Design and Implementation of Intelligent Personalized Digital Assistant Using Deep Reinforcement Learning: Summarization, Prioritization, and Suggestion

Nowadays, people use digital media applications such as email, social media, micro blogs, mobile social networks, video streaming, etc., to share their feelings, ideas, opinions, or comments with each other. These applications comprise a lot of digital media data (e.g., text, voice, image, or video, and content for short). Due to this increasing usage of the digital media applications, people face an issue that they do not have an efficient way to view or summarize all the contents one-by-one as it consumes a lot of time. Existing summarization techniques, are confined with respect to the original textual content, face with the out-of-vocabulary, the coverage, and the repetition problems. Although digital assistants, e.g., Siri, Google Now, and Cortana, use artificial intelligence (AI) approaches to help people analyze the content from the digital media applications, they lack the capabilities of interpreting and extracting the information from the content, and do not provide suggestions for smart reply to people based on the content. Apart from that, prioritizing the content from the digital media applications based on its importance level is a challenge. As per our knowledge, we are the first to explore the area of prioritizing the content from the digital media applications based on their importance level. Hence, we will investigate and propose an intelligent personalized digital assistant (PDA) which can help people in summarizing the content, in prioritizing the content, and in providing suggestion for smart reply based on the content from the digital medial applications. We use deep neural network (DNN) and deep reinforcement learning (DRL) approaches in designing the PDA. The PDA comprises of 3 modules: summarization, prioritization, and suggestion. In the summarization module, the PDA will summarize the textual and non-textual content. For the textual content summarization, we design a DRL based long short term memory (DRL-LSTM) approach. For summarizing the non-textual content to text, we design a DRL based convolutional neural network (DRL-CNN) approach. In the prioritization module, the PDA will identify the personal preferences of each person by analyzing the person's interaction with the digital media applications. We use the action identification API from the Watson workspace, to identify the actions present in the content. The personal preferences, and the actions identified, are combined with the DRL-LSTM and DRL-CNN approaches for prioritizing the content. In the suggestion module, we extract latent features based on a person's response to the content. The latent features and the identified actions are combined with DRL-LSTM and the DRL-CNN approaches to provide suggestions as smart reply based on the understanding of the content from the digital media applications. As we use DRL in the PDA, every module has a different reward function which is updated according to the values obtained with in that module. We designed a cloud based distributed storage and processing system architecture for the PDA. To train the PDA, we deployed the CNN/Daily mail dataset and the New York Times, China Times, and Liberty Times datasets on the system architecture.

Project ID:PB10708-2132
External Project ID:MOST107-2221-E182-072