Research On Personalization Of Dialog Systems

In recent years,the human-machine dialogue System has been an important research direction in the field of artificial intelligence.In the dialogue system,there are various addressable research problems and the generation of responses is an important task.The very common and serious issue in the response generation task is that the current response generation models produce inconsistent and general responses.Personalized response modeling of the human-machine dialogue model(Persona)is an effective means to solve the above problems.Persona focuses on the relationship between human-machine dialogue systems and users and how to generate personalized responses.In addition,applying language style to generated text is also a unique and interesting issue in dialog generation.Most researchers think it's a part of Persona,but others think it's a language attribute and should be handled separately.In human-machine dialogue technology,the speaking style or writing style is an attribute of Dialog Persona,so it should be dealt within the constraints of the Persona Modeling while generating responses.Several attempts has been made to apply Persona modeling and style modeling to Human-Machine conversational systems.Typically two problems are dealt independently.Both of the problems are complex and need to be solved separately,but Persona and Language Style independently cannot contribute to human like natural conversation generation.By allowing them to be combined in one system can significantly produce consistent and polite responses.This Thesis proposes a model for generating personality and style based dialogues at the same time,which can improve the quality of response generation.We present three models called DialogModel,Persona-Model and Politeness-Model.We train them independently on their respective data sets and then we consider the three models in combination during inference time.Our Dialog and Persona models are standard sequence-to-sequence(Seq2Seq)models,the encoders are composed of two layers of LSTM,and our Politeness model uses a self-encoder(Auto-Enc).We use a shared vocabulary,which is obtained by joining three data sets for used for training and decoding three models.Our work has a clear effect on enhancing Personality and language style information.Our model can produce more diverse responses against identical response.