Open Domain Chatbot Based on Attentive End-to-End Seq2Seq Mechanism

Abstract

Chatbot as a conversational system that can interact with human naturally is a Natural Language Processing task that require modeling semantics of complicated relationships of the language for communication. Various attempts have been made to reduce the complexities (language understanding, feature extraction, domain recognition, intent detection, semantic slot filling and language generation) of training text-based Chatbot. While Traditional machine learning models are usually unable to be truly generic, recent advances in deep learning allow end-to-end models to be trained with large dataset. We train an open domain Chatbot end-to-end by directly mapping input tags as sequences to generate optimized output sequence tags. Our model is trained on Google Tensorflow framework running GPUs with deep neural architecture of sequence-to-sequence and attention mechanism. The interaction of the chatbot with human performed common sense reasoning. It also achieved a competitive result when quantified based on Human Performance Evaluation. However, we suggest networked reinforcement learning to incrementally update the consistency drawbacks towards achieving a generic Chatbot.

Ähnliche Arbeiten