| Knowledg graph multi-hop question answering is a research field that uses knowledge graph as a data source to answer multi-hop questions.In the reasoning process,the incompleteness of the knowledge graph often makes the reasoning not get the correct answer or get the answer through the wrong reasoning path.The multi-hop questionanswer method based on knowledge graph embedding embeds the natural language question text and knowledge graph into the semantic space,and uses the score function for similarity calculation to obtain the answer to the question,which can effectively solve the problem of incompleteness of the knowledge graph.However,the embedding of questions and knowledge graphs faces challenges such as inaccurate semantic representation and inconsistent semantic space of embedding.In this paper,we give corresponding solutions to the above problems based on the method of knowledge graph embedding,and the main work is as follows:(1)To address the problem that multi-hop question answering ignore the different contributions of different words of the question text to the semantics of the question in the question text embedding,a method TIP is proposed to weight different words in the question text embedding according to their different contributions to the semantics of the question.he method balances the embedding of the question semantics by penalizing generic words to boost the weights of scarce words,i.e.,the semantic contributions in the question text The larger the value of the word has the greater TIP weight.Statistical experiments on the TIP values of the problem texts in the publicly available Web QSP dataset show the importance of distinguishing different word semantic contribution values in the problem text embedding process.(2)To address the problem of inaccurate semantic embedding of problematic text in the dataset,this paper uses a long short term memory network to obtain contextual information of word embedding and differentiates the embedding semantics by TIPweighted word representation,and proposes an effective multi-hop question answering model,enhance QE.The Hit@1 indexes of the model in the Meta QA 1-hop,2-hop and 3-hop Kg-half sets were 84.4%,92.0% and 71.5%,respectively,which exceeded Graft Net,Pull Net,KV-Mem and embed KGQA.(3)To address the problem that the question text embedding space of the dataset is inconsistent with the knowledge graph embedding space,we propose a model TIPNet that uses the relational information in the knowledge graph triad to shorten the semantic space distance between the question embedding and the knowledge graph embedding.The model integrates the TIP module into the filtering of candidate answers and designs an answer filtering module that distinguishes the semantic contributions of words.In the mainstream evaluation metric Hit@1,the KG-full setting of the Web QSP dataset reached 71.1% and the KG-half setting reached 56.0%,which exceeded the best KG-full setting of the comparison model,Pull Net,by 3 percentage points and the best KG-half setting,embed KGQA,by 2.8 percentage points,respectively.The CWQ dataset reached 37.4% for the KG-full setting and 33.2% for the KG-half setting,which improved the accuracy of answer inference. |
