Research On Controllability Chat-oriented Dialogue System

In the era of the Internet and artificial intelligence,dialogue systems promote the connection between humans and information in a natural way.Specifically,chit-chat dialogue systems in open domains have showed promising potentials in research and applications.Current neural chit-chat dialogue systems are usually hard to control the response attributes,such as generating a happy response.The weak controllability of response attributes has greatly limited the in-depth research and the application of the dialogue system.Therefore,controllable chat dialogue systems are key for further research and application.Existing chit-chat dialog systems mainly based on sequence-to-sequence framework and the encoder-decoder structure,is often trained end-to-end on the large corpus.Such models often have difficulty controlling the specific attributes of the response,such as emotions and sentence functions.Specifically,the emotional dialogue task aimes to generate a response that conforms to the target emotion.As for sentence functions,there are mainly four types in the language,which are interrogative,declarative,imperative,and exclamatory.We can express the dialog intent through controlling the sentence function.Therefore,this paper focuses on emotional dialogue generation and sentence function controlling tasks,and diving into the research of the modelling of conditioned generation.And the contributions are as follows:(1)For sentiment word embeddings in the emotional dialogue generation,this paper proposes a multi-task learning based model to train sentiment word vectors.This model combines the language model and sentiment classification tasks,which effecetively incorporates the representations of semantics and emotions using the pre-training and fine-tune fashion.The proposed sentiment word vectos can balance the representations of emotions and semantics:for general words,it focuses on embedding the semantic information;for sentimental words,it incorporates fine-grained emotional information.Experiments results show the proposed sentimental word vectors can effectively embed emotional and semantic information,which in turn accelerates the down-stream emotional dialogue task.(2)For the emotional dialogue generation task,this paper designes a decoupled representation model of conditional variational autoencoders.Based upon the representation learning and controlling factors,this paper explicitly decouples the emotional representation and content representation.And a newly designed paired decoder is employed to combine the emotional and content states,Experiments show that the proposed model is capable of improving the emotional accuracy of responses,while maintains the dialogue diversity.(3)For the sentence function controlling in dialogues,this paper puts forward a hierarchical generate model with the guidance of global control signal.Specifically,the decoder is divided into a manager module and a worker module.And the latent variable in the conditional variational autoencoders and the leaked feature from function classification are used as the global controlling signal.The manager module takes charge of formulating sub-goals and then the worker module is guided to generate words.As a result,the proposed model does not require hand-engineered lexicons,which allows it trained in an end-to-end way and the controlling ability is enhanced.Based upon the research of emotional dialogue generation and sentence function controlling in dialogue,this paper makes an comprehensive analysis on controlled generation in single-turn chat-oriented dialog systems.This work has great significance to build a more controllable dialog system.