ALICE in Cyberland:
Figure 1: How ALICE could be used to work out a problem
When adding information to the system, or when navigating through the links to solve a particular problem, the user will be assisted by prompts by the computer. In this example the task is to agree on a legal regulation concerning liability for damages caused by the release off genetically modified food into the environment. those sides agreed on imposing strict liability. However, and unknown to them, strict liability has a systematically different meaning in their relative legal cultures. As long as this does not cause any concrete problems however, this partial understanding is sufficient, and they can say to have agreed on the relevant regulation, but if a problem emerges, either side can question the meaning attributed to strict liability by the other side. Once the question is raised, the common ground has to be re-established. This will only be the case if both sides have no further questions.
Back to the example:
Y asked X what he means by strict liability. X then uses legal advice, for instance from a textbook. At this point, Y can agree, agree to disagree, or continue to question X on the meaning of strict liability. Only in the last case, he is then referred to further explanations, for instance the leading case. Again, he can now decide that he has now a sufficient understanding of the legal meaning attributed to strict liability in the foreign jurisdiction, and of the basis of this, renegotiate is the legal regulations on GM food, or else, he can request further information.
In our example, it might turn out that the difference between the two jurisdictions was historically caused by a small difference in the cultural background, say a different attitude to domestic animals. Having an industrial economy, damages caused by animals where so rare that it seemed unproblematic tot leave the loss always with the owner. Provided with this background information, Y can decide that the difference is after all immaterial for the case at hand. In this case, they can decide to retain the original formulation, despite the difference in interpretation discovered in the process (Clark and Gibbs, 1986). Harmonising laws need not mean to superimpose one identical meaning for all participating jurisdictions. Or they can decide to introduce an explicit definition of strict liability, which however exempts damages caused by animals. Either way, common ground has been re-established. During this process, the computer remembers that there are is still an open problem, a conceptual mismatch (Visser and Bench-Capon, 1998) and advises other users accordingly.
Conversely, in deciding a concrete case on the basis of the agreed formulation, the user would be guided from the text to more detailed explanations until she decides that for this purpose, she has gained sufficient understanding. The common technique in automatic translation is to provide user profiles, based on past experience with individual users. To some extent, this has also been the feature of Elenchos.
Let's assume that both sides in our example agreed on the formulation proposed initially, that is imposing strict liability without qualification, but bearing in mind that the two jurisdictions might interpret the relevant cloth differently. In our case, this interaction would result in two different profiles, one for the typical X-lawyer, one for the typical Y lawyer.
An X lawyer, searching for information which relates to his own jurisdiction only would not automatically be provided with any additional information. However, if he has to solve the problem which has a connection to the jurisdiction of Y, the system would automatically remind him of the potential for problems. Initially, it is assumed that ALICE will have predefined profiles, which rank for instance legal systems according to their remoteness. The user would identify himself for instance as coming from a common-law jurisdiction, and receive additional informative prompts if and only if he has to deal with a civilian jurisdiction. At a later stage, we will experiment with profiles for individual users.
Comparative law is obviously not the only subject that deals with the interpretation of texts from a different cultural context, and philosophy in particular has recently discovered the usefulness of computers and the internet to assist researchers in the interpretation of texts from a different cultural context. ALICE will build on the experience of two projects in particular, the Elenchos-study within the 'Project Archelogos' at Edinburgh University, and HSI (Hermeneutics, Semiotics, Informatics) at the Technical University Chemnitz.
HSI started as a colloquium organised by Ferdinand Fellmann, at the University of Chemnitz. Its aim is to:
'develop alternative methods for the understanding of texts, and to address the shortcomings of traditional hermeneutics through developments in computer technology' (Rolf 1998).
For our purpose, the most important aspect of this research is its attempt to start with an investigation of the concept of 'understanding' and its cognitive implications. This will lead to a pragmatic and operational notion of understanding which can be supported efficiently through the use of computer technology. This emphasis on the notion of understanding rather than knowledge representation or argumentation mirrors the shift indicated above.
