Re19:读论文 Paragraph-level Rationale Extraction through Regularization: A case study on European Court

简介: Re19:读论文 Paragraph-level Rationale Extraction through Regularization: A case study on European Court

1. Background


  1. rationalization by construction方法论:直接用constraint来正则化模型,对模型决策基于正确rationales的情况给与reward,而非事后根据模型决策结果推理可解释性

the model is regularized to satisfy additional constraints that reward the model, if its decisions are based on concise rationales it selects, as opposed to inferring explanations from the model’s decisions in a post-hoc manner

  1. 可解释性的意义:right to explanation
  2. 执法过程:

image.png


2. 模型


2.1 Novelty

  1. previous work on word-level rationales for binary classification→paragraph-level rationales
  2. 第一个在端到端微调预训练Transformer模型中应用rationale extraction的工作
  3. 不需要人工标注的rationales


2.2 模型

constraint:以前就有的sparsity, continuity(实验证明无效), and comprehensiveness(需要根据multi-label范式进行修正),本文新提出的singularity(能提升效果,而且鲁棒)

baseline HIERBERT-HA:text encoder→rationale extraction→prediction

image.png


在视频中放的图是:

image.png


词级别的正则器

①分别编码每个段落:context-unaware paragraph representations

②用2层transformer编码contextualized paragraph embeddings

③全连接层(激活函数selu)

K→用于分类

Q→用于rationale extraction→每个段落分别过全连接层+sigmoid,得到soft attention

scores→binarize,得到hard attention scores

④得到hardmasked document representation(hard mask+max pooling)(不可微,所以有一个训练trick)

⑤全连接层+sigmoid

baseline HIERBERT-ALL:不mask事实

constraint:

①Sparsity:限制选择出的事实的数目

②Continuity:于本文模型无用,但还是实验了

③Comprehensiveness:留下的段落生成的结果比扔掉的要好多少,或者比较两种段落的余弦相似度

④Singularity:选出的mask比随机的要好

Rationales supervision:noisy rationale supervision

image.png

image.png


3. 实验


3.1 数据集

提出ECtHR数据集,英语案例文本,silver/gold rationales,事件有时间顺序,决策包括违背的法条和援引的先例


3.2 实验设置

超参数:

image.png

网格搜索,Adam,学习率2e-5

贪心调参

LEGAL-BERT-SMALL:

50 paragraphs of 256 words


3.3 实验结果

指标:

micro-F1

Faithfulness: sufficiency and comprehensiveness

Rationale quality: Objective / subjective (mean R-Precision (mRP) Precision@k)

image.png

image.png

image.png

image.png


4. 代码复现


等我服务器好了再说。

相关文章
|
3月前
|
算法 数据挖掘
文献解读-Consistency and reproducibility of large panel next-generation sequencing: Multi-laboratory assessment of somatic mutation detection on reference materials with mismatch repair and proofreading deficiency
Consistency and reproducibility of large panel next-generation sequencing: Multi-laboratory assessment of somatic mutation detection on reference materials with mismatch repair and proofreading deficiency,大panel二代测序的一致性和重复性:对具有错配修复和校对缺陷的参考物质进行体细胞突变检测的多实验室评估
32 6
文献解读-Consistency and reproducibility of large panel next-generation sequencing: Multi-laboratory assessment of somatic mutation detection on reference materials with mismatch repair and proofreading deficiency
|
7月前
|
机器学习/深度学习 自然语言处理 算法
[UIM]论文解读:subword Regularization: Multiple Subword Candidates
[UIM]论文解读:subword Regularization: Multiple Subword Candidates
60 0
|
数据挖掘
MUSIED: A Benchmark for Event Detection from Multi-Source Heterogeneous Informal Texts 论文解读
事件检测(ED)从非结构化文本中识别和分类事件触发词,作为信息抽取的基本任务。尽管在过去几年中取得了显著进展
72 0
|
机器学习/深度学习 数据采集 人工智能
|
机器学习/深度学习 自然语言处理 计算机视觉
【计算机视觉】MDETR - Modulated Detection for End-to-End Multi-Modal Understanding
对于图像模型,MDETR采用的是一个CNN backbone来提取视觉特征,然后加上二维的位置编码;对于语言模态,作者采用了一个预训练好的Transformer语言模型来生成与输入值相同大小的hidden state。然后作者采用了一个模态相关的Linear Projection将图像和文本特征映射到一个共享的embedding空间。 接着,将图像embedding和语言embedding进行concat,生成一个样本的图像和文本特征序列。这个序列特征首先被送入到一个Cross Encoder进行处理,后面的步骤就和DETR一样,设置Object Query用于预测目标框。
《Towards A Fault-Tolerant Speaker Verification System A Regularization Approach To Reduce The Condition Number》电子版地址
Towards A Fault-Tolerant Speaker Verification System: A Regularization Approach To Reduce The Condition Number
89 0
《Towards A Fault-Tolerant Speaker Verification System A Regularization Approach To Reduce The Condition Number》电子版地址
|
自然语言处理 算法 数据可视化
Re21:读论文 MSJudge Legal Judgment Prediction with Multi-Stage Case Representation Learning in the Real
Re21:读论文 MSJudge Legal Judgment Prediction with Multi-Stage Case Representation Learning in the Real
Re21:读论文 MSJudge Legal Judgment Prediction with Multi-Stage Case Representation Learning in the Real
|
自然语言处理 Oracle 关系型数据库
Re32:读论文 Summarizing Legal Regulatory Documents using Transformers
这篇文章提出了一个英文法律规范文件摘要数据集。模型就是很简单地把抽取式摘要建模成每一句的二分类任务,还测试了在此之前用TextRank先抽取一遍的效果。(这个指标甚至没有做人工的) 看起来非常简单,这样就能发SIGIR吗,那我怎么不行…… 所以可能本文的贡献重点在数据集上吧!
Re32:读论文 Summarizing Legal Regulatory Documents using Transformers
Re1:读论文 C&S (Correct and Smooth) Combining Label Propagation and Simple Models Out-performs Graph Ne
Re1:读论文 C&S (Correct and Smooth) Combining Label Propagation and Simple Models Out-performs Graph Ne
Re1:读论文 C&S (Correct and Smooth) Combining Label Propagation and Simple Models Out-performs Graph Ne
PAT (Advanced Level) Practice - 1118 Birds in Forest(25 分)
PAT (Advanced Level) Practice - 1118 Birds in Forest(25 分)
112 0