Re18: Read the paper GCI Everything Has a Cause: Leveraging Causal Inference in Legal Text Analysis

诸神缄默不语-个人CSDN博文目录

论文名称：Everything Has a Cause: Leveraging Causal Inference in Legal Text Analysis
论文ArXiv下载地址：https://arxiv.org/abs/2104.09420
论文NAACL官方下载地址：https://aclanthology.org/2021.naacl-main.155/（The website has an official explainer video）
官方GitHub项目：xxxiaol/GCI: Code for Everything Has a Cause: Leveraging Causal Inference in Legal Text Analysis.

本文是2021年NAACL论文,The author is from Peking University.
This paper applies the method of causal inference on legal text data,构建了因果图,to help make decisions.This paper conducts experiments on the task of identifying and analyzing similar crimes,It is proved that this paradigm is valid,Injecting causal knowledge into neural networks can improve model performance,and can provide interpretability,尤其在few-shot条件下.
Causality is also used in classification tasks.
Discrimination and analysis of crimes with similar taskssimilar charge disambiguation：multi-class分类,But tag sets aresimilar charge set（The input is the fact description text）

本文解决了2difficult task：①Unsupervised extraction is associated with prediction resultsfactor,会有噪音.②Combining traditional causal inference models with modern neural network architectures.

1. Background

1.1 因果推理

自变量treatment,因变量outcome,The amount of change in the independent variableintervention,Calculate whether and how the dependent variable changes when the independent variable is perturbed,It's causal inference
Confounder：变量,Both independent and dependent variables are affected
treated group是自变量为1的,反之是untreated group

1.2 因果图

factor（本文中所有factor都是二元变量）和charge是节点,Cause and effect are edges
The traditional causal inference method combined with text is relatively simple,Just treat the text as a node directly,without considering the differenceaspectsuch as events.

1.3 PAG

2. 模型

Automatically build cause-and-effect graphs from factual description text,Use causal reasoning to assist legal decision making,本文中similar charge disambiguationThe effect of the framework was tested on the task.
①用关键词抽取（用YAKE+IDFCalculate the importance of the word to the count）,to identify the factual descriptionkey factors.②将相似的key factorsCluster into groups,Each group is treated as an independent node.（Each group and count of nodes in the graph）③Robust to unidentified variables（Unsupervised extraction may lead to incomplete keywords,Causal discovery was unrecognizedconfounder）的causal discovery algorithm（Greedy Fast Causal Inference (GFCI)）to construct a causal diagram.（输出是Partial Ancestral Graph (PAG)）（can be seen in the appendix,This algorithm can identify hiddenfactor）（限制：1. Prohibition of the accused node out of the edge.2. 以案例（Event description text）chronological order to limit causality）（Sampling cause and effect diagrams）④Estimate for each edgecausal strengthto reduce the impact of unreliable edges.（保持Confounder不变）（方法：Average Treatment Effect (ATE)）（估算ATE的方法：Propensity Score Matching (PSM) 在treated/untreated groupto construct similar sample pairs）
Incorporate causal knowledge into NN中：①在NN attention weights上加入causal strength限制（Add loss function）.②Used on causal chains extracted from causal diagramsRNN.

3. 实验

3.1 数据集

本文使用的是CAIL数据集.

3.2 实验结果

proven results：(1) The constructed causal diagram is reasonable.(2) Can capture subtle differences in text,Especially when there is very little training data.

Demonstrate causal superioritybaseline：GCI-co（correlation-based graph,factorPreviously such as the Pearson correlation coefficient＞0.5from the earlier ones with higher frequencyfactorConnect one edge to the other）
Explain why paradigms such as multitasking are not used、No pretrained model is used,Previous work incorporating causal inference failed to capture causal relationships within the text and therefore did nothingbaseline.
The performance of the model under different training set scales is compared.
Every experiment is in3run on a random seed,以平均ACC和macro-F1作为指标.

3.3 Cause-and-effect graph quality analysis

Robustness of the causal discovery process,Sensitivity analysis of cause and effect diagrams

1. Random Confounder
2. Placebo Treatment
3. Subset of Data

3.4 人工评估：看attention

3.5 讨论

Coarse granularity due to clustering、Negative semantics、pronoun resolution、intent

3.6 Gender equity issues arising from data imbalances

False Positive Equality Difference (FPED) and False Negative Equality Difference (FNED)