基于因果图分析的可解释司法判决预测方法研究

doi:10.11959/j.issn.2096-0271.2024023

摘要/Abstract

摘要：

随着人工智能技术的发展和海量司法数据的公开，面向“智慧司法”服务的司法判决预测（legal judgment prediction，LJP）任务受到了学术界和工业界的广泛关注，该任务旨在根据有限的案件事实描述文本来预测案件的罪名、法条和刑期。然而，现有工作缺乏对易混淆司法案件的智能决策的研究，且相关模型通常缺乏可解释性，这会导致模型预测严重依赖领域专家，阻碍LJP在不同法律体系中的应用。为此，提出了一种基于因果图分析的司法判决预测（prediction of legal judgment based on causal graph analysis， CGLJ）方法，首先从非结构化的法律事实描述文本中挖掘要素之间的因果关系，然后采用易混淆罪名聚类的构图方法构建因果图，既考虑了相似事实描述之间的差异，又增强了事实描述和法律法规之间的相互作用，最后将构建好的因果图融入深度神经网络进行联合推理，得到判决预测结果。此外，还对模型预测过程中的因果图推理过程进行了可视化，为判决结果提供了更好的可解释性。在2018中国“法研杯”司法人工智能挑战赛（CAIL2018）司法判决预测数据集上的实验结果表明，该方法相比基线模型取得了更好的效果。

关键词: 司法判决预测, 因果图, 可解释性

Abstract:

With the development of artificial intelligence technology and the disclosure of massive judicial data, the LJP task for ＆quot;smart justice＆quot; services has received widespread attention from academia and industry.The task aims to predict the charges, laws, and sentences of a case based on limited factual descriptions of the text.However, existing work lacks research on intelligent decision-making in easily confusing judicial cases, and related models often lack interpretability, which leads to heavy reliance on domain experts for model predictions and hinders the application of LJP in different legal systems.To this end, this article proposes a judicial judgment prediction method CGLJ based on causal graph analysis.Firstly, the causal relationships among elements are mined from unstructured legal fact description texts.Then a causal graph is constructed using a composition method of easily confused accusation clustering.It not only considers the difference among similar fact descriptions, but also enhances the interaction between fact descriptions and laws and regulations.Finally, the constructed causality diagram is integrated into a deep neural network for joint inference to obtain the decision prediction result.In addition, this paper also visualizes the causal diagram inference process in the model prediction, providing better interpretability for the judgment result.The experimental result on the CAIL2018 judicial judgment prediction dataset shows that the proposed method achieves better result than the baseline models.

Key words: legal judgment prediction, causal diagram, interpretability

中图分类号:

TP18

张虎, 张振, 范越, 郭佳钰. 基于因果图分析的可解释司法判决预测方法研究[J]. 大数据, 2024, 10(2): 109-121.

Hu ZHANG, Zhen ZHANG, Yue FAN, Jiayu GUO. Research on interpretable legal judgment prediction method based on causal graph analysis[J]. Big Data Research, 2024, 10(2): 109-121.

