We know what you are going to post: User Status Generation Based on Personality and Topic

Automatic sentence generation is an easy task for humans but an extremely challenging task for computers. Recent research has shown that Recurrent Neural Networks (RNN) based language model offers a promise to perform this task. However, traditional RNN is not able to generate semantically different sentences for people with different personalities and on different topics. In this project, we propose a model called Recurrent Neural Network based on Personality and Topic (RNNPT) to generate social media user status condition on user personality and status topics. In order to do so, we use a huge set of user personality and status data, and initialize the hidden layer in our models using the personality and topic representations. The personality representation are collected using Five-Factor Model in Psychology [11] while the topic representations are learnt using Latent Dirichlet Allocation (LDA) [3] which separates the status into a certain number of topics. We concatenated the topic representation and the personality representation to be the initialization vector for the hidden layer at time step 0. Two variants of our proposed models, namely RNNPT and RNNPT1 (with one-hot topic representation) are trained and used to output sentences based on different initialization vectors which indicated different user personalities and topics. We analyzed the performance of our proposed models both qualitatively and quantitatively and concluded that our model was able to generate high-quality user status which reflects the input personality and topics at a satisfactory level. Furthermore, our models outperform baseline RNN models in terms of both perplexity and BLEU scores.