Legislative Knowledge Management with Property Graphs Andrea Colombo1,∗ , Francesco Cambria1 and Francesco Invernici1 1 Politecnico di Milano, Dipartimento di Elettronica, Informazione e Bioingegneria, Via Giuseppe Ponzio, 34, 20133, Milan, Italy Abstract The sparse nature and intricate set of relationships between legislative acts pose a significant challenge in the choice of the underlying database model, which both allows for performing structured queries and developing intuitive and smooth knowledge management. In this paper, we propose to use Property Graphs as a powerful alternative for managing legislative knowledge. First, we discuss how graph queries are a valid alternative solution to standard legislative knowledge management by showing how our data model fully captures the problem of law versioning (i.e., the existence of many versions for the same law). Then, we analyze, propose and implement innovative ways for monitoring the legislative system using Property Graph tools that have been recently standardized and developed, such as triggers and graph-based association rules, which empower our model of advanced ways of handling legislative data. For instance, we will show how we can use these tools to develop intelligent warning systems that inform stakeholders of critical changes in legislation through active rule reasoning or to detect shifts in graph patterns via continuous monitoring of significant association patterns. Keywords data management, property graphs, triggers, law 1. Introduction logic (e.g., handling events as document create, document update, document delete, property changes) harms the po- Legislative data comprises the set of acts, bills, or other nor- tential of reasoning over legislative data, leveraging relation- mative documents produced by an authority that detain a ships among acts to discover and detect anomalous patterns legislative power to regulate some domains on a national, re- that require attention. For instance, abrogations of some gional, or international level. In most democratic countries, portions of legislation can inadvertently create voids in legal such a power is in the hands of parliaments, be they na- foundations of third acts. Thus, a data model that cares more tional, regional, or federal, according to the political system. about relationships among acts is crucial in monitoring such One of the main challenges when dealing with such tex- scenarios. tual documents is converting them into machine-readable In this paper, we present and discuss an innovative way formats that enable the adoption of structured approaches of conducting legislative knowledge management based on when querying and analysing legislative acts. Property Graphs (PG). By leveraging state-of-the-art ad- To tackle this, the computer law community has devoted vancements from the database community in the formal many efforts to proposing appropriate international stan- standardization of Property Graphs and their tools [7, 8, 9], dards to capture the common grounds of laws enacted in we present an alternative approach for handling legislative different legislative systems. Most of previous works have acts, which maintains the ability to manage documents as been based on the eXtensible Markup Language (XML) for- hierarchical structures, with the creation of dedicated part- mat, a semi-structured data model that has been naturally hood graph relationships, but that enriches it with tools of used for representing textual and hierarchical data, such as additional utility. First, we demonstrate how our proposal laws [1]. Relevant XML-based proposals include the Legal seamlessly handles textual documents by focusing on one Knowledge Interchange Format (LKIF) [2], LegalRuleML [3] of the main issues of legislative knowledge management, and Akoma Ntoso [4]. that is, the law versioning problem, i.e., retrieving the text However, the need for a user-friendly storing of textual of a law in force in a certain timestamp, thus accounting for documents (with XML tags and hierarchical structures being successive modifications and/or abrogations. ideal for such aims) collides with the possibility of conduct- Then, we show how, by explicitly modelling references ing quantitative analysis of the complexities and features among acts as graph relationships, we enable more sophisti- of the legislative system [5], which strongly depends on cated tools to analyze and monitor trends in the legislative the ability to traverse relationships (e.g., dependencies and activity [10], such as by the development of intelligent sys- references between laws and articles). In fact, while the tems that can reason over legislative data through the use of creation of XML formats to represent the components of active rules in the form of Property Graph triggers [8]. To an individual law (i.e., article or section structure, tags for this aim, we develop a trigger-based warning system that references to other laws) is extremely useful for structured monitors the legislation, capable of detecting harmful sce- access to a document, it becomes cumbersome to manage narios within the legislative system, such as the creation of and analyze trends and the complexity of the legislation. legislative voids. We demonstrate how we capture the latter Furthermore, while XML-native databases generally support scenario by using active rules in the form of triggers ex- tools such as