A Preferential DL Approach to Model the Non bis in idem Principle for the Legal Domain

Tracking #: 2447-3661

Cleyton Mario de Oliveira Rodrigues
Fred Freitas
Ivan Varzinczak1
Italo Oliveira

Responsible editor: 
Guilin Qi

Submission type: 
Full Paper
Description Logics (DLs) are a family of formalisms that emerged to balance the trade-off between expressiveness and decidability for classical monotonic logic. Often, the research developed under the umbrella of AI & Law has relied on full synergy with DL to support argumentation reasoning, decision systems, legal compliance checking, and axiomatization of rationales and assumptions in the legal domain. Nevertheless, in many legal scenarios, regulations are defeasible. Inferences within the legal field are not purely deductive in nature, but retractable and ampliative since generalizations mostly hold for normal or typical cases. This is absolutely true in the criminal domain, where general criminal types are usually described in the caput of the norms (e.g., a robbery), and other specific types unfold from these (e.g., robbery followed by death, which is known as “Latrocínio” in Brazilian Criminal law). Although the classical subsumption relation may seem a correct way to model the hierarchy of laws at first glance, if no contradiction arises between the more general and more specific, what it should be pointed out that the penalty for specific crimes cancel out the penalties foreseen by the more general laws. In other words, a hierarchy of norms must not rely on classical subsumption relation; instead, a non-monotonic approach suits better in this setting. Therefore, in this paper, we show that Preferential DL, a defeasible version of Description Logic, is better suited than classical DLs for a faithful representation of the content of legal regulatory knowledge; in particular, w.r.t. the representation of the principle of Double Jeopardy (a.k.a. Non/Ne bis in idem). In this paper, we make the case for the application of ontologies represented in defeasible, preferential DLs for modelling laws and penalties. Our solution focuses on overrule relations to organize a set of defeasible axioms in terms of specificity criteria.
Full PDF Version: 

Major Revision

Solicited Reviews:
Click to Expand/Collapse
Review #1
Anonymous submitted on 14/May/2020
Minor Revision
Review Comment:

The paper proposes an ontological approach with Preferential Description Logics (DL) to model the criminal types of Brazilian so that the reasoning can be compatible with Double Jeopardy principle (a.k.a. Non/Ne bis in idem). The authors present conceptualizations of some crimes (e.g., Theft, Robbery, Robbery followed by Death) from the Brazilian penal code through a slight modification in the implication operator of DL. Compared with original DL, Preferential DL is better suited than classical DLs for a faithful representation of the content of legal regulatory knowledge such as the representation of Non/Ne bis in idem. Besides, they develop a prototype by the existing Protégé plugin, in which an ordinary user can simulate lawsuits in real or fictitious cases through a friendly experience, without bothering with the low formal level of DL. The results indicate that the proposed theoretical framework could be expanded to deal with other principles.

Nevertheless, several parts of this paper, including “Theoretical Background” and “Suitability of Preferential DL to the Legal Domain” should be improved.
Major issues:
1) There exist definitions that are not easy for readers to comprehend.
a. The introduction of antecedents and consequents of defeasible subsumption axioms (page 13, line 16-21) is not clear, which is helpful to understand Definition 4.1. Please give a concrete example of this definition.
b. Definition 4.2 is the key in this paper, so one example with a detailed explanation is necessary. In addition, employing \mathcal{F}_1 and \mathcal{F}_2 as concepts may not be suitable. In general, these symbols often represent axioms.

2) The verification of the method in this paper is not comprehensive.
a. The developed prototype is tailored only for the examples in this paper, but it still lacks concrete illustrations of time consumption.
b. Although Preferential DL is decidable, there still exists a scalability problem for them because its computational complexity of entailment is EXPTIME-complete. Hence, I hope to see the real computational complexity of entailment for the legal domain rather than an upper bounder.

Minor issues:
1. There still exist some typos in this paper such as Ontologies -> ontologies (page 2, line 16).
2. "The third reason relates to the way we humans actually reason under incomplete information (page 7, line 49)" is spoken language.
3. One “.” is missed of the sentence “In the following example, we address these particularities” (page 6, line 36).
4. The format of citation is not well such as [43], [44], [45]-> [43-45].
5. The original definitions from references should be provided citations such as Section 2.1.3 (page 4).

1. The proposed method is novel for solving existing problems in the legal domain.
2. The authors presented conceptualizations of several crimes from the Brazilian penal code through a slight modification in the implication operator of Preferential DL.
3. The developed prototype with encouraging results can verify the proposed approach to some extent.
4. The structure of the paper is acceptable, and its figures and examples are well-readable.

1. Some definitions are hard to be comprehended and lack corresponding examples.
2. Insufficient analysis of time consumption in real criminal cases.

