In this paper, we present CRIKE, a data-science approach to automatically detect concrete applications of legal abstract terms in case-law decisions. To this purpose, CRIKE relies on the use of the LATO ontology where legal abstract terms are properly formalized as concepts and relations among concepts. Using LATO, CRIKE aims at discovering how and where legal abstract terms are applied by judges in their legal argumentation. Moreover, we detect the terminology used in the text of case-law decisions to characterize concrete abstract-term instances.
Law is general and abstract by de nition. On the opposite, court case law decisions are speci c and concrete, in that they provide a peculiar interpretation of law applied to the considered single cases. Legal interpreters, such as for example judges and lawyers, are daily involved in analysis and evaluation of court case law with the aim to extract/derive possible suggestions for incoming case applications by relying on the experience of past applications that can be considered as a sort of consolidated legal knowledge.
According to the Italian law, the legal terminology can be distinguished into three main categories, that are i) statutory terms, i.e., terms directly or indirectly de ned by law; examples of statutory terms are public o cer, illicit drug, and consumer; ii) descriptive terms, i.e., terms featuring actions, human activities, and any real-life object; examples of descriptive terms are escape, car, and year; iii) abstract terms, i.e., terms featuring something indeterminate that requires a concrete application for being really de ned; examples of abstract terms are good faith, long-term cohabitation, and dangerous driving. Consider the abstract schema of a legal action provided in Figure 1. When a new case law is received for Copyright c 2019 for the individual papers by the papers authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors. SEBD 2019, June 16-19, 2019, Castiglione della Pescaia, Italy.
New case law to judge judge Decision/Verdict on new case law (concrete law interpretation) Law
Courthouse History of Case-Law Decisions
(CLDs) judgement, the expected evaluation process has to take into account i) the law, for understanding the terms, either statutory, descriptive, or abstract, that can be relevant for the current case, and ii) the history of case-law decisions, for detecting possible relevant interpretations and concrete applications of abstract terms that can be useful to support the decision/verdict to eventually deliver.
In this paper, we present CRIKE (CRIme Knowledge Extraction), a datascience approach to detect concrete applications of legal abstract terms in large case-law decisions. To this purpose, CRIKE relies on the use of LATO (Legal Abstract Term Ontology) where legal terms are properly formalized as concepts and relations among concepts. Using LATO, CRIKE aims at discovering how and where legal abstract terms are applied by judges in their legal argumentation.
The paper is organized as follows. In Section 2, the CRIKE approach is introduced. The LATO ontology and the CRIKE techniques for legal knowledge extraction are discussed in Section 3 and 4, respectively. Related work are discussed in Section 5. Concluding remarks are provided in Section 6. 2
The CRIKE approach (see Figure 2) is conceived to support extraction of legal knowledge from a (possibly large) dataset of Case-Law Decisions (CLDs) coming from di erent, o cial sources, such as for example First Grade and Court of Appeal judgements. CRIKE embeds the LATO ontology where relevant law concepts of a given domain of interest are properly formalized. To enforce knowledge extraction, CRIKE exploits a given dataset of CLDs in input by adopting a conventional data-science process where each CLD is indexed and stored in a digital format. In particular, the CLDs of our dataset are acquired from the Court and the Court of Appeal of Milan and they are usually provided in image format with highly heterogeneous quality. The indexing and storage activity exploits data cleaning and tokenization techniques to obtain a pure textual version of each CLD as well as a focused set of metadata. By exploiting the indexed CLDs metadata, knowledge extraction is enforced with the aim at classifying a CLD with respect to the LATO ontology knowledge. In particular, extraction expert-based ontology design indexing/storage of
CLDs Law
LATO ontology (abstract term specification)
legal
knowledge extraction
concrete applications
of abstract terms
is focused on detecting the concrete applications of legal abstract terms in the
text of the considered CLDs. The crucial idea of CRIKE is that the detection
of a given abstract term AT is not only concerned with the recognition of single
terms featuring AT , but also with the recognition of terms associated with the
ancillary concepts related to AT , that we call abstract-term context.