图/表 6

图1

图2

表1

表2

表3

图3

参考文献 27

[1]	PEARL J . The seven tools of causal inference,with reflections on machine learning[J]. Communications of the ACM, 2019,62(3): 54-60.
[2]	RIBEIRO M T , SINGH S , GUESTRIN C . Why should I trust you?:explaining the predictions of any classifier[C]// Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York:ACM, 2016: 1135-1144.
[3]	LOFTUS J R , RUSSELL C , KUSNER M J ,et al. Causal reasoning for algorithmic fairness[EB]. arXiv preprint, 2018,arXiv:1805.05859.
[4]	BAREINBOIM E , PEARL J . Causal inference and the data-fusion problem[J]. Proceedings of the National Academy of Sciences of the United States of America, 2016,113(27): 7345-7352.
[5]	LIU C L , HSIEH C D . Exploring phrase-based classification of judicial documents for criminal charges in chinese[C]// International Symposium on Methodologies for Intelligent Systems. Berlin:Springer, 2006: 681-690.
[6]	LIU C L , CHANG C T , HO J H . Case instance generation and refinement for case-based criminal summary judgments in Chinese[J]. Journal of Information Science and Engineering, 2004,20(4): 783-800.
[7]	LUO B , FENG Y , XU J ,et al. Learning to predict charges for criminal cases with legal basis[EB]. arXiv preprint, 2017,arXiv:1707.09168.
[8]	HU Z , LI X , TU C ,et al. Few-shot charge prediction with discriminative legal attributes[C]// Proceedings of the 27th International Conference on Computational Linguistics. Santa Fe:Association for Computational Linguistics, 2018: 487-498.
[9]	ZHONG H X , GUO Z P , TU C C ,et al. Legal judgment prediction via topological learning[C]// Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Stroudsburg:Association for Computational Linguistics, 2018: 3540-3549.
[10]	YE H , JIANG X , LUO Z ,et al. Interpretable charge predictions for criminal cases:learning to generate court views from fact descriptions[EB]. arXiv preprint, 2018,arXiv:1802.08504.
[11]	CROCE D , ROSSINI D , BASILI R . Auditing deep learning processes through kernelbased explanatory models[C]// Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing. Hong Kong:EMNLP-IJCNLP, 2019: 4037-4046.
[12]	SUN Z , FAN C , HAN Q ,et al. Selfexplaining structures improve NLP models[EB]. arXiv preprint, 2020,arXiv:2012.01786.
[13]	HAN X C , WALLACE B C , TSVETKOV Y ,et al. Explaining black box predictions and unveiling data artifacts through influence functions[EB]. arXiv preprint, 2020,arXiv:2005.06676.
[14]	WALLACE E , FENG S , BOYD-GRABER J . Interpreting neural networks with nearest neighbors[EB]. arXiv preprint, 2018,arXiv:1809.028478.
[15]	LEI T , BARZILAY R , JAAKKOLA T ,et al. Rationalizing Neural Predictions[EB]. arXiv preprint, 2016,arXiv:1606.04155.
[16]	NARANG S , RAFFEL C , LEE K ,et al. WT5? ! training text-to-text models to explain their predictions[EB]. arXiv preprint, 2020,arXiv:2004.14546.
[17]	JIANG Y C , BANSAL M . Self-assembling modular networks for interpretable multihop reasoning[C]// Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP). Stroudsburg:Association for Computational Linguistics, 2019: 4474-4484.
[18]	JIANG Y C , JOSHI N , CHEN Y C ,et al. Explore,propose,and assemble:an interpretable model for multi-hop reading comprehension[C]// Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Stroudsburg:Association for Computational Linguistics, 2019: 2714-2725.
[19]	ZHONG H X , WANG Y Z , TU C C ,et al. Iteratively questioning and answering for interpretable legal judgment prediction[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2020,34(1): 1250-1257.
[20]	LANDEIRO V , CULOTTA A . Robust text classification in the presence of confounding bias[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2016,30(1): 186-193.
[21]	PONTI E M , KORHONEN A . Eventrelated features in feedforward neural networks contribute to identifying causal relations in discourse[C]// Proceedings of the 2nd Workshop on Linking Models of Lexical,Sentential and Discourse-level Semantics. Stroudsburg:Association for Computational Linguistics, 2017: 25-30.
[22]	PAUL M J . Feature selection as causal inference:experiments with text classification[C]// Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017). Stroudsburg:Association for Computational Linguistics, 2017: 163-172.
[23]	LIU X , YIN D , FENG Y S ,et al. Everything has a cause:leveraging causal inference in legal text analysis[C]// Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics:Human Language Technologies. Stroudsburg:Association for Computational Linguistics, 2021: 1928-1941.
[24]	PEARL J , MACKENZIE D . The book of why:the new science of cause and effect[M]. London: Penguin, 2018.
[25]	SUN M , CHEN X , ZHANG K ,et al. Thulac:an efficient lexical analyzer for Chinese[J]. Retrieved Jan, 2016,10:2022.
[26]	CAMPOS R , MANGARAVITE V , PASQUALI A ,et al. YAKE! Keyword extraction from single documents using multiple local features[J]. Information Sciences, 2020,509: 257-289.
[27]	OGARRIO J M , SPIRTES P , RAMSEY J . A hybrid causal search algorithm for latent variable models[C]// The Eighth Edition of the International Conference on Probabilistic Graphical Models,Lugano:PGM, 2016,52: 368-379.

数据集	训练集样本数	测试集样本数	法条数	罪名数
CAIL-big	1 587 979	185 120	118	130
CAIL-deal	1 395 496	164 173	67	85

模型	法条预测
模型	1%1%	10%10%	30%30%	50%50%	100%100%
LSTM	0.34	0.41	0.51	0.66	0.79
CNN	0.32	0.43	0.57	0.74	0.83
BERT+CNN	0.35	0.48	0.62	0.80	0.89
LSTM+Casual	0.51	0.57	0.64	0.75	0.87
CNN+Casual	0.54	0.61	0.73	0.87	0.91
CGFJ(BERT+CNN+Casual)	0.55	0.62	0.75	0.89	0.937

模型	罪名指控预测
模型	1%	10%	30%	50%	100%
LSTM	0.29	0.34	0.43	0.54	0.69
CNN	0.31	0.37	0.48	0.62	0.79
BERT+CNN	0.34	0.41	0.56	0.69	0.84
LSTM+Casual	0.46	0.51	0.57	0.64	0.76
CNN+Casual	0.47	0.54	0.61	0.71	0.83
CGFJ(BERT+CNN+Casual)	0.51	0.56	0.65	0.77	0.87