The next generation ultra-reliable and low-latency communications (xURLLC) need novel design to provide satisfactory services to the emerging mission-critical applications. To improve the spectrum efficiency and enhance the robustness of xURLLC, this paper proposes a robust beamforming and rate-splitting design in the finite blocklength (FBL) regime for downlink multi-user multi-antenna xURLLC systems. In the design, adaptive rate-splitting is introduced to flexibly handle the complex inter-user interference and thus improve the spectrum efficiency. Taking the imperfection of the channel state information at the transmitter (CSIT) into consideration, a max-min user rate problem is formulated to optimize the common and private beamforming vectors and the rate-splitting vector under the premise of ensuring the requirements of transmission latency and reliability of all the users. The optimization problem is intractable due to the non-convexity of the constraint set and the infinite constraints caused by CSIT uncertainties. To solve it, we convert the infinite constraints into finite ones by the S-Procedure method and transform the original problem into a difference of convex (DC) programming. A constrained concave convex procedure (CCCP) and the Gaussian randomization based iterative algorithm is proposed to obtain a local minimum. Simulation results confirm the convergence, robustness and effectiveness of the proposed robust beamforming and rate-splitting design in the FBL regime. It is also shown that the proposed robust design achieves considerable performance gain in the worst user rate compared with existing transmission schemes under various blocklength and block error rate requirements.