Motivating example. Consider the Italian law about drugs and related drug
o enses, as reported in [
To formalize the knowledge related to abstract terms and their interpretation, we introduce LATO in CRIKE. LATO is a legal ontology where relevant law terms to exploit knowledge extraction in CLDs are de ned; it contains concepts to represent general law terms, either abstract, statutory and descriptive terms.
LATO is manually de ned by domain experts and implemented according to the SKOS formalism. In particular, the concept hierarchy is based on a root concept Term with three main subconcepts, namely AbstractTerm, DescriptiveTerm, and StatutoryTerm (see Figure 3(a)). In addition to general law terms, the LATO ontology contains concepts that represent the Italian legislative structure, such as for example the concepts Law, LawArticle, and LawParagraph. Furthermore, the concepts Conviction and Discharge are also speci ed in LATO to represent the possible Court decisions (i.e., the verdict) of a given case law. In particular, the concept Conviction denotes a verdict in which the Court judges the defendant guilty, while the concept Discharge denotes a verdict in which the facts have a penalty relevance, but no punishment is nally delivered. Finally, the concepts Quantity and UnitOfMeasure are de ned in LATO for allowing to represent the quantitative estimation of substances that can appear in legal documents.
AbstractTerm is the core concept of the LATO ontology since it represents the target of the knowledge extraction functionalities of CRIKE. The related construct of SKOS is exploited to enrich the speci cation of an abstract term AT by formalizing the ontology relationships between AT and the other concepts of the LATO ontology composing its context. In particular, given a considered abstract term AT , related is used to connect AT to ancillary concepts of LATO representing i) an objective judgment element OBJ usually expressed through the connection of AT with a descriptive/statutory concept; ii) a subjective quantitative evaluation SU BJ usually expressed through a relationship between AT and Quantity/UnitOfMeasure concepts; and iii) a legislative reference LREF usually denoted with a connection of AT with a speci c law or regulation (i.e., Law, LawArticle, and LawParagraph concepts). According to SKOS, each LATO concept is associated with a preferred label (prefLabel) as well as with one or more alternative labels (altLabel) and hidden labels (hiddenLabel) to enrich the concept de nition with a label-set of literal descriptions that is very useful for subsequent knowledge extraction, to capture possible synonyms, acronyms, and abbreviations in the text of CLDs.
Example. An example of SKOS de nition for the abstract term AT = MinorOffense is shown in Figure 3(b) according to the Italian drug-tra cking law. MinorO ense is related to the concepts Drug and DrugTra ckingVerb that represent the OBJ relationships since they are subconcepts of StatutoryTerm and DescriptiveTerm, respectively. The relationships with the concepts Quantity and UnitOfMeasure represent the subjective judge evaluations SU BJ . The concepts Par5, Art73, and DPR309/1990 are subconcepts of the LawParagraph, LawArticle, and Law, respecTerm
AbstractTerm
MinorOffense
DrugMinorOffense DescriptiveTerm
DrugTraffickingVerb StatutoryTerm
Drug
Cannabis Cocaine
Heroin UnitOfMeasure Quantity (a)
Law
DPR309/1990 PenalCode PrivacyCode
Weapon LawArticle
Art648
Art73 LawParagraph
Par3
Par5 Conviction Discharge
Drug
Drug Trafficking Verbs OBJ
Quantity MinorOffence (b)
Unit SUBJ Measure tively, and they express the legal references LREF of MinorO ense in the Italian criminal code where the drug tra cking crime is de ned. 4
Knowledge extraction in CRIKE is based on the idea to exploit text analysis techniques for detecting the concrete applications of legal abstract terms belonging to LATO throughout the stored/indexed case-law decisions CLDs. To this end, for a given abstract term AT , we introduce the notion of abstract-term context CtxAT containing, besides the AT term, all the concepts of LATO that are ancillary to AT , namely OBJ , SU BJ , or LREF concepts:
CtxAT = fCi j r(AT; Ci)g where r(AT; Ci) denotes a SKOS related relationship between the abstract term AT and the concept Ci.