triggers [6], their focus on the document-level pressed in Cypher, the closest language compliant with the Graph Query Language standard. We demonstrate the effec- Published in the Proceedings of the Workshops of the EDBT/ICDT 2025 tiveness of this warning system over the Italian legislation, Joint Conference (March 25-28, 2025), Barcelona, Spain for which we have implemented our PG-based schema and ∗ Corresponding author. stored it within a Neo4j database. We experiment how warn- Envelope-Open andrea1.colombo@polimi.it (A. Colombo); ings signaled by our reasoning-based approach significantly francesco.cambria@polimi.it (F. Cambria); positively correlate with ex-post-legislative interventions francesco.invernici@polimi.it (F. Invernici) Orcid 0000-0002-7596-8863 (A. Colombo); 0009-0002-0268-4840 that aim at fixing the legislative void, demonstrating their (F. Cambria); 0009-0002-5423-6978 (F. Invernici) predictive power for a timely detection of legislative voids. Copyright © 2025 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0) Finally, we analyze how association rules can be a useful CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings tool for knowledge discovery, providing insights into trends in the evolution of the legislative system. To this aim, by leveraging the recently developed association rule opera- tor for property graphs [9], we present an application in which association rules can help in detecting the attitude of governments towards a topic of particular interest, i.e., whether there is discontinuity among successive govern- ments. We focus on legislation related to “procurement”, and we find that, in Italy, there is a continuous significant stream of creations and abrogations of rules governing this topic, certifying a certain degree of instability in the topic in recent years. Overview. The remainder of the paper is organized as fol- lows. Section 2 presents the Property Graph Schema that we adopt for modelling legislative knowledge, discussing its Figure 1: Graphical visualization of the Property Graph Schema. implementation in the Italian system and how it can retrieve Properties are omitted for better visualization. time-dependent versions of laws. In Section 3, we present how tools like PG triggers and graph association rules can be used to facilitate knowledge discovery and automated within a law document than articles, are also modelled as monitoring systems. In Section 4, we perform our experi- distinct nodes with their characteristics. Law, article and ments to demonstrate the benefits of using a graph database attachment nodes are interconnected using a set of directed and its rich set of tools to manage legislative knowledge. edges that capture references, dependencies, and amend- ments that denote the evolution of a legislative system. We included four fundamental types of relationships which we 2. Property Graphs for Legislative identified as generic enough across multiple legislative tra- ditions: an “is the legal basis of” dependence, an amends, an Data abrogates and a more generic cites reference. Modelling legislative data and the interconnections between The formal PG-Schema, as defined in [7], that we use to laws and articles in a property graph offers a structured model legislative data is the following: and dynamic approach to understanding and managing the CREATE GRAPH TYPE lawsGraphType STRICT{ knowledge of a legislative system. (lawType: Law {id STRING, title STRING, In this section, we present a property graph schema that lawNum INT, typeLaw STRING, publicationDate aims to handle legislative data and discuss how we can DATE, numArt INT, numAttach INT}), leverage the expressiveness of a graph query language, such (articleType: Article {id STRING, title STRING, as Cypher, the closest GQL-compliant language, to handle a number INT}), fundamental feature in legislative knowledge management, (attachmentType: Attachment {id STRING, title the retrieval of the correct version of a law based on a desired STRING, type STRING}), timestamp. (:lawType)-[hasArticleType: has_article]-> (:articleType), (:lawType)-[hasAttachmentType:has_attachment]-> 2.1. Property Graph Schema (:attachmentType), (:lawType)-[referenceType: is_legal_basis_of] In [11], we proposed the first approach for representing ->(:lawType), the Italian legislative system, in terms of documents and (:articleType)-[referenceType:is_legal_basis_of| references, through a Property Graph, together with its amends|abrogates|cites]->(:lawType), schema [7]. While our proposal was country-specific, its (:articleType)-[referenceType: amends|abrogates| foundation relies on an internationally adopted and OASIS cites {paragraph STRING, newText STRING}]-> standard [12], which aimed to define common elements (:articleType), that characterize legislative acts across multiple legislative (:articleType)-[referenceType: amends|abrogates traditions [13]. With no or little adaptations, our schema {paragraph STRING, newText STRING}]-> applies to multiple countries; thus, we can consider our (:attachmentType), schema powerful enough to represent any legislation. (:attachmentType)-[referenceType: is_legal_basis_of|cites]->(:lawType), In Figure 1, we visually present the schema. In a graph, (:attachmentType)-[referenceType: cites nodes can represent laws connected to their articles (or, ac- {paragraph STRING}]->(:articleType), cording to the legislative