Abstract: RNN based networks have been widely used in various applications to obtain impressive performance, and CARU has more advantages in NLP tasks. However, the RNN exerts a great pressure on the CARU unit when a single layer is used. In this work, we propose to implement a multi-layer design, which can gradually extract the main features through multiple CARU units. The advantage of this is that it can consider part of speech and content. It allows each layer to perform its work clearly while alleviating long-term dependencies. Using seven popular data streams, the performance of multi-layer CARU is compared and evaluated with many state-of-the-art technologies. Experiments show that our design can improve the classification performance of various data sets. In addition, the design and implementation can be easily deployed in RNN based systems.