A Subsequent Words Recommendation Scheme for Chinese Input Method based on Deep Reinforcement Learning
Jingyun Yang, Hengjun Wang, Kexiang Guo · 2020
In order to improve the accuracy of subsequent word recommendation for Chinese input methods and increase users' input speed, we propose a recommendation scheme for subsequent word input based on deep reinforcement learning and build a model for this scheme. In this paper, we train the model based on the deep Q-network (DQN) procedure. Firstly, the state feature extraction module is used to obtain state features with different lengths of historical input information. Secondly, the users are required to select the target items from the recommendation list generated by the word prediction module. Then, the model feeds back and updates the status of the user's selection according to the reward and punishment rules designed in this paper. Finally, through continuous human-computer interaction, the model is optimized and adjusted to obtain the optimal strategy. The experiments show that the proposed scheme makes a certain improvement in the accuracy of subsequent word recommendation compared with traditional methods.