Construction Of Task Type Dialogue System Based On Multimodel Fusion

Human-computer dialogue and interaction are becoming more and more common in people's daily lives,and even become an indispensable lifestyle for people.The manmachine dialogue includes three types: chat-type dialogue,question-answer dialogue,and task dialogue.Different types of human-machine dialogues require different algorithm models.Thesis studies and designs a task-based dialogue system based on multi-model fusion,and uses Python to simulate multiple models.Thesis first studies the existing Chinese representation methods,including TF-IDF,word2 vec,CWE,JWE,BERT,etc.,and proposes a Hierarchical JWE model.By jointly predicting different levels of information such as words,characters,and sub-characters in the target word,HJWE makes full use of the semantic information within the Chinese word and improves the performance of word embeddings in tasks such as text classification.Secondly,on the basis of various word embedding models,thesis studies the algorithm flow of retrieval matching model,including BM25,WMD,SMN,etc.In the context of multi-round dialogue application,a sentence semantic representation algorithm based on weighted SVD decomposition is proposed.This algorithm first weights the sentence word embedding matrix,and then performs SVD decomposition on the weighting matrix,retaining the right vector corresponding to the top k largest singular values Add up to get a vector representation of the sentence.This algorithm can retain the real and effective semantic information in the sentence and remove the noise embedded in the sentence word matrix.The simulation on the JDDC data set shows that,under BLEU as the evaluation index,this algorithm is improved by 0.10 ? 0.13 compared with the method of directly averaging the sentence embedding matrix,and is improved by 0.05 ? 0.09 compared with the TF-IDF weighted average method.Finally,thesis studies the application of Seq2 seq model based on LSTM,Transformer and other modules in the dialogue system.Aiming at the high-frequency scenes in the JDDC dialogue data set,mining and refining,obtaining dozens of scenes such as invoices,logistics,etc.and building a rule-based dialogue module.Based on the retrieval matching model,generating model,and rule model,a new multi-model fusion dialog frame is proposed.This framework enters the user's questions into the rule model to generate replies.If no corresponding replies are matched,they are entered into the retrieval model and the generated model to generate multiple replies,and then input into the rearrangement model,and the highest matching response is selected.This framework makes full use of the advantages of each model and improves the stability and robustness of the dialogue system.