| Relational extraction is one of the classic tasks in natural language processing.With the development of deep learning technology,relationship extraction methods tend to be automated,and deep learning frameworks are trained through big data to complete relationship classification.At present,there has been a large amount of research focused on independent sentence completion relationship extraction,mainly focusing on obtaining contextual information of entities in sentences.In contrast,document level relationship extraction has a wider range of application scenarios,but faces more problems.In document level relationship extraction tasks,due to the large number of entity pairs and the diversity of sample categories,the proportion of incorrect and useless information in the text is relatively higher,requiring the model to have strong ability to synthesize text information,especially the ability to aggregate key information in multiple sentences.At the same time,there is a phenomenon of category imbalance in the relationship extraction dataset.Some relationship types correspond to too many sample types,while most other relationship types contain fewer and evenly distributed sample types,resulting in the classifier ignoring the existence of minority class relationships.However,in practical life,the prediction of minority classes may be more critical.Therefore,how to extract and process imbalanced relationship datasets to reduce their skewness is of great significance for relationship classification.The main research content of this article is:1)Research on the representation of entity to entity relationships in documents.In order to capture words with strong correlation with entity relationship information in complex texts,this paper proposes a relationship representation method based on dependency graph networks.This method uses syntactic dependency analysis to learn the correlation information between each word in a sentence,uses graph convolutional networks to model the correlation of key word nodes,maps relevant words to high-dimensional space,and then uses classification models to predict the relationship types of entity pairs.Finally,in the process of relationship prediction,a multi example learning method is used to replace the information in a single text with information from multiple texts to combat noise in the text.The experimental results on a publicly available document relationship extraction dataset show that in cross sentence relationship extraction tasks,the model outperforms the comparison algorithm in both feature representation and noise removal tasks,with an F1 score improvement of 0.16%on the test set.2)Research on the imbalance of relationship categories.In order to solve the problem of imbalanced relationship categories in the dataset,this paper proposes a data enhanced imbalanced relationship classification model.In order to balance the sample size of different relationship categories,this paper proposes a text enhancement algorithm based on entity pair replacement.This algorithm utilizes a text pre training model to select contexts from the original sentence samples that can correctly express relationships,and combines them with different entity pairs to generate new sentence samples and expand the sample with fewer categories to achieve category balance.Finally,the model uses comparative learning to optimize the relationship classifier,reducing the information interference of the generated new samples on the classification model.Finally,the results on two publicly available remote supervised datasets show that the proposed model can improve the accuracy of relationship classification in small sample categories,while overall performance is improved by 0.82%and 2.63%,respectively,compared to other text enhancement methods. |
