Generating Legal Arguments and Predictions from Case Texts | Stefanie Brüninghaus and Kevin D. Ashley

  • "automatically finding abstract,legally relevant concepts in case texts" (65)
  • "demonstrate how they can be used to make predictions of case outcomes" (65)
  • experiments with different document representations
  • findings w.r.t representation:
    • replacing names with roles in the case texts is useful for prediction due to better indexing
    • adding syntactic and semantic information ("propositional patterns")
    • concept extraction using nearest neighbours
  • use case:
    • 1) represent case
    • 2) predict outcome of a case given other cases
    • 3) explain the prediction

Given this abstract what interests me is that this project combines 1) document similarity with prediction with 2) explanation.

Introduction

"This paper reports an attempt to develop an AI and Law approach that can go “all of the way” from textual descriptions of a problem’s" (65)

Combines SMILE (SMart Index LEarner) and IBP (Issue-Based Prediction).

Representation

Task: "The focal task of the research reported here is automatically processing textual descriptions of the facts of legal cases and assigning appropriate indices from among a fixed set of indexing concepts." (65)

So note that the indexing concepts are fixed. I.e. we need a data set on a specific domain with an appropriate set of concepts. The data needs to be labeled with

Data: "From the CATO system (Aleven 1997), we have a knowledge base of 146 Trade Secret Law cases, represented in terms of Factors, together with squibs, summaries of the case facts." (65)

BoW-representation is common, but unlikely appropriate here:

A bag-of-words representation often works very well for relatively simply tasks; however, it does not seem ideal for finding complex and abstract concepts like Factors in legal case text (66)

A richer representation is desirable, but which one is an open question.

  • adding semantic information
    • helps generalize to text by different authors, courts, procedural settings, jurisdictions
    • Brüninghaus & Ashley 1999: incorporate thesaurus
  • Hypothesis 1: replace individual names with their roles, because they are often case specific and can lead to erroneous inferences (e.g. the name Smith occurring in different cases that are different and with different persons; Smith is just a common name)
    • roles are meaningful: plaintiff, defendant, product
    • But how do you replace this automatically?
    • Can we based on linguistic/syntactic information determine who does what?

Idea is to use linguistic analysis to 1) capture patterns of action and 2) capture negation.

Three used representations:

  1. BOW
  2. First replace names and case-specific info, then BOW
  3. Propositional patterns