Avoiding Green and Colorless Ideas: Text-based Color-related Knowledge Acquisition for Better Image Understanding

In this paper we introduce a simple text mining method which could be helpful for automatic image understanding process, especially in object recognition. By using colors as an example of visionrelated feature category, we describe how word frequencies, dependency parsing and quasi-semantic filtering help to acquire more accurate knowledge which usually requires costly and time consuming annotation process to be properly obtained. We describe retrieved data, preliminary experimental results and after analyzing errors we suggest possible solutions for improvement. We conclude the paper with a discussion about using the retrieved knowledge in fields like image processing or metaphor generation and understanding.