Two of the findings of this research group will be directly relevant for ALICE. The first concerns the importance of 'background knowledge' for the interpretative task. Fellmann uses as example the task of interpreting Goethe's Werther, which requires an understanding of the knowledge available to the author and his average reader regarding e.g. sociology and psychology. Not in the sense of an objective account of the social conditions of the 18th century, but as an account of the subjective knowledge of the historical author.
For comparative law, these concerns are expressed by Legrand, who draws our attention to the need of a cultural background knowledge that transcends the information provided by the legal text. Again, this knowledge will typically be 'subjective' in the sense described here. If we want to understand, say, certain developments in the English law of real property, we can not measure it against a yardstick of 'objective' knowledge of the economic conditions of the 19th century - this would be Mattei's economic approach to comparative law - but we need to understand the economic knowledge and philosophy of 19th century England which would have informed judiciary and Parliament.
For research in comparative law, and even more, for concrete legislative projects based on findings from comparative law, this poses a serious problem. Since potentially, the most minor detail of e.g. 19th century culture might be necessary to properly understand a particular piece of legislation, the amount of knowledge required to co-ordinate the cross-border understanding of several jurisdictions presupposes the availability of such huge amounts of specialist expert knowledge that for mere reasons of logistics, the traditional methods of conferences and edited books for the acquisition and dissemination of it becomes infeasible. This has been indeed the experience of our collaborators from Mauritius.
Internet based systems offer an ideal solution to this problem. HTML allows to 'annotate' the legal provisions with 'interpretative' links to relevant sites, representing the collective world knowledge or 'horizon' for the interpretative effort. While it will be an obvious (and technically unproblematic) requirement for ALICE to be Internet-based, most of the 'blue sky research' associated with the project will develop techniques to deal with the unwelcome side effect of this solution, i.e. attempts to reduce the available knowledge to those parts which are relevant for a given query or problem.
A second, and more unusual, element of HSI is its idea that the interpretative act is not only supported by a (hidden) computer programme, but indeed means writing one (Fellmann 1999). ALICE will to a certain extend reflect this insight. While its users will not directly be asked to 'write codes', they will have the possibility not only to add to the database, but also to access the formal representation of this information, and to rearrange directly the structure ('ontology') of the knowledge represented.
The author's own experience with computer aided interpretation originated in his involvement with the Elenchos -project at the Department of Philosophy at Edinburgh University. Elenchos is a sub-study of Project Archelogos, which, while older and developed independently from HSI, can be seen as one of the most developed application for a computer-supported textual interpretation (Schafer, 2000).
Archelogos itself is primarily a 'dumb' database, with arguments of Platonic dialogues as objects, restructured by experts on Plato using a methodology which emphasises the logical structure of the arguments (Scaltsas, 1997). Hyperlinks link philosophical propositions to their supporting premises, allowing the reader to navigate through the internal structure of an argument. In addition, and here Archelogos comes closer to HSI, researchers are able to navigate from the original Platonic texts to various, often contradictory interpretations found in the secondary literature. These again are linked to comments on the validity of this argumentation by other scholars, and indeed back to the war with the Platonic text if a commentator supports his interpretation through a cross-reference to other textual passages.
From the beginning, the idea behind Archelogos was to use this material for teaching purposes, to help students appreciate the logical structure of philosophical arguments. LogAnalysis was a first study within the Archelogos frame to use the collected data for interactive tutorials. It has been specifically designed and created for university and advanced high school students. It provides an Archelogos-type of argument analysis, but also much more in terms of philosophical instruction in contemporary and classical Philosophy for the student. In order to attract the student's interest and provide the background for the philosophy of Plato it also contains a substantial cultural archive about the philosophy and the culture of the times that the dialogue was written. Again, this feature correspondence to the requirement of contextual background knowledge identified in HIS and in comparative law. ALICE will consequently use a similar multi-medial approach to annotate information. But unlike Elenchos, users will be able to create direct links to information they consider of interest, marking the move from a passive system to a dynamic one, and so addressing the problem of organising the necessary expert knowledge, identified above as one of the obstacles in the process of discussing legal harmonisation.
