Models of sentences from natural languages like English are an important component of many natural language processing technologies, including speech recognition, machine translation, and text retrieval. Due to the complexity and size of natural languages, such language models are necessarily very complex and high dimensional. I will first discuss some recent advances using hierarchical and nonparametric modelling approaches giving state-of-the-art predictive results, then address the problems of adapting language models to specific domains, and the syntactic and semantic representations of words.