Research Of Legal Element Extraction Based Reasoning

The combination of judicial analysis and natural language processing is an important way for artificial intelligence to be applied to the legal industry.These technologies can reduce the complicated work of professionals.The construction of effective LegalNLP algorithm system requires two parts of algorithms:first,the method based on label symbols needs to use professional legal knowledge to define professional label symbols,and the human decision-making process is reflected in a more interpretable way through label symbols;Secondly,based on the embedding method,an effective neural network model needs to be designed to achieve better performance of downstream tasks.At present,the methods based on embedding have been widely studied and applied in many practical scenarios,but the interpretability of the answers given is poor,and the answers given need to be interpreted by the label symbols of the legal profession,of which the more important is the legal element.Legal elements are key elements in the composition of decisions.They not only bring intermediate supervisory information to the decision prediction task,but also make the prediction results of the model more interpretable.At present,there is still no large-scale dataset based on legal element annotation,and there is no targeted research for legal element extraction.The method of research of legal element extraction based reasoning proposed in this paper solves the above two problems,including the method of legal element extraction based on inverse learning and gradient guidance and the semi-supervised method of legal element extraction based on the joint representation of elements and polarity.(1)The legal element extraction based on inverse learning and gradient guidance uses inverse learning to reconstruct the information of the predicted results to form new sentences,which are consistent with the predicted results in terms of element correlation.The features of the new sentence and the original sentence calculate the mean square error and derive a gradient containing correlation errors.The gradient is incorporated into the main prediction module through a circular structure,and the final prediction result of the model is obtained through repeated cycles,enabling the model to use correlation for reasoning to obtain implicit legal elements.(2)Semi-supervised legal element extraction based on the joint representation of element polarity utilizes a small amount of labeled information to train models with high accuracy and versatility,uses an element-polarity label propagation algorithm,a small amount of labeled information and a large amount of sentence similarity information to conduct label propagation to obtain pesudo distribution;Using a small number of tags to fine-tune BERT and use it to predict sentences without labels,another set of pesudo distributions is obtained.Synthesizing the results of the two distributions to obtain high confidence pseudos and conducting self training,enables the model to use labeled text for reasoning to obtain high confidence pseudos for unlabeled text.Sufficient ablation and comparative experiments have demonstrated that the reasoning based legal element extraction methods proposed in this article have achieved optimal performance in their respective comparative methods,and the task metrics of legal element extraction has increased by an average of 5%compared to other methods in the field.Through experimental analysis,the proposed method solves the two problems currently faced by the task of extracting legal elements:the lack of labeled data and the implicit legal elements in the fact description.