However, even LogAnalysis remained relatively 'dumb'. Feedback was limited, and students were confined to static representation of the Platonic arguments and a predefined set of questions and answers. However, it is able, within limits, to check if the student has properly understood the text at hand. If 'unconvinced', it will prompt further elucidations from the user, or provide further information. a similar quasi-dialogue in ALICE will allow, again severely limited, checks of understanding by the system. Where the likelihood of misunderstandings has been identified beforehand, ALICE will ask leading questions and, if the answers are mistaken, provide further information.
LogAnalysis links the various arguments from the primary text to several, often contradictory interpretations in the secondary literature. This corresponds in ALICE to the idea that a primary text, the agreed treaties, has links to the various interpretation it receives in the participating jurisdictions.
Elenchos, an expert system for Plato's political and legal philosophy, tries to incorporate a more dynamic and 'intelligent' use of the material. (see the Archelogos webpage) Its original brief was intentionally wide: To investigate the potential of expert system technology to offer both students and academic philosophers an optimised use of the Archelogos database. The result was a dialogue game not dissimilar to dialogue based expert systems for legal education. It allows students to have a (partially open ended) 'discussion' with the computer which represent the Platonic argument for democracy. The user will be able to question the system, force it to justify its answers, and can challenge it with (limited) counter-arguments. The user can also choose pre-formulated counter-arguments (for or against democracy) derived from the subsequent philosophical discussion, challenging e.g. Plato with an argument typical for Rawls of Nozick. The brief for Elenchos was wide:
To develop a model for the presentation of European cultural heritage in text using innovative artificial intelligence techniques so that the European cultural heritage will be accessible to both specialists and non-experts.
The inclusion of both specialists and non-specialists was of specific significance for ALICE. It turned out that certain features of the programme caused problems for the specialists only. Students typical struggle with the contend of the text, but are prepared to take the information at face value. For them, the feedback mechanism was satisfactorily reliable. i.e., we were able to foresee correctly likely problems and difficulties. Experts however would often mis-construct the question, or, in a rather diffuse sense, feel unable to answer it, and left feeling unhappy with the way the question was posed (see also on this Isaacs and Clark, 1987).
In these cases, there was typically a disagreement with the argument structure chosen by the project-designers, and the prior conceptualisation of the expert. Different ontologies and conceptualisations of user and system can result for instance in presupposition failure, both in the sense that a question posed by the computer carries a presupposition not shared by the user, or a presupposition of the user not being represented in the system's ontology. As a problem, this again mirrors the situation we face in cross-border legal studies - it is possible to educate any student in any legal system, but once the conceptual scheme pertaining in a given system is learned, the acquisition of a second system is fraught with problems.
Elenchos uses two strategies to deal with this problem: firstly, we encounter again 'user-profiles'-the system learns from the answers the user is giving in a dialogue game, and can adjust the 'difficulty' level, e.g. by displaying more or less additional information and background knowledge, or by more or less elliptical argument structures. This reflects the positive side of pre-knowledge; the expert user will get easily bored, and won't need information he considers as trivial. 'The obvious goes without saying', and a long as this does not cause problems, the pragmatic solution is to restrict again the amount of information provided.
When pre-knowledge gives rise to misunderstanding, instruments are needed to re-establish the lost common ground. In Elenchos, this happens by allowing the expert user access to the formal deeps-structure, where she can:
a) see explicitly the conceptual choices made by the system designers; and
b) can alter these conceptualisations by providing alternative models.
In the Elenchos example, she can for instance change the status of a proposition X from 'supporting proposition Y' to 'is supported by proposition Y' if she disagrees with our analysis of the argument structure.
We will discuss a short legal example for ALICE below, after we have discussed the nature of this 'formal deep structure'. However, we should note already here that this is indeed the core of ALICE - the ability to deal with more than one conceptualisation at the same time is the one distinguishing feature of all the applications we have intended, from abstract research in comparative law, to the solving of IPL problems, to the design of unified legal frameworks which replace national legislation.
Clark's research and its further development in the theory of grounding and computer modelling of the emergence of understanding forms the theoretical foundations of ALICE. Little has been said so far however about the technical solutions which corresponds to the findings of this approach. Concluding, I want to give a short outlook at two formal approaches in particular which not only seem to match most closely the theoretical requirements following from research in grounding and the theoretical work on comparative law, but which also show avenues for overcoming some of the problems encountered in HIS and Project Archelogos.
