Skip to content

Dixit Data

On twitter I saw a mention of an interesting paper, one that is based on the popular dixit card game.

The goal of the game is to describe your card with a perfect clue.

The paper lists some examples of cards. You can imagine the wide scope of captions that are possible.

This clue needs to be chosen such that most, but not all (!), players guess the right surreal image card. It involves a lot of creative, lateral thinking and it's a joy to play.

The paper lists an example of a round too.

This creative captioning is also something that is incredibly hard to do for an AI, which is why the paper is suggesting that it might be an engaging benchmark.


Dixit, it turns out, is a popular board game to study. The paper lists many examples of research studies that involve the board game. The effect of Dixit as part of therapy is studied as well as how the game might elicit more lexical diversity in children. It's because of the broad creativity that is involved that the paper argues it might make for an interesting variant of the Turing test. If a game is played and none of the human players can distinguish between the AI and the fellow players, it would indicate a new milestone.

It's an interesting idea for sure and you would touch a few interesting topics in AI. When you consider the task of creative captioning, you suddenly get to combine many fields. The paper mentions:

  • Computer Vision
    • Object recognition: What is in the image?
    • Scene analysis: How are the objects related?
    • Affective analysis: What is the emotional content of an image?
  • Natural Language Processing/Understanding/Generation
  • Story Reasoning/Narrative Understanding/Generation

I'm curious how you might go and practically implement this though. I guess we'll see.