INFORMATION explosion highlights the need for machines to better understand natural language texts. In this paper, we focus on short texts which refer to texts with limited context. Many applications, such as web search and microblogging services etc., need to handle a large number of short texts. Obviously, a better understanding of short texts will bring tremendous value.One of the most important tasks of text understanding is to discover hidden semantics from texts. Many efforts have been devoted to this field. For instance, named entity recognition (NER) 1, 2 locates named entities in a text and classifies them into predefined categories such as persons, organizations, locations, etc. Topic models 3, 4 attempt to recognize “latent topics”, which are represented as probabilistic distributions on words, from a text. Entity linking 5, 6, 7, 8 focuses on retrieving “explicit topics” expressed as probabilistic distributions on an entire knowledge base. However, categories, “latent topics”, as well as “explicit topics” still have a semantic gap with humans’ mental world. As stated in Psychologist Gregory Murphy’s highly acclaimed book 9, “concepts are the glue that holds our mental world together”. Therefore, we define short text understanding as to detect concepts mentioned in a short text. Fig. 1 demonstrates a typical strategy for short text understanding which consists of three steps: 1) text segmentation – divide a short text into a collection of terms contained in a vocabulary (e.g., “book dis- neyland hotel california” is segmented as fbook disneyland hotel californiag); 2) type detection – determine the types of terms and recognize instances (e.g., “disneyland” and “california” are recognized as instances, while “book” is a verb and “hotel” a concept); 3) concept labeling – infer the con- cept of each instance (e.g., “disneyland” and “california” refer to the concept theme park and state respectively). Overall, three concepts are detected from short text “book Disneyland hotel California” using this strategy, namely theme park, hotel, and state in Fig. 1.