Interpretation of Imputed Behavior in ALIBI (1 to 3) and SKILL

• Pagine: 213-241

Page 213

ALIBI is a program that analyzes an accusation and finds excuses (or, more generally, alternative explanations: not ´ALIBISª), ALIBI analyzes an accusation, decomposes actions imputed, and then tries to exculpate the defendant, by separating ascertained events from deontic (i.e, legal or moral) interpretations, and by recomposing the constitutive elements yielded by the recursive decomposition of the actions, into a less reprehensible plan of action. ALIBI2 has, as interface, a parser that allows the input accusation to be stated as a set of simple English sentences. ALIBI3 includes an explicit mechanism to calculate costs of excuses and parts thereof in terms of deontic liability, skill is a program that extends justification to areas other than in the legal domain: skill in performing at some task is judged according to common-sense knowledge (including widespread prejudices) about the task, on classes of performers, and on the environment

@1. Computer generation of exculpatory excuses

According to a Persian anecdote, a certain person was asked, one day, by one of Ms neighbors to lend him his rope. He replied, ´I have spread millet on itª. The neighbor wondered: ´How can one spread millet on a rope?ª The answer he got, was: ´For an excuse, any reply would doª.

Roles and addresses:

Dr. Ephraim Nissan (advisor & contact-author): Department of Mathematics & Computer Science, Ben-Gurion University of the Negev, P.O. Box 653 Beer-Sheva 84105, Israel

TSVI KUFLIK (ALIBI1): with the Israel Air Force (on study leave in the U.S.).

Gilad Puni (ALIBI 1): 17/14 Kallanit St., Kfar Saba

Roni Salpati (ALIBI 2): 36/16 Bea-Yair St., Arad 80700

Yuval Shaul (ALIBI2): 6 Ben-Yosef St., Bat-Yam 59401

Auni Spanioli (ALIBI3): St. 202/10, Nazareth 16000

Fadel Fakher Eldeen (skill): Majdal Shams, Golan Heights 12438,

Page 214

(Haim 1956: p. 34). Any excuse would do, because the rope owner expected no punishment would come to him, from not having his excuse believed. He was not standing trial for a punishable charge. A felon would be more wary on proposing an excuse, once he (or she) is caught. What makes for an apt excuse, as opposed to an insipid one? The excuse of the anecdote.

´I have spread millet on my ropeª, defies common sense, whereas

´I have spread millet on my sheetª,

or

´I have spread bird lime on my ropeª, would make sense, instead, from the viewpoint of physical equilibrium, as well as of uses found for objects in civilization.

In this paper, we are concerned with the generation of justification for actions that are imputed because reprehensible, ALIBI, a planner coded in Prolog, generates and proposes pretexts (in the following, improperly termed ´ALIBISª) for behavior reprehended and ascribed to a defendant, with whom ALIBI sides or identifies, (The malapropism ´ALIBIª in the name of the system has been deliberately adopted, out of the advertising tactics of selecting the most easily recognized or catchy term in a domain, as an attention-getter.) ALIBI is not meant to simulate the way an earnest lawyer would try to defend his (or her) client; certain deontic constraints on delegated legal defense are violated. We are specifically interested in alternative interpretations of behavior; excuses the way ALIBI generates them can be considered to be:

- false excuses that a guilty defendant may concoct by him- (or her ) self,

- or the actual way things happened, contradicting the incriminating interpretation,

Even delegated defense sometimes works that way, in very particular circumstances: one of the authors (E. Nissan) happened to describe the workings of ALIBI to a certain person who could tell, according to his own biography; a former Soviet physicist who now works on mathematical aspects of artificial intelligence in Tel Aviv, this man related that he had himself to perform the task of ALIBI: while still in Moscow, in less liberal times, as a refusenik, he had been put together with common criminals in prison, seemingly in order to break him psychologically. These inmates were ai too happy to have a scientist to find excuses for them, so he had to comply. A defendant is likely to invent excuses. A lawyer should not. A cell mate of a defendant may have to. However, a lawyer still has to manoeuvre in the relam of interpretations, and from both perspectives - of the legal application, and of artificial intelligence (ai) the mechanism ofPage 215 justification in general certainly deserves to be researched computationally. Actually, such a topic of inquiry is relevant to other areas of AI, such as the automated generation of narratives¹. Much research has been reported, in the literature, about expert systems embodying legal knowledge. ´One kind is based on a representation of the law- and attempts to support reasoning with the law, while the other kind is based on practical experience of legal practitioners. Following [Susskind 87] we shall call the former an "academic" system and the latter an 'experiential' system. (Schild 1989: Sec. 2). However, little has been done concerning the pre-legal phase of exculpation, which Is the application of ALIBI, ALIBI is not meant to simulate the wayPage 216 an earnest lawyer would try to defend his (or her) client: certain deontic constraints -on delegated legal defense are violated. We are specifically interested in alternative interpretations of partially observed behavior; excuses the way ALIBI generates them can be considered to be: (i) false excuses that a guilty defendant may concoct by him- (or her ) self; or (ii) the actual way things happened, contradicting the incriminating interpretation.

