Let me give you a simple definition for grounded semantics, there are two ways to look at it:
- Mapping between symbols in language (words, phrases, and etc) and their corresponding perceptual representations. (color words to color representations such as RGB)
- Finding the common ground meaning for words in dialog interaction. (negotiation of word meanings on the fly)
They are trying to model vagueness for color representations and give it possibility of being used in pragmatic reasoning (the examples are in demo section of their published code).
There are two recent work a short paper and a poster showing up in EMNLP 2016, which they are doing things that I am interested in. They basically exposed the idea of using encoder-decoder model for colors. It’s like machine translation for colors. (translate the given color code to English, or with a given word what would be the most likely color). In recent machine learning seminar in Chalmers, I talked about these works, here is the slides. I wasn’t aware of the word-to-color work at that time, they have a nice demo here.
What is my problem now?
Well, I am trying to make sense of problems that I see if we consider principles of compositionality for color semantics. For starter, I think negation is the most problematic part (what does “not red” mean?, can you incrementally learn compositional “not” in “not red”?). I will come with some code and examples later, this might be interesting for representation learning community.
I am not going to publish any post next week. But hopefully the week after that, I will have some codes and explanation of what I am working on :D.