Shared task on multimodal story generation at the Generation Challenges at INLG 2024 in Tokyo, Japan!
Introducing the Visually Grounded Story Generation (VGSG) Challenge, a shared task at the Generation Challenges at INLG 2024 in Tokyo, Japan! Our goal is to generate coherent and visually grounded stories from multiple images using vision and language models.
Here is our call for participation. Feel Free to join the Discord Server if you have any questions for the organizers. More details can be found in the following.
This task invites enthusiasts in vision-and-language integration to push the boundaries of vision-based story generation. This shared task aims to generate stories grounded on a sequence of images. Participants will explore the intricate relations between visual prompts and textual stories, fostering innovations in techniques for coherent, visually grounded story generation. This task is particularly challenging because: 1) the output story should be a narratively coherent text with multiple sentences, and 2) the protagonists in the generated stories need to be grounded in the images. In addition, we also elicit submissions to fine-tune existing multimodal models efficiently, which is exciting for both industry and academics.
Major progress has been made by the industry in vision-based text generation in the last few years. While generating descriptions from images or videos is widely available, grounded story generation is not evaluated with a benchmark. Investigating vision-based story generation requires several capabilities including commonsense reasoning and content planning.
We're calling for submissions that demonstrate novel approaches to generating narratives that are not only coherent but also visually grounded. We particularly welcome participants to explore the following aspects of multimodal story generation:
Both automatic and human evaluations will be used. Participants can submit a 4-page report, which will be peer-reviewed and included in the INLG 2024 proceedings if accepted. Participants will present their systems in a poster session at the Generation Challenges workshop.
Whether it's through advanced computational models, innovative narrative techniques, or new data processing methods, your contribution can help illuminate the paths toward more intuitive, human-like story generation from visual inputs. Join us in this endeavor to bridge the gap between visual prompts and story generation, and help set the stage for the next leap in vision and language research.
To demonstrate the idea and help you get started quickly, we built a very simple baseline using OpenAI GPT-4V.
We use a high-quality crowdsourced dataset Visual Writing Prompts for training, validation, and testing.
Participants need to submit their generated stories together with a system description to the shared task page.
Here is a table of our submission sites for data splits, formats, and different tracks:
| Data Split | Format | Tracks | Submission Site |
|---|---|---|---|
| val | non-anonymised | Strict/Open | https://huggingface.co/spaces/VGSG/val |
| val | anonymised | Strict/Open | https://huggingface.co/spaces/VGSG/valAnonymised |
| test | non-anonymised/anonymised | Strict | https://huggingface.co/spaces/VGSG/testStrict |
| test | non-anonymised/anonymised | Open | https://huggingface.co/spaces/VGSG/testOpen |
TBA
Xudong Hong (Saarland University)
Asad Sayeed (University of Gothenburg)
Vera Demberg (Saarland University)