Review #2
By Katarina Britz submitted on 28/May/2020
Major Revision
Review Comment:

Review: Submission 2447-3661

The paper proposes a computational approach to modelling and resolving certain normative conflicts in legal reasoning, in particular those arising from defeasible rules. In such cases, different rules are applicable in different situations, and the appropriate rule has to be selected to determine, for example, which penalty is applicable. A solution to this problem would also address the double jeopardy principle, which determines that an individual may only be penalised once for any given offence.

The reasoning support required needs to incorporate preferential exception handling in a systematic way, something that classical ontologies cannot provide. The paper makes a persuasive case that preferential description logics do provide suitable reasoning support to model and reason about these legal regulations. The topic is relevant to the scope of the journal and makes a useful contribution to the application of defeasible ontologies. However, the paper is not publishable in its current form, due to a large number of (often minor, but sometimes important) issues that should be addressed. These include some technical issues that need to be ironed out.

The authors sometimes use the terms “exceptional” and “typical” in the technical sense in which these terms are defined in defeasible DLs, when their use is however linguistically counter-intuitive. For example, on p2-c2-line 8 (page 2, column 2, line 8), robbery is deemed more exceptional than theft. This creates the impression that theft (without any further knowledge of whether it was accompanied by violence) is more common than theft with further knowledge that it was accompanied by violence. What is presumably meant is simply that theft is a more general category, and robbery is a subcategory, and hence more exceptional. So robbery is a more specific crime category than theft, rather than more exceptional. Using the term “exceptional” here is misleading. There are a number of similar instances, also related to the use of “typical”. If most acts of stealing are characterised by violence, how can theft without violence be called typical, and robbery an exception? On p5-c1-line 20, defeasible subsumption is defined “to axiomatise exceptions for the typical objects”. This makes the notion of typicality synonymous with absence of conflicting specific information. For example, a typical theft is one without violence. There is no reason to assume this. Typicality is a technical term used to refer to “minimal object in the preference order in any model”, but using it like this in an applied setting is in my view problematic and misleading.

The DL language chosen is SHIQ, but on p3-c2-line 9 you have nominals in the concept language. If they are not needed anywhere they should be removed here. Disjunctions are needed, and in Section 4.1 a normal form is introduced (though not formally) where the antecedent of axioms are only conjunctions. These include defeasible axioms, and there need to be some explanation (either in the background or in Section 4.1) justifying why the concept language is still SHIQ. Can one always rewrite a defeasible axiom in this form, or does this impose some restriction on the concept language, or on what can be defeasibly stated?

In Definition 2.2, a model of a concept C is defined as any interpretation in which C has a non-empty extension. “In addition, …” There are two problems with this definition: Firstly, the semantics of axioms (subsumption axioms and instance axioms) haven’t been defined yet, so one cannot define the notion of a model of a TBox or ABox yet. Secondly, all concepts are interpreted in all interpretations, sometimes as the empty set. It doesn’t make sense to link the notion of a model of a knowledge base to some concept having a nonempty interpretation. The definition should move down to after line 29, and the two notions should be separated.

The discussion on p6 after Definition 2.5 doesn’t capture the point of why preferential reasoning is too weak. The fact that monotonicity fails doesn’t in itself show that preferential reasoning is too weak. The issue is the absence of the additional precondition of (RM) which is ignored in preferential entailment.

Definition 2.6 misses the criterion that a minimal model should exist and must either be unique, or there must be some predefined way to select one of the minimal models. Then on p6-c2-line 10, it would be useful to relate the presumption of typicality to the notion of a minimal model. On p6-c2-line 20, Is Figure 1 any possible interpretation, or how does it relate to the minimal model?

The focus of the paper is on one specific type of conflict, but presumably the penal code contains all three kinds discussed in Section 3.1. So the usefulness of the defeasible DL formalism would be very limited if the other types cannot be handled. Can they already be modelled in classical DL? That would be both important and useful to point out and briefly discuss.

There is something wrong with the classical example in Figure 3. At present, the classical ontology does not entail that Robbery is subsumed by Theft. Here is a similar example: Every girl is female. Every girl with blue eyes is a person. But it doesn’t follow that every person is female. The penalties also don’t follow. Perhaps a missing disjointness statement between RobberyPenalty and TheftPenalty?

It would be useful to refer to rationality (RM) on p12-c2 a bit earlier than is done at present, in the inference on line 16. In the discussion on lines 30-34, there is something more that needs to be explained. The RM property ensures that the typicality of all objects are maximised. The axiom on line 37 targets a specific set of objects — namely those in the extension of Event \sqcap \exists realizedThrough.Steal. But the axiom isn’t what effects the maximisation of typicality — instead, it constrains them to also belong to the RHS concept. So the preceding explanation doesn’t really explain the addition of the axiom.

