A Study And Implementation Of Multi-level Semantics Model For Multi-turn Dialogue System

It is a great challenge that to make the computer understanding the utterances that full of complexion and abstraction.We used to design a chatbot based on template or statistic,which are commonly inflexible and unscalable.These years saw the VUI(Voice User Interfaces)arise as an alternative to the traditional graphic interfaces and most of mobile devices are equipped with a voice assistant when they are released to the market.In addition,the smart sound box come into our life and become our daily routine.Nowadays the smart assistant can interact with smart home devices,and also chatting with users like a real human.NLP(Natural Language Process)plays an important role in these processing and acts like a “Neutral center”.Never before have we make such a rapid progress since the neutral network applying to the NLP with the help of GPU accelerating.In in paper,we combined deep learning and information retrieving techniques and dived into the implementation of retrieval-based multi-turns dialogue system.By proposing a multi-level semantic model,which extracts different levels of information from words to utterances with three submodels.First of all,a word sequence model was proposed based on Recursive Neural Network,to encode the context and candidate responses into vectors,which would be used in a evaluation model to compare similarity.With the advantages of LSTM Neural network,the word sequence model will capture information of semantics in global level.An utterance sequence model was then to be introduced,making use of information of turns,which was discarded during the processing of word sequence.Apart from RNN,the model is composed of convolutional network and attention mechanism.The last one is the keyword model based on traditional information retrieval technique.When it comes to a long context,the word sequence model failed to keep previous information,which will be handled by the keyword model,by encoding the entire context into a semantic vector.We tested our hybrid method on public dataset,the result of which shows a great surpass to the other baseline methods.In addition,more experiments had been conducted for the further analysis of relationship between submodels and hyperparameters.We also do some visualization work to conclude the principle of the model.