Research On Question Answering And Question Generation Of Complex Questions Based On Knowledge Graph

Knowledge graph is an important data structure for storing information.Questions and answers that use knowledge graph as a data source are knowledge graph question and answer.Knowledge graph question answering system analyzes natural language questions and returns answers directly.At present,the knowledge graph question answering technology has been widely used in related intelligent search and recommendation services.It can be seen that the knowledge graph question answering has important practical value.At present,academia has done a lot of research on knowledge graph question and answer,but most of the research only focuses on single-hop questions or chain multihop questions,while there are relatively few researches on more general and complex questions.On the one hand,it is for complex questions.It is difficult to understand and the query structure corresponding to complex problems is usually a non-chained graph structure,which makes the question and answer more difficult;on the other hand,there is less training data for complex problems,resulting in insufficient data to train the model.In response to the above problems,this article has conducted the following three aspects of research:(1)Improve the current SOTA abstract query graph generation model AQGNet.The abstract query graph is an abstract representation of the query graph.When the abstract query graph is determined,the number of candidate query graphs can be significantly reduced.This article first redefines the concept of abstract query graphs,which makes the abstract query graph's representation of the structure of the query graph more accurate,and further reduces the number of candidate query graphs under the condition that the accuracy is basically unchanged;in view of the characteristics of Levi Graph,This paper proposes a novel graph representation model BGGAT to encode abstract query graphs,which further improves the performance of the abstract query graph generation model and achieves a new SOTA performance.(2)Propose a new knowledge graph question answering model AQG4 KGQA.The above-mentioned AQGNet can generate abstract query graphs corresponding to natural language problems,thereby greatly reducing the number of candidate query graphs,and can form a Pipeline with the ranking model to complete the knowledge graph question and answer task.Based on the abstract query graph generation framework,this paper proposes a novel end-to-end query graph generation model AQG4 KGQA.AQG4KGQA uses the auxiliary learning method to train the model with query graph generation as the main task and abstract query graph generation as the auxiliary task.Although it has not reached the new SOTA performance,the experimental results are still competitive.(3)Propose a new knowledge graph problem generation model BGGAT4 KGQA.At present,the main reason for the further improvement of the performance of the knowledge graph question answering model is the lack of high-quality complex problem data sets.This paper proposes a knowledge graph problem generation model BGGAT4 KGQG based on the BGGAT proposed in this paper.The model takes the knowledge graph subgraph and the answer as input,and automatically generates natural language questions related to the input knowledge graph subgraph and answer.Related experiments prove that the performance of the model is better than that of the current SOTA.obvious improvement.