Introduction to Legal Technology, lecture 2 (2015)

TLS0070 Introduction to Legal Technology

Lecture 2 Artificial intelligence and law: the 20th century University of Turku Law School 2015-01-27 Anna Ronkainen @ronkaine [email protected]

Computing prehistory in general (20th C and before)

Vaucanson automata -  Jacques Vaucanson (1709–1782), France -  1737: The Flute Player -  1738: The Tambourine Player, The Digesting

Duck

-  1745: the first completely automated loom

The Jacquard loom -  Joseph Marie Jacquard (1752–1834), France -  developed Vaucanson’s loom further by

making it programmable -  exchangeable weaving patterns input using

punched cards

Babbage Analytical Engine - Charles Babbage (1791–1871), England -  designed the Difference Engine for

tabulating polynomial functions -  based on it, designed the Analytical Engine,

a mechanical general-purpose computer -  none were built at the time

The first programmer - Ada Lovelace (1815–1852), England - wrote the first programs written specifically

for the Analytical Engine (which did not exist at the time), generally considered the first computer algorithms

Algorithm for computing Bernoulli Numbers on the Analytical Engine (page 1), Ada Lovelace, 1843

Hollerith tabulating machines -  Herman Hollerith (1860–1929), USA -  generalized the use of punched cards into increasingly

general-purpose mechanical data processing (1886-) -  founded the Tabulating Machine Company (-> IBM) -  widespread use of Hollerith machines across many

applications through the 1st half of 1900s -  downside: Hollerith machines facilitated the Holocaust

(and afterwards gave rise to data protection legislation)

ENIAC and the end of the mechanical age -  the first electronic general-purpose

computer -  built at Penn in 1943–1946, commissioned

by the US Army -  initial use: calculating artillery trajectory

tables -  17468 vacuum tubes, 65 m3, 150 kW -  data input/output with punch cards,

programmed with rewiring

Things start getting smaller: On to semiconductors -  transistor developed in 1947 by Bardeen,

Brandain and Shockley (Bell Labs) -  first transistorized computer built in

Manchester 1953 -  first integrated circuit constructed in 1958:

Jack Kilby (Texas Instruments) -  first microprocessors in 1971 (Garrett CADC,

TI TMS 1000, Intel 4004)

The beginnings of data processing in law

The first search-and-replace ever: s/retarded child/exceptional child/g -  terminology change in the Pennsylvania health

code in the late 1950s -  legislative technique required all instances of

textual changes to be enumerated individually -  the legislature turned to prof Horty at Penn -  first tried to solve this manually, too unreliable -  solution: input text into computer, index the

position of each word to find all occurrences of the word in question

-  obviously generalizable into textual information retrieval in general

Next steps -  M.U.L.L. (later Jurimetrics) journal 1959– -  case law retrieval experiments by Colin Tapper

(Oxford) through the 1960s -  Centre d’études pour le traitement de

l'information juridique (IRETIJ, Montpellier) 1965

-  CREDOC (Belgium) 1967 -  OBAR (Ohio) 1964 -> LEXIS 1970 -  NORIS (Norway) 1970 -  Westlaw 1975

First expert systems: mid-1980s -  inspired by systems from other fields (e.g.

MYCIN) -  Latent Damage Law (Susskind and Capper) -  British Nationality Act (Bench-Capon and

Sergot) -  SHYSTER (Popple)

Where did all the lawyers go? -  the PC revolution (1980s) and the launch of

the commercial Internet (1993) -> computer-related legal problems everywhere!

-  expert systems were considered a failure – not just in law – for good reason -> the AI winter of late 1980s

-  leaving the field to computer scientists and legal theorists made AI & law

Major threads of AI & law research (non-exhaustive)

Information retrieval (1-st gen) -  normal database search (exact match or

wildcard characters) -  Boolean search operators - modest practical advances since the 1980s

(with some recent exceptions) -  legal AI contributions negligible

Administrative automation -  has been with us since the 1960s (or 1890s if you count

the use of Hollerith machines for the US census...) -  an absolute must for effective administration on a large

scale -  works well if the rules to be applied are straightforward

enough (rather hopeless with discretionary rules) -  seems that implementing new rules in these kinds of

systems is still a major PITA -  (also an occasional subject of doctrinal work in

administrative law, rule-of-law issues etc., e.g. Kuopus 1988)

