Insights from Jurisprudence for Machine Learning in Law

2 Overview Of Machine Learning In Law

Philipps (1989) was among the first to demonstrate the application of neural networks in law with a hypothetical example from Roman Law. The will of a hypothetical citizen whose wife was pregnant read thus: “If a son is born to me let him be heir in respect of two thirds of my estate, let my wife be heir in respect of the remaining part; but if a daughter is born to me, let her be heir to the extent of a third; let my wife be heir to the remaining part”

Philipps trained a feed forward neural network with backpropagation of errors to deliver the correct output when exposed to scenarios that involved the birth of a boy and of a girl but not both. He then put forward a case that necessarily defeats these rules; one in which twins, a boy and a girl are born. In this case, the network that had not been exposed to this scenario during training, produced an outcome that indicated the mother receives two shares, the son receives three and the daughter receives four. Philipps argued this outcome is reasonable in that it represents an equilibrium based on past cases. However, Hunter (1994) pointed out that the notion of equilibrium with past cases is jurisprudentially flawed. There is neither a notion of moral correctness nor any appeal to rationales that reflect higher principles.

Another instance of the application of connectionism for modeling defeasible rules in law can be seen in the work of Thagard (1989). He advanced a theory of explanatory coherence that modelled the way in which competing hypotheses are supported, to a greater or lesser extent, by available evidence. Some nodes in the network represent propositions that represent each hypothesis. Other nodes represent available evidence. Links exist between evidential nodes and hypothesis nodes, which have an associated weight. These weights may be excitary or inhibitory. To determine which hypothesis has more support, the network is activated. Nodes feed activation (or inhibition) to other nodes that feed back to each other until equilibrium is reached. The network is then said to be settled.