Deep Learning for Classification of Bi-Lingual Ads in Online Classifieds

摘要

Classification of ads in online classifieds is a domain well suited for applying deep learning for understanding images and text. Since many different items are in the same category, classification based on images alone is hard. Adding the title of the ad increases classification accuracy significantly. This paper describes a system developed for an online classifieds site in Thailand (kaidee.com), where titles are often a mixture of Thai and English. To achieve machine understanding of bi-lingual text, a character-level neural embedding was used. Both ID convolution and bidirectional long short-term memory (BLSTM) were examined, with convolution being both more accurate and quicker to train. The Inception v3 model was used to extract visual features from images. Visual features and character embeddings are concatenated and fed into a classifier. The results show that this approach is better than classifying based on either image or text alone. A focus of this paper is the simplicity of the solution, yielding an accuracy of 86.0% applied on real-world data.