Shared-Private LSTM for Multi-domain Text Classification

Shared-private models can significantly improve the performance of cross-domain learning. These methods use a shared encoder for all domains and a private encoder for each domain. One issue is that domain-specific knowledge is separately learned, without interaction with each other. We consider tackling this problem through a shared-private LSTM (SP-LSTM), which allow domain-specific parameters to be updated on a three-dimensional recurrent neural network. The advantage of SP-LSTM is that it allows domain-private information to communicate with each other during the encoding process, and it is faster than LSTM due to the parallel mechanism. Results on text classification across 16 domains indicate that SP-LSTM outperforms state-of-the-art shared-private architecture.