Expert systems -  a big thing in AI in the 1980s -  basic idea pretty straightforward: -  you take an expert in some domain (e.g.

some area of law) -  make them turn their domain expertise into

computable rules -  add a reasoning engine -  and voilá, you have a computer giving

expert advice or making expert decisions

Example: British Nationality Act 1-[1] A person born in the United Kingdom after commencement shall be a British Citizen, if a t the time of birth his father or mother is:

a) a British Citizen, or b) settled in the United Kingdom

Represented as Rule 1: X acquires British citizenship on date Y

IF X was born in the UK AND X was born on date Y AND X is after or on commencement of the act AND X has a parent who qualifies under 1.1 on date

Rule 2: X has a parent who qualifies under 1.1 on date Y IF X has a parent Z AND Z was a British citizen on date Y

Rule 3: X has a parent who qualifies under 1.1 on date Y IF X has a parent Z AND Z was settled in the UK on date Y

Expert systems work (sort of) -  if the legal rules are straightforward enough: -  no ambiguity or vagueness regarding the inputs -  clarity about which rule applies in each situation

-  even in the best case, formalization of rules is far from trivial (knowledge-acquisition bottleneck)

-  also requires expertise on what to model and what to leave out (and how to make sure the system isn’t used beyond its design limits)

-  how much of the expertise really lies in the system and how much in the user?

-  in a sense, expert systems are doing just fine, it’s mainly the term that’s fallen into disuse...

Case-based reasoning -  one possible approach: analyze legal cases in

terms of factors (very common in US doctrine)

-  use factors to find best match for case at hand

- map factors into a network to find

Soft computing: Fuzzy logic and neutral networks -  both highly fashionable in AI in the 1980s -  also some experiments within legal AI in the

early 1990s -  fuzzy logic was also popular among legal

theorists (mostly on a metaphorical level) since Reisinger 1972

‘We suggest that fuzzy logic is no more than (over)sophistication of the approximation approach, that it may give good results in some very special applications, but its philosophical basis is uncertain generally and very uncertain when applied to open-textured legal concepts. Both the appearance of precision and the appearance of generality are spurious.’ (Bench-Capon and Sergot 1985/1988)

Your basic neural network

Ontologies -  the philosophical meaning of ontology: the

study of the nature of being (what is and isn’t)

-  in computer science: a way of formalizing entities in an universe of discourse (concepts and their relationships etc.)

-  the Semantic Web (Berners-Lee et al 2001) - Cyc 1984– (OpenCyc 2002–) - WordNet 1985–

Ontologies contain (in very general terms) -  individual entities -  classes of entities -  attributes for entities -  relations between entities -  function terms -  restrictions -  rules -  axioms -  events (changes to entities)

Ontologies in law -  Valente’s functional ontology (1995): -  norms (normative knowledge) -  things, events, etc. (world knowledge) -  obligations (responsibility knowledge) -  legal remedies (reactive knowledge: penalties,

compensation) -  rules of legal reasoning (meta-legal knowledge,

e.g. lex specialis) -  legal powers (creative knowledge)

-  (and several others)

Segment from the E-Courts ontology

(Breuker et al 2002)

E-courts top-level ontology

(Breuker et al 2002)

Use of ontologies -  always exist in a specific context, built for that

(no Begriffshimmel and no point in aiming for one)

-  can be generated by hand or by machine -  two very different ontologies can work just as

well (no Right Answer!) -  very useful for information retrieval (find similar

things that are called something else) -  can also be used e.g. for similarity metrics -  categorization hierarchy also interesting from a

cognitive perspective (basic-level concepts etc.)

Argumentation: Wigmore (1905)

Argumentation frameworks (Dung 1995) -  a set of arguments, and attack relations

between pairs of arguments (A attacks B) -  general semantics for argument trees -  plus specific rules for finding which attack

relation dominates (in case of conflict)

Pros and cons -  argument maps can illustrate how things are

made (and sometimes also show that some valid arguments are actually always ignored)

-  easier as a theoretical than a practical exercise

-  a lot easier when you already have a decision and have to find a matching argument scheme

Questions?

All images PD or CC-BY-2.0 Wikipedia unless otherwise indicated, see the obvious Wikipedia pages for details