The Knowledge Graph And Human-Machine Interaction Based Requirements Elicitation Methods In Soft-Robots For Services

With the rapid development of the Internet of Things and the Internet of Services,the complexity of services connected to the Internet has brought challenges to the matching of requirements to services.To this end,Soft-Robots for Services(e-SBOT)is developed for improving service delivery ability.The e-SBOT has the capacity of sensing and predicting user requirements anytime and anywhere,and connects to a back-end big service network for constructing personalized service solutions smartly.This dissertation focuses on the requirement elicitation for e-SBOT.The conventional methods suffer from two issues when they are applied to requirement elicitation for e-SBOT.First,the knowledge graph of user(KGU)is generally generated by manual construction,which may lead to incomplete user information collection;second,most of the existing methods utilize the single-turn interaction strategy,which is difficult to handle personalized requirements.To solve these problems,this dissertation conducts research from two aspects: the user knowledge graph completion and multi-turn human-computer interaction method for requirement elicitation.Specifically,the main research contribution includes the following aspects:(1)To solve the problem of fine-grained user preference recognition,KGCaps AN is proposed in this dissertation.KGCaps AN introduces the syntactic dependency tree as the prior knowledge to solve the problem of misconnection between the preference targets and related words.Moreover,KGCaps AN utilizes the capsule attention network to integrate such prior knowledge into the forward propagation.The experiment results on benchmark datasets demonstrate that KGCaps AN achieves higher performance than compared models.Second,this dissertation proposes an S-HGKT model to solve the cross-domain task.S-HGKT integrates external knowledge,such as sentiment lexicons,to help to bridge the differences between domains.To effectively integrate the external knowledge,S-HGKT introduces a KE-LSTM method to reduces the excessive forgetting of external knowledge by the forget gates.Experiments on multiple tasks reveal the effectiveness of the S-HGKT model.(2)To solve the problem of the intent tree construction,this thesis proposes an endto-end generative framework P-Mem2 seq.Specifically,the P-Mem2 seq model contains four layers,the language understanding layer is used to recognize the requirements from user input;the intention reasoning layer utilize KGU to predict the potential requirement of users;the dialogue state management layer is used to track the state of dialogue,then generates the next-turn interaction strategies;the interactive generation layer is used to generate the response content.The four layers are constructed by deep neural networks and trained in an end-to-end format.The experiment results on benchmark datasets demonstrate that P-Mem2 seq achieves higher performance than compared models.(3)To solve the problem of requirement pattern mapping,TAN is proposed in this dissertation for the intent tree reconstruction.Second,this dissertation proposes a humanmachine interaction model based on requirement patterns: Pmem2seq-RQ.The extensive experiments demonstrate that our algorithm outperforms other compared models.In summary,this dissertation focus on research of requirement elicitation of e-SBOT based on KGU and human-machine interaction.The proposed models and algorithms provide new ideas and methods for solving the requirement elicitation of e-SBOT.