Research On Relation Extraction Of Drug-related Judicial Texts Based On Multi-channel Decoding Structur

In recent years,China’s judiciary has been committed to further building smart courts,modernizing the judicial system,enhancing judicial capabilities,and promoting judicial digitalization and intelligence.Therefore,both the Supreme Court and the local courts are accelerating the pace of construction of Smart Court 3.0.With the application of new technologies such as big data and artificial intelligence in the judicial system has been widely explored in China,drug-related crimes are one of the most serious crimes and one of the most harmful types among all judicial cases.Moreover,the processing of drug-related judicial text is also a relatively difficult task.Fortunately,in natural language processing,a key issue is how to make computer systems understand these difficult-to-understand information and convert these unstructured data into structured data.On the one hand,the entity relationship extraction technology studied in this paper is an important method to solve this problem.On the other hand,this technology provides the basis for downstream tasks such as reasoning based on knowledge graphs,evidence chain construction,and judicial assisted sentencing.The main research work and innovations of this paper are as follows:1.We propose a feature extraction method that decouples entity features and relational features.Redundant features are reduced by enabling different features to perform their duties during the learning process.Meanwhile,different strategies are implemented for relational and entity features according to different prediction tasks,which improve the interpretability of features and the extraction performance of models.2.We propose a novel structured relation extraction model(MCREL),which avoids the exposure bias problem that arises in multi-task structures.At the same time,the model uses a novel multi-channel labeling strategy to achieve finer-grained entity splitting,which can effectively solve various relationship overlapping problems.3.We propose a generative relationship extraction model(MCTG),which uses a Non-Autoregressive network to generate Triple from multiple channels,and matches the generated result with Ground Truth through the KM algorithm,thereby transforming the relationship extraction problem into a set Forecasting problems.This model is based on independent Triple generation channels,which avoids the impact of entity pointer positioning errors on other Triple that emerge in the MCREL model.Secondly,it avoids the limitation of only using the left side features of the Triple in the autoregressive model.Non-Autoregressive model considers the influence between Triples and uses the features both of the left and right sides.Finally,the loss is calculated based on set matching algorithm,which avoids the influence of the order of Triples generating.4.We applied the above models to the relation extraction tasks of Chinese drug-related judicial texts.In more complex scenarios,the model still has good robustness.Compared with other models,the F1 value has increased by 6.4%.