For each concept C 2 CtxAT , we de ne the concept label set LC that contains the whole set of labels, either preferred, alternative, or hidden, associated with C. Furthermore, based on the notion of LC , we de ne the extended label set LC where the concept label set of C is enriched by including the concept label set of the concepts Cj subsumed by C:
LC = LC [ LCj j Cj
C
Consider the goal to detect the concrete applications of a certain abstract term AT in a dataset of case-law decisions CLDs. CRIKE knowledge extraction is enforced by exploiting the extended label sets LC of the concepts in the context CtxAT . For each document d 2 CLDs, we de ne a vector representation d where each element corresponds to a concept in the context CtxAT . The value d[i] 2 d is set to 1 when a label hit is detected, meaning that at least one occurrence of a label in LCi is found in d for the concept Ci 2 CtxAT , and 0 otherwise (i.e., label miss). A threshold based mechanism is de ned to specify the minimum number of label hits required to consider that a concrete application of the abstract term AT is detected in the document d.
Deal deal To deal
Quantity amount Deal Quantity
Drug Trafficking verbs kg. kilogram Kilo art. 73 article73
A. 73 Kilo Weight
Unit OfMeasure
MinorOffense
DPR Art73 309/1990
DPR 309/90 L. 309/1990 public safety
Act Par5 Drug par. 5 Par. five paragr. V Cocaine cocaine Crack crystal The police officers, during a routine inspection, discovered four bags of white powder onto the toilet floor (weighing 3.6 kg. in total). The powder tested positive for cocaine. […] the deal.
Both males were arrested. For instance, considered matters, mode of action, quality and quantity of the substance, the penal relevance is clear, but the offence is minor.
For the foregoing reasons, the court applies the art. 73, DPR 309/90, par. 5 […]. Example. Consider the abstract term AT = MinorO ense and the corresponding context CtxMinorO ense = fDrug, DrugT raf f ickingV erb, DP R309=1990, Art73, P ar5, Quantity, U nitOf M easureg. Moreover, consider the extended label set LDrug = LDrug[ fLCocaine; LHeroin; LCannabisg. In Figure 4, we show an example of knowledge extraction based on the concepts and corresponding extended label sets in the context CtxMinorO ense. An example of vector-based document representation for the abstract term AT = MinorO ense is shown in Figure 5. If we consider a threshold of 80% of label hits, we have that a concrete application of MinorO ense is detected in document d1 since 6 hits are found over the available 7 concepts in CtxMinorO ense.
Work related to the issues addressed in CRIKE regards legal argumentation
mining and legal ontology design. Legal argumentation mining refers to the
capability to automatically detect and classify the role of possible argumentative
units within a considered legal case text [
A survey on legal ontology design is presented in [
Original contribution of the proposed CRIKE approach is related to the enforcement of a data-science process with the support of an expert-based law ontology to extract knowledge from CLDs. A further peculiar feature of CRIKE is related to the formalization of an abstract term as a legal ontology concept with a corresponding context of related concepts. Ontology concepts with associated contexts are used to drive the identi cation of concrete applications/interpretations of corresponding abstract terms in the text of CLDs. 6
CRIKE support to practices and concluding remarks In this paper, we presented the CRIKE approach for legal knowledge extraction. We envisage the following main practices that can be supported by using CRIKE i) knowledge-assisted verdict writing, where the concrete terminology extracted for abstract terms can support the judge in the preparation of new case-law decisions; ii) history-based verdict prediction, where the knowledge extracted by CRIKE is used to train a machine learning mechanism with the aim to predict the possible decision on a new incoming case-law to judge; and iii) legal analytics, where the results of knowledge extraction are exploited to detect possible trends and common abstract-term interpretations.
A preliminary experimentation of CRIKE has been performed based on a
dataset provided by the Courthouse of Milan, Italy, whose results are described
in [
Di erent research directions are currently being investigated. On the one side, we are working on a bootstrapping approach to enforce enrichment of the LATO ontology, so that the context of abstract terms can be progressively augmented with new relevant terms and literals as long as they are detected in CLDs during extraction. On the other side, machine learning techniques are being developed to enforce a supervised classi cation of CLDs based on abstract terms, by exploiting a training set of CLDs manually annotated by domain experts.