Research On Relation Extraction Based On Weakly Supervised Learning

The technology of relation extraction can quickly and effectively extract structured relation facts from massive text sentences,and provide people with more precise information retrieval and auxiliary decision-making services.So far,relation extraction based on full supervision has achieved considerable results.However,full supervision has strict requirements on the authenticity of training samples,and manual labeling takes time and effort.In contrast,weakly supervised learning is more suitable for the needs of large-scale data processing because of its low cost of labeling.However,the problems of wrong labeling and uneven labeling caused by weakly supervised signals affect the performance of extraction,which seriously hinders the application of weakly supervised relation extraction in actual production and life.Aiming at the problem of weakly supervised mislabeling,this paper proposes a neural network framework through a trade-off mechanism for distantly supervised relation extraction.The framework attempts to predict relation facts by combining the semantic information contained in the word embedding representations of sentences and entity pairs at the sentence level.Therefore,the paper designs a trade-off mechanism.The mechanism can adaptively assign impact weights to text and entity pairs according to their content.Then,the multi-layer soft attention is used to filter out valuable semantic information to help constraint weakly supervised signals in relation extraction task.Aiming at the imbalance of the sample number for different relation categories in the distantly supervised training dataset,the paper further optimizes the extraction model through a trade-off mechanism,and proposes a weakly supervised relation extraction model,which integrates multi-source semantic representations to enhance the robustness under the sample imbalance.Firstly,the model adopts first-order logic inference to integrate human-perceived and non-text discrete supervised knowledge into a low-dimensional continuous vector to reflect text characteristics,and expand the feature space of weakly supervised relation extraction.Then,different learning methods are used to mine the potential semantic information in text statements from two aspects of text content and text characteristics.Finally,different forms of weakly supervised knowledge are fused in the neural network to help the model identify various relation facts.The experimental results on the widely used dataset NYT-Freebase show that distantly supervised relation extraction framework through a trade-off mechanism effectively alleviates the effects of erroneous labeling noise,and achieves substantial improvement.In addition,theweakly supervised relation extraction method fused with multi-source semantic network performs better than the state-of-the-art algorithms on the precision,and can mine more types of relation instances with few training samples.The method can find 10 different types of relation facts in top 300 correct results,which has high practical value.