Re20:读论文的先例:普通法的信息理论分析

Gods are silent-personal CSDN blog directory

Paper name: What About the Precedent: An Information-Theoretic Analysis of Common Law
Paper ArXiv Download Address: https://arxiv.org/abs/2104.12133
The official NAACL download address for the paper: https://aclanthology.org/2021.naacl-main.181/ (There is the author's official NAACL explanation video. The author's accent is the most correct English I have heard in the past few days!)
Official GitHub project: valvoda/Precedent

This article is a 2021 NAACL paper with authors from Cambridge.
It is very important to cite precedents in the legal system of the sea. This article analyzes which of the two legal theories is more in line with the facts: one is that the reason for the ruling of the precedent has a greater impact on the result of the ruling in this case, and the other is that the factual part of the precedent has a greater impact on this case.The verdict has more impact.This paper adopts the viewpoint of information theory and models the legal decision task as a classification task. By calculating the mutual information between the two parts and the decision result, it is finally proved that the reason for the decision that is a precedent has a greater impact.In addition, the reasons for some cases in which another theory is established are also analyzed.

Compare the information entropy/the size of the mutual information between the two and the case results, specifically comparing these two things:

That is, you want to separate the reason for the decision from the decision result.

Uh, this one is, special mathematics...

1. Background

1. Halsbury's: Arguments of precedent are considered important (experimental results: √)
   Goodhart's: Facts of precedent are considered important
   (The former emphasizes reasoning more than the latter, which assumes that if one does not want to be the same as precedent, it is easier to say that the precedent is factually different than to overturn it)
2. ratio decidendi: The reason for the decision, the text refers to the arguments or facts of the precedent as the reason for the decision.

2. Model

Approximate with a neural network:

3. Experiments

3.1 Dataset

ECHR dataset, case text can be divided into facts, arguments (the judge's interpretation) and outcome. Arguments cite relevant cases, also known as the precedent
Sub-corpus: contains deduplicated case citation relationships

3.2 Experiment Settings

Longformer
Hyperparameters: Due to limited resources, the hidden layer dimension is 50, the batch size is 16, and each case is truncate to 512 tokens.
4 Nvidia P100 16GiB GPU
Up to 6 hours

3.3 Experiment Results

Small cross entropy, big MI

Discusses how some laws differ from othersThe same situation, such as being truncated because it happened too late...