Recent study shows successful results in generating a proper language description for the given image, where the focus is on detecting and describing the contextual relationship in the image, such as the kind of object, relationship between two objects, or the action. In this paper, we turn our attention to more subjective components of descriptions that contain rich expressions to modify objects – namely attribute expressions. We start by collecting a large amount of product images from the online market site Etsy, and consider learning a language generation model using a popular combination of a convolutional neural network (CNN) and a recurrent neural network (RNN). Our Etsy dataset contains unique noise characteristics often arising in the online market. We first apply natural language processing techniques to extract highquality, learnable examples in the real-world noisy data. We learn a generation model from product images with associated title descriptions, and examine how e-commerce specific meta-data and fine-tuning improve the generated expression. The experimental results suggest that we are able to learn from the noisy online data and produce a product description that is closer to a man-made description with possibly subjective attribute expressions.