Topic models, a type of statistical model used to uncover hidden structures in data collections, have seen significant advancements over the last decade. Originally, these models were primarily utilised for text data, but their application has now expanded to include other forms of data such as images and music. The two main types of topic models are Latent Dirichlet Allocation (LDA) and Hierarchical Dirichlet Process (HDP). LDA, the simpler of the two, assumes a fixed number of topics, whereas HDP does not.
Despite their usefulness, topic models have certain limitations. They struggle with short, ambiguous documents and can produce topics that are difficult to interpret. Additionally, they are unable to handle the temporal dynamics of topics. To overcome these challenges, researchers are developing new models that can take into account temporal changes and semantic coherence.
The future of topic models is promising with the potential for applications in many areas. For example, they could be used to track public sentiment over time, or to analyse the evolution of scientific fields. As technology advances, the power and reach of topic models are set to increase.
Go to source article: http://radar.oreilly.com/2015/02/topic-models-past-present-and-future.html?imm_mid=0cdd47&cmp=em-data-na-na-newsltr_20150304