Is it reasonable to impose the modelling restriction that the consequents in the ovverrule relation be disjoint? This seems to be required for the lemmas to follow, but was that the real reason for introducing the restriction, or does it also make intuitive sense? Some explanation of whether this is just a technical ploy or whether there is some other motivation is required. The lemmas themselves require a few important tweaks:
Lemma 1: The assumption should be that < is not irreflexive.
Lemma 2: Up to line 13 is fine, from which line 19 follows by the (AND) rule. So A(\alpha) is defeasibly subsumed by \bot. And that, I think, is all one can say. Indeed, if the antecedent is inconsistent, then asymmetry fails.
Lemma 3: The term intransitive usually means never transitive for any objects. You mean “not transitive”.

In Example 4.1, it is stated that it doesn’t make sense to claim the existence of two crimes, Robbery and Murder (p15-c1-line 50). It seems reasonable to have Robbery (without Murder) and also to have Murder (without Robbery). Is the point not rather that these are not required in the example?

Section 5 covers related work. It is quite extensive, and some of it is not directly linked to the application considered in the present paper. It could be trimmed but some might disagree.

Editorial comments:

Referencing: Don’t mix referencing styles, as in e.g. “Horrocks (2005)[17] …”. Either write “Horrocks (2005) …”, or just “Horrocks [17] …” This should be fixed throughout the paper.
p1-line 27: what it should -> t should
line 32: suits better -> is better suited
p1-c2-line 46: use -> used
line 48: reference or explanation required — why should a wider audience be served?
p2-c1-line 3: This is out by a decade — should be the start of the present century
line 7: allows to axiomatize -> allows the axiomatization / allows axiomatizing of
p2-c2-line 31: preferential DL is not a superset of classical DL — it is an extension of classical DL
p3-c1-lines 29-31: DLs are well-behaved fragments of .. Clarify that not all DLs are equally expressive — some are quite inexpressive, and others very expressive.
p4-c1-line 1: Avoid unnecessary capitalisation (e.g. “An Interpretation” should be “An interpretation”). Also throughout the paper, as with “Model” in line 8.
p5-c1-line 4: iot -> IoT (in the acronym the capitalisation is required)
p6-c1 lines 47-51: Remove “P \Vdash “
p6-c2-line 3: Remove “= T \cup A“ as this has already been established in the previous definiton.
p7-c1-lines 44-46: The issue is not really that it is computationally expensive, but rather that it is difficult, requires intensive maintenance and remodelling, etc.
p7-c46-line 46: Compliance -> compliance
p8-c1-line 49: e1 and e2 individuals -> individuals e1 and e2
p10-c1-line 5: on -> of.
line 19: Conversely, Perdurants … -> In contrast, perdurants … (“Conversely” means something more specific.) Ditto p10-c2-line 8.
p12-c2-line 2: The latter -> These
p14-c2-line 2: Only for -> For
p15-c1-line 43: whether -> when
p15-c2-line 51:call -> calls / a call?
p16-c1-line 2: DIP tab -> the DIP user interface (or something a bit more descriptive than “DIP tab”).
p16-c2-lines 25-36: Either remove “proposed by” and just keep the reference in each case, or insert the name of the author(s) in each case.
p17 Figure 8: This hand-drawn line and arrow are really ugly.
Throughout the paper: The font used in many of the displayed knowledge bases is very small, and without any good reason. E.g. on p18-c1-line 48, the font is quite readable, but then in c2-lines 13-14 it is tiny. Why not make the font the same size throughout?

Review #3
Anonymous submitted on 01/Jun/2020
Major Revision
Review Comment:

This manuscript presents a work of applying Preferential Description Logics [13] to represent legal regulatory knowledge. It models the criminal types as well as their corresponding penalties of the Brazilian law, especially w.r.t. the Double Jeopardy principle. The contribution is well presented with a well discussed related work section.

The paper is well written. It introduces a sound background, from the classical DL to the preferential DL. The advantage of the preferential DL in dealing with exceptions is well presented by an example on Bird, Fly, etc. In the methodology part, a case study about crimes against property is adopted, and the suitability of preferential DL to the case is well explained and its suitability to the Double Jeopardy principle is proved.

Although the application of the preferential DL is presented by a mini demo by a Protégé plugin (DIP Reasoner), I still concern the evaluation part of the manuscript. A result report with more evaluation on the suitability (maybe with user feedbacks) can be added. What is the readability of the axioms? What is the accuracy in inferring the criminal type? What is the statistics of the ontology for modeling the case study about crimes against property? Any comparison with other DL extensions in dealing with different conflicts?