Multi-level Weight Optimization Based Distant Supervision Relation Extraction

Distant supervision relation extraction has become one of the basic methods for extracting structured relations from unstructured natural text.The goal of entity relation extraction is to extract structured relation data from massively growing data,and it is an important supporting technology in knowledge graph construction and intelligent search engines.This task usually requires a large-scale high-quality labeled data to train extraction model.In order to reduce the cost of data annotation,distant supervision assumption was proposed to generate training data.However,the automatically generated dataset contains many mislabeled sentences.Even by constructing sentence bags,the problem caused by mislabeling cannot be solved well.At the same time,in the sentences extracted from natural language,the number of words irrelevant to the expression of the relationship is large,so that the feature extractor cannot effectively extract the features of the sentence.Based on the above analysis,this thesis proposes a relation extraction method with multilevel weight optimization,which proposes a sentence bag reorganization method and optimizes the weight of sentences at the sentence bag level.And this thesis proposes an attention mechanism with the shortest dependency path and entity type information.The extraction method optimizes the weight of words at the sentence level,which improves the effect of distant supervision relation extraction.Both weight optimization schemes improve the performance of relation extraction.The major contributions of this thesis can be listed as follows.(1)To address the problem of unreasonable sentence bag division in the multi-instance learning strategy of distant supervision relation extraction,this thesis proposes a relation extraction method based on sentence bag reorganization of sentence similarity,which optimizes the weights of different sentences from the sentence bag level.The method first uses a neural network that does not require additional annotation data to extract the representation vector of the sentence,and then the sentence bags are split or merged according to the size of the sentence bags and the similarity between sentences.A large sentence bag will be split into several small sentence bags,which can increase the weights of the sentences in these high-quality sentence bags.At the same time,multiple smaller sentence bags will be merged into one large sentence bag according to the similarity between the sentences,which helps to reduce the probability that the sentence bag does not contain any correctly labeled sentences.Finally,on the reconstructed sentence bag,a feature extractor based on a piecewise convolutional neural network is used to extract the features of sentences to achieve relation classification.The experimental results show that the proposed method can significantly improve the effect of entity relation extraction.(2)To address the problem that the sentence length in the remote supervised relation extraction dataset is long,this thesis proposes a distant supervision relation extraction method based on dependency path and entity type,which optimizes the weight of words at the sentence level.Based on the method of sentence bag reorganization,this method develops a dependency path attention mechanism to enhance the capability of neural network via producing complementary representation for the extraction model and rich the features extracted by the model by adding entity type information.Experiments on public datasets show that this method has a performance improvement of 3% compared with the current mainstream relation extraction methods,which proves that this method can more effectively achieve the goal of relation extraction.