Re21: Read the paper MSJudge Legal Judgment Prediction with Multi-Stage Case Representation Learning in the Real

Gods are silent-personal CSDN blog directory

Paper name: Legal Judgment Prediction with Multi-Stage Case Representation Learning in the Real Court Setting
Paper SIGIR official download address: https://dl.acm.org/doi/abs/10.1145/3404835.3462945 (there is an official SIGIR explanation video, this guy’s accent is not heavy, but his voice sounds very tired and hard to hear,After overdrawing the body...)
Thesis ArXiv download address: https://arxiv.org/abs/2107.05192
Official GitHub project: mly-nlp/LJP-MSJudge

This article is a 2021 SIGIR paper.The author is from Peking University and Ali.

This paper believes that the previous LJP work paradigm is too simplified for real scenarios, so this paper proposes a new task and corresponding data set that is more in line with the real situation:
The task is to simulate the real court trial scene, through the plaintiff's claim and legal debateThe content implements the LJP task (three classification task).The auxiliary task is to separate the facts from the debate (classification task, whether certain fact labels appear).
Propose a new dataset LJP-MSJudge, which contains plaintiff’s claims and court debate data (multi-role dialogues of the court debate) (the result of judgment in civil cases is whether various claims are supported).

1. Background

case life-cycle information

Difficulty:

1. The lexical space of different roles may be different, and it is difficult for traditional NLP algorithms to consume this.
2. The gap between the parties' statements and the facts identified by the final judgment.

2. Model MSJudge

Multitasking
MSJudge: Simultaneously identify legal facts from court debate and predict judgment result for each claim

(I am a little curious about the extracted fact part here, can I use the factual description text in the final judgment to do teacher forcing?)

Visualize interactions between components ("debate and fact", "fact and claim" and "across claims")

Multi-Stage Context Encoding: imitating judges to understand court debate and pre-trial claims
Debate Utterance Encoder: word embedding + role embedding (random initialization, joint training)→Bi-LSTM+attention→utterance embedding
Debate Dialogue encoder: Bi-LSTM, modeling to obtain the global representation of utterance
Pre-trial Claim Encoder: Bi-LSTM+attention (debate and claim shared word embedding matrix)
Multi-Stage Content Interaction: Modeling debatesand claims, facts and claims, claims, strengthen the claim representation
  Debate-to-Claim
  Debate-to-Fact
  Fact-to-Claim
  Fusion
  Across-Claim
Fact Recognition and Judgment Prediction

3. Experiments

1. word embeddings (skip-gram) and role embeddings dimension: 300
   Bi-LSTM hidden layer dimension: 256
   Adam learning rate 0.001
   batch size 16
   dropout rate 0.8
2. grid search tuning method and cross-validation
3. Add all the debates to each claim and make predictions

Others are omitted, to be added.