Respondent Moot Memorial including Charges and Argument Advanced.docx
Addressing the Great Bottleneck in Legal AI: Harnessing Law Students to Produce Open Access Human-annotated Data Sets
1. Addressing the Great Bottle-
neck in Legal AI: Harness-
ing Law Students to Pro-
duce Open Access Human-
annotated Data Sets
Arthur Dyevre
SAIL
1 June 2021
3. 1 Outline
1 The Challenge: Supervised Learning in Law
2 Solution 1: Small Is Beautiful
3 Solution 2: Legal Education and Legal AI
4 Conclusion: Clearing the Bottleneck
3 Harnessing Law School
4. Most important end-user tasks in legal AI require human-labelled
data:
I Information retrieval
5. Most important end-user tasks in legal AI require human-labelled
data:
I Information retrieval
I Document classification
6. Most important end-user tasks in legal AI require human-labelled
data:
I Information retrieval
I Document classification
I Argumentation mining
7. Most important end-user tasks in legal AI require human-labelled
data:
I Information retrieval
I Document classification
I Argumentation mining
I Overruling and jurisprudential change
8.
9.
10. 1 Where is the legal equivalent of ImageNet?
I Typical datasets in published research small
11. 1 Where is the legal equivalent of ImageNet?
I Typical datasets in published research small
I Often proprietary
12. 1 Where is the legal equivalent of ImageNet?
I Typical datasets in published research small
I Often proprietary
I Annotating legal documents is expensive and time-consuming
13. 1 Where is the legal equivalent of ImageNet?
I Typical datasets in published research small
I Often proprietary
I Annotating legal documents is expensive and time-consuming
I Requires domain expertise
16. 2 Outline
1 The Challenge: Supervised Learning in Law
2 Solution 1: Small Is Beautiful
3 Solution 2: Legal Education and Legal AI
4 Conclusion: Clearing the Bottleneck
10 Harnessing Law School
28. 3 Outline
1 The Challenge: Supervised Learning in Law
2 Solution 1: Small Is Beautiful
3 Solution 2: Legal Education and Legal AI
4 Conclusion: Clearing the Bottleneck
22 Harnessing Law School
42. "...this course was a very nice change from the classic practical
class. A regular seminar normally means doing assignments
and then it is done. This class however was a completely new
way of working. For once it felt like our work was serving
a purpose other than just grading. It felt like we were con-
tributing something to legal research. This was very new and
refreshing and gave me motivation for the class. Lastly, I be-
lieve that this class has made me a better jurist. We have had
to analyze judgments before in other classes, but they were
always handpicked decisions about a certain topic. This class
has given us a wider variety of judgments to read, no matter
the topic. Reading and annotating these judgments has given
me a better understanding of Belgian decisions in general."
44. Next step:
I Develop integrated app
I Broaden scope: contract clauses, terms of service...
45. Next step:
I Develop integrated app
I Broaden scope: contract clauses, terms of service...
I More tasks: entity recognition, arguments,
overruling/distinguishing...
46. Next step:
I Develop integrated app
I Broaden scope: contract clauses, terms of service...
I More tasks: entity recognition, arguments,
overruling/distinguishing...
I Create annotation-ready tasks
47. Next step:
I Develop integrated app
I Broaden scope: contract clauses, terms of service...
I More tasks: entity recognition, arguments,
overruling/distinguishing...
I Create annotation-ready tasks
I Scoring and certification
48. 4 Outline
1 The Challenge: Supervised Learning in Law
2 Solution 1: Small Is Beautiful
3 Solution 2: Legal Education and Legal AI
4 Conclusion: Clearing the Bottleneck
35 Harnessing Law School
50. 4 Conclusion
I Progress in Legal AI depend on advances in supervised learning
I Transformer and transfer learning are part of solution
51. 4 Conclusion
I Progress in Legal AI depend on advances in supervised learning
I Transformer and transfer learning are part of solution
I Opportunity to create win-win collaborations with law schools to
create new data sets
52. 4 Conclusion
I Progress in Legal AI depend on advances in supervised learning
I Transformer and transfer learning are part of solution
I Opportunity to create win-win collaborations with law schools to
create new data sets
I Thanks!