'Active logic' is the name of a family of inference engines developed at the University of Maryland by, amongst others, Don Perlis (Gurney et al, 1995) The distinguishing feature is that they incorporate a history of their own reasoning as they run, making them more flexible than traditional AI systems and therefore more suitable for commonsense reasoning.
Common sense agents are defined as:
'people who are not necessarily very good at any particular task but who are able to maintain a focus, an assessment of what the task or topic is, or that is has changed, or that it is not clear, that help is needed, that the task should be given up etc. A familiar human setting is that of trying to follow an expert discourse, or one in a foreign language, and missing in a lot of the details'.
Already this definition coincides obviously with the scenario ALICE has to deal with. Typically, lawyers representing one country will listen to the expertise provided by the colleagues from the other participating member states. Initially, they will rather crudely try to understand the new information within their own conceptual framework, adjusting as they progress, and eliminating initial mistakes in light of new evidence.
Traditional AI systems showed a lack of flexibility in the face of nonsense, contradictions, and conceptual mismatch. According to the active logics group, this mistake is the result of the inability of traditional systems to use the knowledge they have to recover from their mistakes. Recovery from mistakes is a separate process from reasoning about the world and therefore cannot make use the power of the inference machine and its world knowledge (Perlis et al, 1998).
Active logic on the other hand is able to reason about its own believes and mistakes in the same way it reasons about the external word. Intended applications in particular are misidentifications, contradictions, context shifts, new words, new meaning for old words, and believe revisions. Again, most of these problems will be encountered by people engaged in the process of negotiating a new harmonised legal order. Initially, they will encounter new words describing legal concepts of a foreign jurisdiction, which they might misidentify at first, only to realise the initial 'mistake' if this initial hypothesis results in contradictions (this is what Salter described as the dialectical procedure of comparative law). 'Mistakes' can be tolerated, and are indeed 'encouraged', if this does not result in the breakdown of communication ( Perlis and Purang, 1996). Eventually, once the new legal text is agreed, it might well use concepts which also can be found in the legal systems of the participating countries, but with a changed meaning. In return, the new supra-national law will feed back into the participating legal systems, changing the meanings of the concepts which gave originally rise to it (Schafer and Bankowsi, 2000).
Some of the things Alice has to do, and active logic promises to offer, are:
Noting and resolving identification error;
Noting and resolving contradictions;
Learning new words and new meanings for old words.
A particularly advanced application for active logic is the problem of presuppositions. Starting, as a we did, with the emphasis on the user and his problems, the active logic group noted that the standard solution of Prolog, to interpret failure as negation will often result in presupposition errors. Presuppositions, as we noted above, play a crucial role as a source of understanding and misunderstanding alike. The user used to the conceptual frame of his legal system will typically assume that the foreign legal system has equivalent concepts and ideas. Initially, as a hypothesis, this assumption is necessary and indeed helpful. Only if these presuppositions or prejudices fail will we run into trouble if we are not flexible enough to adjust the initial assumption.
Traditional expert systems where rather inflexible in this regard. When for instance sending a query whether a specific judgement relies on the notion of the precautionary principle, the traditional database would interpret the failure to recognise this concept simply as a negative sentence: 'no, the precautionary principle was not used', but it would not automatically alert the user that the precautionary principle does indeed not even exist in the jurisdiction in question. Active logic is able to make this distinction and therefore offers an important contribution to the avoidance of culturally induced misunderstandings (Gurney et al, 1997).
While the inference engine will make use of active logic, the even more important question for ALICE is how to represent the legal information. In the same way as the emerging theory of grounding provides an interdisciplinary exchange where computer science, comparative law, linguistics, and cognitive science come from very different directions to the identification of similar problems and solutions, so does the emerging field of formal ontologies.
