Learning subconcepts in user-generated text from social media sites: A case of exploring food concept

Learning concepts from social media textual data is challenging due to its informal communication style. A task was proposed for modeling time-varying conceptual characteristics of a core food concept by learning and tracking its associated time-dependent subconcepts as well. The preliminary experimental study demonstrated that the Latent Dirichlet Allocation model can learn meaningful subconcepts from microblog data. In the meantime, it was found that in contrast to information retrieval, choice of term weights will either suppress or lift certain term types so as to radically interrupt relevance of subconcepts and participating terms. The term frequency weight suits better in constituting semantically reasonable subconcepts which are more relevant to the core concept. The generated subconcepts can help construct a collective picture of the core concept from the social media perspective.