基于深度学习的司法智能研究
发布时间：2018-11-23 16:09
【摘要】：本课题为基于深度学习的司法智能研究,任务主要以司法领域的自动量刑、相关法条预测和相似案例推荐为主。旨在以深度学习技术为主,解决司法领域智慧化问题,开展人工智能与法律领域的结合。在研究过程中,以单人单罪的刑事案件作为实验数据。自动量刑指的是在给定案情描述的情况下,预测案件的罪名、刑期和罚金。在实验中分别采用词袋模型、fast Text和卷积神经网络模型,对刑期和罚金任务,对比使用了预测值变换和数字离散化等方法。罪名预测上卷积神经网络模型效果最好,准确率为96.22%。刑期预测上,最好结果为平均绝对误差5.42个月,平均绝对比例误差36.60%,一致率43.04%。罚金预测上,最好结果为平均绝对误差5199元,平均绝对比例误差52.36%,一致率34.06%。相关法条预测指的是在给定案情描述的情况下,预测案件引用的法条信息。在实验中分别尝试了多种实验思路,如比照法条文本、多标签分类和通过相似案件的法条预测。同时也尝试了融合更多信息的模型,如罪名预测结果和案件要素抽取结果。其中以融合更多信息的多标签分类结果最好,在平均覆盖率@5上结果为92.34%,宏平均准确率为89.43%,宏平均召回率为87.02%,宏平均F1值为88.21%,微平均准确率为88.08%,微平均召回率为84.23%,微平均F1值为86.11%。相似案件推荐指的是在给定案情描述情况下,通过文本相似度的计算在已有的案件库中推荐部分相似案件。在研究中分别尝试了词频-逆向文件词频、doc2vec、词频-逆向文件词频和word2vec融合等方法,其中词频-逆向文件词频和word2vec融合的效果最好。在模型评估上,通过采用人工打分的方法,以avgDCG@5作为评价指标,最好结果为18.51。
[Abstract]:The task of this paper is to study judicial intelligence based on deep learning. The main tasks are the automatic sentencing in the judicial field, the prediction of relevant laws and the recommendation of similar cases. The purpose of this paper is to solve the problem of wisdom in judicial field and to combine artificial intelligence with law field. In the course of the study, single-person single-crime criminal cases as experimental data. Automatic sentencing refers to the prediction of charges, sentences, and fines given a description of the case. In the experiment, word bag model, fast Text and convolution neural network model are used to compare the term of imprisonment and the task of fine by using the methods of predictive value transformation and numerical discretization. Convolution neural network model is the best in charge prediction, and the accuracy is 96.22. The best result is the average absolute error of 5.42 months, the average absolute proportion error of 36.60 and the consistent rate of 43.04. In the prediction of fine, the best result is the average absolute error of 5199 yuan, the average absolute proportion error of 52.36 and the consistent rate of 34.06. The relevant law prediction refers to the law information quoted in the case given the case description. In the experiment, we try many kinds of experimental ideas, such as comparing the text of the law, classifying the multi-label and forecasting the similar cases. At the same time, we try to integrate more information models, such as the result of charge prediction and the result of case element extraction. The results of multi-label classification with more information are the best, with 92.34 on the average coverage @ 5, 89.43 on the average accuracy of macros, 87.02 on the average recall of macros, 88.21 on the average F1 values of macros. The average accuracy was 88.08, the recall rate was 84.23, and the F1 value was 86.11. Similar case recommendation refers to the recommendation of some similar cases in the existing case base through the calculation of text similarity under the given case description. The methods of word frequency-reverse file word frequency, doc2vec, word frequency-reverse file word frequency and word2vec fusion are tried respectively in the research. Among them, word frequency-reverse file word frequency and word2vec fusion are the best. In model evaluation, the best result is 18.51 by using artificial scoring method and avgDCG@5 as evaluation index.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：D926;TP18
本文编号：2352014
[Abstract]:The task of this paper is to study judicial intelligence based on deep learning. The main tasks are the automatic sentencing in the judicial field, the prediction of relevant laws and the recommendation of similar cases. The purpose of this paper is to solve the problem of wisdom in judicial field and to combine artificial intelligence with law field. In the course of the study, single-person single-crime criminal cases as experimental data. Automatic sentencing refers to the prediction of charges, sentences, and fines given a description of the case. In the experiment, word bag model, fast Text and convolution neural network model are used to compare the term of imprisonment and the task of fine by using the methods of predictive value transformation and numerical discretization. Convolution neural network model is the best in charge prediction, and the accuracy is 96.22. The best result is the average absolute error of 5.42 months, the average absolute proportion error of 36.60 and the consistent rate of 43.04. In the prediction of fine, the best result is the average absolute error of 5199 yuan, the average absolute proportion error of 52.36 and the consistent rate of 34.06. The relevant law prediction refers to the law information quoted in the case given the case description. In the experiment, we try many kinds of experimental ideas, such as comparing the text of the law, classifying the multi-label and forecasting the similar cases. At the same time, we try to integrate more information models, such as the result of charge prediction and the result of case element extraction. The results of multi-label classification with more information are the best, with 92.34 on the average coverage @ 5, 89.43 on the average accuracy of macros, 87.02 on the average recall of macros, 88.21 on the average F1 values of macros. The average accuracy was 88.08, the recall rate was 84.23, and the F1 value was 86.11. Similar case recommendation refers to the recommendation of some similar cases in the existing case base through the calculation of text similarity under the given case description. The methods of word frequency-reverse file word frequency, doc2vec, word frequency-reverse file word frequency and word2vec fusion are tried respectively in the research. Among them, word frequency-reverse file word frequency and word2vec fusion are the best. In model evaluation, the best result is 18.51 by using artificial scoring method and avgDCG@5 as evaluation index.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：D926;TP18
【参考文献】
相关硕士学位论文 前1条
1 高菲;基于机器学习的计算机辅助量刑初探[D];华东政法学院;2005年
，本文编号：2352014
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2352014.html
