Re8: reading papers Hier spcnet: a legal stat hierarchy based heterogeneous network for computing legal case

The gods were silent - personal CSDN Blog Directory

Title of thesis ：Hier-SPCNet: A Legal Statute Hierarchy-based Heterogeneous Network for Computing Legal Case Document Similarity
The paper ArXiv Download address ：https://arxiv.org/abs/2007.03225
Paper official ACM Download address ：https://dl.acm.org/doi/abs/10.1145/3397271.3401191（ The site also has SIGIR Video of the presentation of papers at the meeting , All the pictures in this article except the pictures attached to the paper are taken from this video ）

This article is about 2020 year SIGIR The passage , It is the similarity task of Indian legal instruments （ It belongs to legal information retrieval Legal IR field ）. The data and code are not public .
Legal information retrieval includes recommendation system and precedent search , It aims to search for cases matching a given scene or case from a large number of precedents .
This paper is a graph based similarity computing paradigm , The first one combines legal knowledge （ Legal information ） To do the work of legal document similarity ： Put legal documents and articles （ Combined with its hierarchical information ） Make a heterogeneous diagram , Through the heterogeneous graph embedding method node2vec¹ and metapath2vec² Get node representation , Calculate node similarity through cosine similarity , That is, the similarity of documents .
in addition to , This paper also proposes that , Text based methods and graph based methods can be combined （ Find the maximum or average ）, To complement each other , Get better results .
The calculated similarity and manually marked data are used to calculate the Pearson correlation coefficient as the model index .
I'm particularly speechless about this article. Its indicators haven't yet 2017 Text only baseline The effect is good , Then follow the text-based baseline Made a vote like hybrid Just brush up the indicators . i don't know what to say ！ This kind of article can also be published SIGIR！ That can take the place of , I also want to go SIGIR！
This paper is to do the task of extracting calligraphy strips from Indian legal texts LeSICiN³ One of the important reference papers of this paper , Some of this blog post is in LeSICiN The existing content in will be briefly written , Please refer to the corresponding blog .

1. Background & Motivation

Supervised learning method is not applicable to legal document similarity tasks , Because there is not enough tag data .
The similarity of legal documents is not strictly defined , Mainly rely on experts . The similarity of legal documents needs to be interpretable .

Legal document similarity tasks are commonly used in two ways: text-based and web-based .
Based on text ：Measuring Similarity among Legal Court Case Documents
Based on the Internet ：⁴ and ⁵
hybrid：Finding Similar Legal Judgements under Common Law System

Model principle ： Quote the same law or precedent , Or quote different laws or precedents 、 But documents that are similar in network structure are similar .

There are two sources of legal knowledge in the common law system ： Written laws and precedents .

The previous network-based similarity calculation method only considers the citation relationship between case documents precedent citation network (PCNet), This will lead to the loss of an important source of legal knowledge ： Hierarchy of legal provisions .

Former PCNet Indicators used to measure the similarity of legal documents ：

1. Bibliographic Coupling⁴：precedent citations (out-citations) A collection of Jaccard similarity index
2. Co-citation⁴: similar Bibliographic Coupling, But with in-citations
3. Dispersion⁵： Measure the out-neighbours (out-citation documents) The similarity of , That is, whether it exists in the same community / In cluster .NetworkX The latest version of the implementation function ：https://networkx.org/documentation/stable/reference/algorithms/generated/networkx.algorithms.centrality.dispersion.html（ I didn't read it , I'll learn more about it if necessary ）

2. Hier-SPCNet The construction of graph

Full name ：Hierarchical Statute and Precedent Citation Network

node ：
Case study
articles of law （5 Kind of ）

Relationship ：
Document quotation law （ You can quote any level of law ）
Mutual citation of documents
The law points to each other
There is a hierarchical relationship between laws （ This is similar to LeSICiN Of , But it's a little different ）：Act → Part → Chapter → Topic → Section/Article（ Not everyone has all levels ）

3. Representation of nodes

node2vec¹： By random walk （BFS or DFS） Generate node representation .
The implementation tool is aditya-grover/node2vec,128 dimension , Other super parameters are default values .
because node2vec Suppose the network is homogeneous , therefore Hier-SPCNet It is also treated as a homogeneous graph .

metapath2vec²： be based on user-defined metapaths
This paper defines 14 Start with a document metapath, Related to legal system .metapath Neighbor relationship implies some similarity .
The implementation tool is stellargraph · PyPI
Follow LeSICiN similar ,Hier-SPCNet Medium metapath It also starts from the same kind of nodes . But this article just needs to do document representation
Then the most outrageous thing is that it is omitted here 10 Kind of metapath The definition of , I'm speechless ！

4. experiment

4.1 Data sets

Data collected from the Supreme Court of India , Climb from Thomson Reuters Westlaw India, Only public data is used .
Extract references from text ： Regular expression based patterns , Such as < [section or article number] of the [Act] >
share 1806 Case documents ,128 individual acts（ And its hierarchy , Quoted by at least one document ）.Hier-SPCNet The Communist Party of China has 22566 Nodes ,31309 side .PCNet There is the same 1806 Case document nodes and 542 Reference edges .
The document similarity label is marked by experts 100 For documents , Details .

4.2 The main experimental results

The evaluation index is Pearson correlation coefficient .

co-citation The value of is the same because in-citations identical （ Because the law does not quote documents ）.
Other analysis is omitted .

average and max It is to average or maximize the two similarity scores

5. Code reappearance

Wait until my server is ready .

---

1. node2vec: Scalable Feature Learning for Networks︎︎

2. Metapath2vec: Scalable representation learning for heterogeneous networks︎︎

3. Re6： Reading papers LeSICiN: A Heterogeneous Graph-based Approach for Automatic Legal Statute Identification fro︎

4. Similarity analysis of legal judgments︎︎︎

5. Finding Relevant Indian Judgments using Dispersion of Citation Network︎︎