Scenario-based question answering (SEQA) is to answer a question which is contextualized by a scenario and expects a paragraph-long answer. A scenario may contain text and images to describe a concrete case in the domain. GeoSEQA is a SEQA dataset containing 4,003 scenario-based essay questions collected from China's high-school geography exams. We randomly split the 4,003 questions into 2,393 for training (59.8%), 781 for validation(19.5%), and 829 for test (20.7%).

The train_raw.json, val_raw.json and test_raw.json files consist question, answer and scenario( material, image and image_text):

{
    "id":"id",
    "background_id":"background_id",
    "src":"question",
    "tgt":"gold-standard answer",
    "material":["text scenario",...],
    "image":[["image title","image annotation by template"],...],
    "image_text":[["image title","image annotation by natural language"],...],
    "object2Geonames":{"value of entity slot in image annotation by template":{"longitude"：longitude,"latitude"：latitude},...}
}