I argued for the importance of formal ontologies for comparative law and questions of legal harmonisation in detail elsewhere (Schafer, 1988). For our purposes here, it is sufficient to note that most of the traditional applications of formal ontologies have a direct bearing on most of the problems discussed here. Uschold (Uschold and Gruninger, 1996), for instance, writes in his introductory text:
'People, organisations, software systems must communicate between and among themselves. However, due to different needs and background contexts, there can be widely varying viewpoints and assumptions regarding what is essentially the same subject matter. Each uses different jargon, each may have differing, overlapping and/or mismatched concepts, structures and methods. The consequent lack of shared understanding leads to poor communication within and between these peoples and their organisations. The way to solve these problems is to reduce or eliminate conceptual confusion and come to a shared understanding. Such an understanding can function as a unifying framework for the different viewpoints'.
Ontologies are particularly helpful in situations where a computer application requires the use of multiple categorisations. This is in practice frequently the case. The new expert system to support courts in deciding the likelihood of re-offending of offenders for instance will receive information from social services, the police, the prison system and psychologists. All of them will use computer systems with embedded conceptual assumptions which might or might not be compatible with those used by the other services. Ontologies serve to make these assumptions explicit by identifying the logical connections between elements across models of the system.
The focus on shared understanding, which we have already encountered in Clark's work, is by no means coincidental. Simon Winter's paper on expectation and meaning, which informed much of the discussion above, cites for instance, the 'part- whole relation', as one of the global properties of lexica which are a prerequisite for establishing common ground, a universally shared expectation. Part-whole relations are then again at the very heart of many systems of formal ontologies, most notably mereology.
The process of legal harmonisation then can be seen as ontological engineering, the establishing of a shared understanding of some domain of interest. The vocabulary of the emerging legal text and the ontology it implies (i.e; its semantic) functions as an inter-lingua.
If we remember the analysis of the necessary steps which precede an agreement of the new treatise, we will note that two stages can be distinguished. The first consists in understanding, as much as possible and necessary, the legal framework of the other participating countries. Secondly, to agree on the basis of this understanding a new legal framework, the harmonised law. The first stage corresponds what in ontology oriented programming would be called ontology based information gathering, the second is similar to the process of ontological integration (Vimercati, et al, 1998). Here, a new standard (for instance on metatags for webpages) is agreed, which retains as much as possible of the original, heterogeneous and localised conceptualisations, so as to be better able to deal with the data inherited from them. Quite often, these new standards will have the form of an explicit agreement (see e.g. Benjamins and Fensel, 1998).
Ontologies are sufficiently flexible to deal with both stages of the negotiation process. The first, establishing shared understanding, employs what Guarino called 'fine grained ontologies', i.e. ontologies which axiomatically describe the intended models in a very precise way (Guarino, 1998). For the newly created legal text, a coarse grained ontology is sufficient, as this allows the participating countries, within certain parameters, to se the harmonised law as natural extension of their own legal system - even if these systems would contradict each other. This, highly desirable, result is due to the fact that coarse grained ontologies underspecify their intended models and are therefore only suitable as means of communication if there is at least some shared understanding of what these models should look like.
To choose ontologies as means of representing the legal conceptualisations of the participating lawyers, and to integrate them into a new, 'harmonised' ontology, has a number of consequences for the project:
Firstly, it means that the legal provisions in question are not represented as sentences, but as objects. Law is seen not a system of rules, but as a network of relations between objects, e.g. the (formal structure of) a contract object and an 'environment object', linked through a formal relation.
Secondly, the emphasis is on the semantics, not the syntax of the texts. HIS introduced the idea that writing a small programme should be part of the interpretative process. However, these programmes were understood syntactically. The original text found a syntactic representation in a logic based computer language, which in turn was manipulated to produce an interpretation. It seemed difficult, if not impossible, to bridge the gap between these syntactic rewriting operations, and the semantical dimension of the text (Fellman, 2000). Since ALICE chooses from the beginning a semantic approach to represent the natural language law texts, this gap does not arise.
Thirdly, ontological representations can be easily transformed into graphical images. The user therefore will not write code, but simply move around symbols and boxes. No logical or computer training will therefore be necessary.
I refer the reader for a more detailed understanding of ontologies to the burgeoning literature in this field. However, a small example might be helpful to understand the potential of this methodology.
