.9%
Dermatological	27	9.7%
Ear, Nose and Throat	19	6.8%
Pneumo	18	6.5%
Oftalmology	14	5.0%
Gynecology	11	4.0%
Pediatrics	11	4.0%

Interpreter Frequencies (Validation):

Interpreter Code	Count	Percentage
TT	61	21.9%
FB	53	19.1%
DT	47	16.9%
NU	26	9.4%
YB	15	5.4%
CW	15	5.4%
FW	15	5.4%
EB	14	5.0%
GW	13	4.7%
DC	11	4.0%
BF	8	2.9%

3. Test Split (279 total rows)

Finally, another 10% of the data was extracted for testing, acounting for 279 videos with 0.4351 hours of active signing. Topics are distributed as follows:

Medical Topic	Count	Percentage
General	72	25.8%
Trauma	54	19.4%
Triage	33	11.8%
Information	27	9.7%
Ear, Nose and Throat	24	8.6%
Gynecology	19	6.8%
Pneumo	17	6.1%
Dermatological	16	5.7%
Oftalmology	10	3.6%
Pediatrics	7	2.5%

Interpreter Frequencies (Test):

Interpreter Code	Count	Percentage
DT	53	19.0%
TT	53	19.0%
FB	47	16.8%
NU	26	9.3%
FW	16	5.7%
BF	16	5.7%
EB	16	5.7%
CW	15	5.4%
YB	13	4.7%
DC	12	4.3%
GW	12	4.3%

---

Catalan Sign Language (LSC)

Catalan Sign Language (LSC) is the natural sign language used by the deaf community in Catalonia. It features its own distinct morphology, grammar, and spatial syntax, independent of spoken or written Catalan. Formally recognized by law, LSC acts as a vital foundation for healthcare accessibility and cultural identity for tens of thousands of regional signers.

---

Dataset Structure

Data Instances

The final dataset is compiled in Parquet format. Each instance maps to a segmented video file and features full textual outcomes as well as linguistic gloss tokens.

{
  "signal": "video_segments/BSC_T3_GEN_000107_FB_20250429.mov",
  "signal_start": "0.0",
  "signal_end": "5120.0",
  "glosses": "PRO1 METGE EXAMEN VENIR",
  "output": "El metge vindrà a examinar-te ara."
}

Data Fields

• signal: Local file path directory to the specific video segment (.mp4 or .mov).
• signal_start: The segment's relative starting point (timestamp in seconds).
• signal_end: The segment's relative ending point (timestamp in seconds).
• glosses: Human-annotated linguistic sign glosses transcribed in sequential order.
• output: Textual clinical translation in written Catalan.

Dataset Creation

Curation Rationale

Deaf individuals routinely face critical communication barriers in healthcare environments due to a scarcity of clinical translation frameworks. LSC-CARE targets this critical resource gap by establishing a curated, clinical dataset containing precise human annotations (glosses) to train highly accurate and robust machine translation models for medical applications.

Source Data

All recording materials originate from standardized patient safety and clinical protocols gathered from collaborating Catalan Hospitals. Videos were captured in a highly controlled studio ecosystem to ensure optimization for computer vision models.

A group of 11 deaf native signers executed the physical recordings. Signers wore dark clothing contrasting against a green chroma backdrop to ensure high visual fidelity for machine learning feature-extraction networks.

Annotation

The video clips were completely cataloged, segmented, and annotated by a specialized team of sign language interpreters at Universitat Pompeu Fabra (UPF) in Barcelona. Alignments between text directives, timestamps, and custom gloss configurations underwent manual validation steps to ensure exact contextual accuracy.

Personal and Sensitive Annotation

To preserve privacy rights, the identities of all 11 participating sign interpreters are completely anonymized across file indexes and metrics using 2-letter randomized identifiers (e.g., FB, DT, TT, NU).

Considerations for Using the Data

Social Impact of the Dataset

LSC-CARE serves as a baseline foundational dataset for generating automated diagnostic assistance tools, interactive hospital check-in kiosks, and real-time medical text-to-sign conversion systems tailored specifically for Catalonia's health network.

Other Known Limitations

This dataset does not contain audio streams. Additionally, the annotations focus exclusively on patient care protocols across specialized medical subdomains. Models trained on LSC-CARE may experience performance degradation if applied to conversational contexts or domains outside of clinical environments.

Additional Information

Recorded and annotated by the Catalan Sign Language Laboratory at Universitat Pompeu Fabra (UPF), Barcelona.

Discussion of Biases

No specific bias mitigation strategies were applied to this dataset. Inherent biases may exist within the datasets, reflecting the biases present in the source data.

Other Known Limitations

The dataset contains almost exclusively data from the medical domain. Application of this dataset in a general domain or in other specialized domains would be of limited use.

Additional Information

Dataset Curators

Catalan Sign Language Laboratory at Universitat Pompeu Fabra (LSC-LAB)

Funding

This work has been promoted and financed by the Government of Catalonia through the Aina Project.

Acknowledgements

We gratefully acknowledge the following individuals and organizations for their valuable contributions to this corpus:

• Machine Translation group at the BSC-AI Institute (https://www.bsc.es/research-development/research-areas/cognitive-computing/machine-translation)
• Prof. Carlos Escolano (FIB-UPC) https://futur.upc.edu/CarlosEscolanoPeinado

Licensing Information

This work is licensed under a Attribution-NonCommercial 4.0 International licence.

Citation Information

[N/A]

Contributions

[N/A]