The output of ALIBI looks like almost natural English, and is composed out of patterns of sentences with variables. The input, instead, is a sequence of Prolog predicates according to a certain pattern, in the first version of ALIBI: a list of imputed actions is fed into the system; it may be conceived as a much simplified ´police reportª. The second version of the automated

Figure 1. Input analysis in ALIBI2

[ GRAPHICS ARE NOT INCLUDED ]

Page 217

exculpator, ALIBI2, has been upgraded with a user Interface, to enable the program to handle a subset of natural English. Moreover, In ALIBI2, the knowledge-base has been reorganized; what In the first version was the syntax of the input, has been made much more expressive, and has become an Internal representation; as more complex Input can be expressively afforded, now, heuristics can be added to the planner, to account for more complex situations. The parser, as presently structured, consists of the cascade of a merely syntactic phase, based on a generative grammar, and of another phase, where semantics comes In: an Instance of an action Is represented In a predicate, by having the parser fill values for deep cases (possibly with other kinds of Information, too) as arguments, for which a pre-defined order Is unnecessary. See Figure 1.

The following phases are shared by all versions of ALIBI. Actions Imputed are decomposed Into constituent actions, recursively according to commonsense knowledge stored in the knowledge-base. This process goes on until terminal actions are reached In the decomposition tree. The program spoils actions of connotational Interpretations pertaining to deontic modalities (that is, legality or morality). For example, stealing is reduced to taking as in given objective circumstances. Then, from the elementary actions, ALIBI recomposes another tree, to form an alternative Interpretation of the behavior Imputed, that Is legitimate or less reprehensible. For each Interpretation generated, ALIBI composes and displays a verbal statement exposing the excuse; simplified natural English Is generated by pasting canned text Including variables that are assigned the suitable lexical value. See Figure 2.

[ GRAPHICS ARE NOT INCLUDED ]

Page 218

Instead of just having incriminating excuses excluded by the planner, ALIBI3, developed in 1990, includes a mechanism that explicitly computes a numeric score of liability for plans, or parts thereof, being generated, (Another difference of ALIBI3 with respect to the previoys versions, is that these are in prolog, whereas ALIBI3 has been receded in Lisp).

The first version of ALIBI, and several directions for further development, have been described in a previous paper by Kuflik, Nissan and Puni (1989). In the present paper, instead, we are going to emphasize the general concepts, and to point out those representational choices that were introduced after the first version of ALIBI. Moreover, we are going to discuss justification in a broader context than the legal domain: in another program we developed in 1990, skill, human skill in performing at some task is judged (based on an input description) according to common-sense knowledge (including widespread social prejudices) about the task, about classes of performers (professionals in specific domains, youngsters versus adults, men versus women, etc.), and about the environment.

@2. Examples and structure of processing

While the first version takes, as input, for example:

done(rob,diamonds_pack,jeweller-s_shop)

done(injure,sub-machine_gun,jeweller).

done(break,body,display_window),

done(sneaks display_wimdow,jeweller-s_shop)

done(take,diamonds_pack,jeweller-s_shop).

ALIBI2 accepts the equivalent English-like statement:

the defendant robbed the diamonds_pack from a jewellers_shop.

he wounded the jeweller by a sub_machine_gun.

he broke the display_window with his body.

he sneaked into the jewellers_shop through the display_window.

he took_away the diamonds_pack from the jewellers_shop.

We intend to refine this, by allowing a more natural expression of nominal phrases, etc. As the input format is evolving, in the following we are going to expose examples, by stating their defining input informally.

Now, let us discuss the example. The defendant is charged with having robbed a pack of diamonds, at a jeweller's shop...