Johansen and Wohed (1998) discuss an ontology based deontic specification pattern to classify legal documents and facilitate information retrieval. Similar structures will be used by ALICE to represent the underlying conceptualisation, the 'meaning' of the legal texts. One structure they introduce is a deontic object tree, which represents top level ontological relations in a legal domain. One of the nodes of this tree (the representation of the formally introduced 'specificity-relation') distinguishes between revocable and non-revokable obligation (objects). Revocable obligations can be subdivided into those which are revocable by both parties, and those which are revocable by only one party. An example for the latter, in Sweden, is citizenship.
The resulting tree looks therefore like this:
Figure 2: Semantic tree of citizenship
However, lawyers from Germany or Greece are a used to system in which citizenship cannot be revoked. In the discussion with the Swedish colleague, this can be the cause for misunderstanding., and ontological mismatch (Visser et al, 1998) Once the reason for the problem is understood, the German lawyer can select the citizenship box and move it to right branch of the tree. ALICE then creates an alternative ontology and keeps a record of this is the potential problem. Again German lawyers would not normally be informed about this conceptual distinction by ALICE as long as they do not deal with a legal system that adopts the Swedish conceptualisation.
Another example. Lawyers from civilian jurisdictions are used to three branches of courts, civil, criminal, and administrative. English and Scots lawyers however do not distinguish in procedural law between private law and administrative law cases. A French lawyer might want to know if there are any decisions by the administrative courts in England on the precautionary principle. Environmental issues typically would be dealt with in France by an administrative court. If using ALICE, the active logic inference engine would insure that in this case, he receives an intelligent and supportive feedback. He would be informed that his query was unsuccessful not because English administrative courts do not recognised the precautionary principle but because there are no designated administrative courts in England. He could then access the ontology which represents the English court system. Its binary divide between private and criminal courts could then be extended by a third branch. This branch could then lead via a hyperlink to supporting texts which would help the user to understand:
a) why the two ontologies differ; and
b) which of the two branches from the English system comes closest to the third branch in the civilian system, allowing a sufficient, if ultimately faulty understanding.
Findings in comparative law on the methodological assumptions of cross-cultural legal understanding mirror closely findings in the theory of grounding. The convergence of results and problems from comparative law, linguistics, cognitive and computer science make it promising to use computer implemented solutions to problems developed by the letter to overcome perceived shortcomings of the former. Alice, using results from similar systems developed in philosophy, will explore in particular the possibility to develop an ontology based expert system which combines a range of technical solutions to minimise the danger of misunderstanding and to maximise the availability of expertise to people engaged in the process of creating transactional legal treaties. If successful, legal AI would have finally identified a field with considerable growth prospectus, and a considerable advantage over its human competitors.
a) books and articles
Aikenhead, M (1986), 'A Discourse on Law and Artificial intelligence', Law Technology Journal, <http://www.law.warwick.ac.uk/ltj/5-1c.html>.
Belkenius, C (1996), 'Generalisation In Instrumental Learning', in, From Animals to Animals 4', Proceedings of the 4th International Conference on Simulation and Adaptive Behaviour, P Maes (ed.) MIT Press, Cape Cod, 1996.
Cole, M (1985), 'The Zone of Proximal Development - Where Culture and Cognition Create Each Other', in, Wertsch, J (ed), Culture, Communication and Cognition - Vygotskian perspectives. Cambridge, Cambridge University Press 1985, pp.146-161.
Kraus, R M, Fussell, S R (1990), 'Mutual Knowledge and Communicative Effectiveness', in Galegher et al (eds.) Intellectual Teamwork - Social and Technological Foundations of Co-operative Work, Hillsdale, Erlbaum Ass, 1990 pp.111 - 145.
Mena, Kashyap, Illarramendi, Sheth (1998), 'Domain Specific Ontologies for Semantic Information Brokering on the Global Information Infrastructure', in, N Guarino, (ed.) Formal Ontology in Information Systems, Amsterdam, IOS, Amsterdam, 1998.
Winter, S (1998),'Expectations and Linguistic Meaning', Ph.D. thesis, Cognitive Science Department, Lund University, 1998, < http://www.lucs.lu.se/People/Simon.Winter/Thesis/.
Open law: <http://eon.law.harvard.edu/openlaw/>
Project Archelogos : <http://www.Archelogos.phil.ed.ac.uk/>