Research Of Knowledge Graph-based Semantic Information Extraction Method And Its Application

In recent years,with the continuous improvement of computer performance and the rapid growth of information contained in Internet texts,it has become a general trend to construct a knowledge base consisting of concepts,entities and relationships and build an automatic question and answer system based on it.This paper examine the problem of question answering over knowledge graphs.According to the characteristics of semantic parsing,we divided it into two subcategories task: entity detection and relationship prediction.In order to solve the problem of labeling confusion caused by BiLSTM model in entity recognition process,we constructs a BiLSTM+CRF model.By using CRF layer to add constraints to the output layer,this problem can be avoided to a certain extent.The model achieves a 93% F-score on SIMPLEQUESTION dataset.Aiming at entity non alignment problem in entity linking process,we have developed a new heuristic extension method,which significantly decreases the effect of noise and increases accuracy.In relation prediction task,most of the existing methods follow so called encoder-compare framework.They map the question and the KB facts to a common embedding space,in which the similarity between the question vector and fact vectors can be conveniently computed.This,however,inevitably loses original words interaction information.To preserve more original information,we propose an ARNN-SMCNN model,which is able to capture comprehensive hierarchical information utilizing the advantages of both RNN an CNN.We use RNN to capture semantic-level correlation by its sequential model nature,and use an attention mechanism to get different expressions of questions.Meanwhile,we use a similarity matrix based CNN to extract literal-level words interaction matching.The model is tested on SIMPLEQUESTIONS dataset,and the validity of the model is verified.In addition,we also completed the annotation and extraction of entities and relationships for the problem of knowledge graph construction.Taking the construction of financial knowledge base as an example,we designs the construction process of knowledge base including data source acquisition,corporate entity extraction,business relationship extraction and so on.Aiming at entity recognition in financial field,we first analyses the influence of feature construction on the results of conditional random field model,and verifies the role of the features extraction in this paper.In addition,the idea of active learning is applied to the condition random field model to solve the problem of entity extraction under the condition of insufficient samples.On relation extraction tasks,we defines a variety of relationships in the financial field,and formulates candidate keywords for them.In order to solve the problem that feature combination consumes a lot of labor costs,we combines Gradient Boosting Decision Tree model and Logistic Regression model by stacking manner.Compared with the comparative algorithm,the model achieves good results in all relational classification experiments.