In a multiuser MIMO broadcast channel, the rate performance is affected by multiuser interference when the Channel State Information at the Transmitter (CSIT) is imperfect. To tackle the detrimental effects of the multiuser interference, a Rate-Splitting (RS) approach has been proposed recently, which splits one selected user’s message into a common and a private part, and superimposes the common message on top of the private messages. The common message is drawn from a public codebook and decoded by all users. In this paper, we generalize the idea of RS into the large-scale array regime with imperfect CSIT. By further exploiting the channel secondorder statistics, we propose a novel and general framework Hierarchical-Rate-Splitting (HRS) that is particularly suited to massive MIMO systems. HRS simultaneously transmits private messages intended to each user and two kinds of common messages that are decoded by all users and by a subset of users, respectively. We analyse the asymptotic sum rate of RS and HRS and optimize the precoders of the common messages. A closedform power allocation is derived which provides insights into the effects of various system parameters. Finally, numerical results validate the significant sum rate gain of RS and HRS over various baselines.