SignQA-2026

SignQA-2026 is a multilingual question-answering dataset built from sign-language sequence annotations. It contains Chinese QA pairs derived from CSL-Daily and German QA pairs derived from PHOENIX-2014T.

The current release contains annotations only. Each record refers to a source video by video_id; video files are not included.

Dataset configurations

Configuration	Language	Train	Validation	Test	Total
csl-daily-qa	Chinese (zh)	92,000	5,385	5,880	103,265
phoenix14t-qa	German (de)	35,480	2,595	3,210	41,285
All	—	127,480	7,980	9,090	144,550

The dataset contains 28,910 unique video IDs. Each video ID has five QA records, one for each module from M1 to M5.

Data format

Every JSONL row contains the following fields:

Field	Type	Description
video_id	string	Identifier of the corresponding source video
split	string	Original split name: train, dev, or test
dataset	string	Source dataset identifier: csl or phoenix
module	string	QA module, from M1 through M5
pattern_id	integer	Identifier of the question-generation pattern
question	string	Natural-language question
answer	string	Natural-language answer
lang	string	ISO language code: zh or de

Example:

{
  "video_id": "11August_2010_Wednesday_tagesschau-1",
  "split": "train",
  "dataset": "phoenix",
  "module": "M1",
  "pattern_id": 3,
  "question": "Kannst du mir sagen, welches Zeichen bei Schritt 4 steht?",
  "answer": "Das 4. Gloss lautet \"DONNERSTAG\".",
  "lang": "de"
}

Intended use

SignQA-2026 is intended for research on sign-language understanding, sequence-grounded question answering, multilingual QA, and evaluation of models that reason over sign-language gloss sequences.

Users who need visual inputs must obtain the corresponding CSL-Daily and PHOENIX-2014T videos separately and follow the original datasets' access terms.

Citation

If you use SignQA-2026 in your research, please cite our paper:

Shiwei Gan, Lichen Wang, Xiao Liu, Yafeng Yin, Kuizhuang Liu, Sanglu Lu, and Lei Xie. Sign Language Question Answering: A New Task, Benchmark, and Baseline for Sign Language Understanding. 2026.

@misc{gan2026signqa,
  title  = {Sign Language Question Answering: A New Task, Benchmark, and
            Baseline for Sign Language Understanding},
  author = {Gan, Shiwei and Wang, Lichen and Liu, Xiao and Yin, Yafeng and
            Liu, Kuizhuang and Lu, Sanglu and Xie, Lei},
  year   = {2026},
  eprint = {2607.27826},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  doi    = {10.48550/arXiv.2607.27826},
  url    = {https://arxiv.org/abs/2607.27826}
}

arXiv: 2607.27826 [cs.AI]

Limitations

• This repository does not include source videos or gloss sequences.
• Questions and answers are tied to source-dataset video identifiers.
• Template-based QA patterns may not reflect the full diversity of naturally occurring questions.
• Performance across Chinese and German should not be compared without accounting for differences between the source datasets.

Licensing and attribution

This dataset is derived from CSL-Daily and PHOENIX-2014T. Users are responsible for reviewing and complying with the licenses and terms of the original datasets. The annotation license should be stated explicitly before public release if it differs from the source-dataset terms.