tradition, sections, clauses, etc.) (:attachmentType)-[referenceType: cites through a parthood relationship. Metadata relevant to the {paragraph STRING}]->(:attachmentType)} law, such as title, publication date, and law domain (indi- cating the ministries responsible for the law’s content), are Regarding edge properties, all defined types of references assigned as properties directly to the law node, enabling a share the common attribute paragraph, which denotes, if structured approach to query and analyze the legislative cor- applicable, the specific portion of the article’s text that is pus. Articles are distinct nodes with properties that detail being referenced. For instance, one could be interested in the specific sections of the law they represent. These prop- citing or amending only a specific portion of text within the erties include the article text, any headings or titles specific article. Thus, such a feature can be captured by inserting the to the article, and a list of policy-related topics governed by paragraph of interest. As a consequence of such a modelling that article. Attachments, which serve a different function choice, an abrogates edge whose paragraph property is NULL Figure 2: Annotated Cypher query deriving the in-force text of law IDLAW at a certain TIMESTAMP . In red, the two variables to be set according to the user’s requests. First, a subquery (within the CALL construct) is used to derive and union the law articles in-force at the desired timestamp. Then, a COLLECT function is used to produce the correct law text as a document. implies a full abrogation of the destination article. A law publication dates available in law nodes, is capable to always becomes abrogated whenever all its articles have been ab- infer the desired law version. By doing so, we avoid storing rogated. Thus, a law abrogation can be inferred by looking multiple textual versions of the same law, achieving a more at whether all its articles have an ABROGATES edge without efficient storage of legislative data and, at the same time, the paragraph property.1 allowing a more structured and straightforward approach Finally, amends and abrogates edges are the only types of for querying statistics or detecting trends of a legislative references that also share a second property: newText. Such system. (textual) property captures the new text of the destination In Figure 2, we present the query in Cypher that we article node, as modified/abrogated by the source article designed to manage law versioning based on our proposed node. schema. Such a query strongly relies on navigating graph The Italian Legislative Property Graph We implemented patterns to combine temporal information. In detail, by our schema for the Italian legislation and derived a graph specifying a TIMESTAMP 𝑇 and a law of interest 𝐿 (i.e., an consisting of Italian national laws. The graph is stored in IDLAW ), we can get the text version of 𝐿 at 𝑇 by performing a a Neo4j database (available at [14]), comprising 74k law union that concatenates articles of 𝐿 in their latest modified nodes, 318k article nodes, 127k attachment nodes, 107k legal version, as per modified by the law with the maximum basis dependencies, 64k abrogations, 80k amendments, and publication date before 𝑇, with articles that have never been 228k generic citations. amended or have been amended only after 𝑇, thus do not have any incoming edge of types ABROGATES and AMENDS . Finally, we note how this approach is replicable even on 2.2. Managing Law Versioning via Graph a more granular level, i.e., by considering paragraphs as Queries graph nodes. The benefit would be even more pronounced Legislative systems evolve over time, generating a continu- since most of the modifications change only some words of ous stream of new laws that amend or repeal existing ones. a law, and a standard document versioning approach would Each change represents a new version of the law, captur- require the storage of an additional document, even for tiny ing the updated text whenever a modification occurs. This substitutions. Here, the challenge would be consistently process often results in an exponential increase in textual modelling paragraph nodes and linking textual references documents, as even minor changes necessitate the storage to such nodes, since references do not always explicitly of additional files with the novel text. mention the paragraph, as we experienced with the Italian By adopting the graph model and the schema that we pro- legislation. posed, we can seamlessly handle such feature by (i) storing the original textual version of each article of law within a 3. Triggers and Association Rules node property, (ii) storing the modified textual version, as per amended by successive laws, as an edge property and Triggers have been a fundamental feature since the incep- (iii) delegating the computation of the in-force text for a law tion of relational databases [15]; they were explored in depth at a certain timestamp to a graph query, that, by leveraging in [16] and formalized in the ISO-ANSI SQL3 Standard [17]. Traditionally, triggers have been utilized to automate tasks 1 In general, an abrogation can also directly repeal an entire law, which that facilitate the development of automation systems [18]. is also permitted by our schema. n the rest of the paper, we will not In graph databases, adding reactive components has also consider this case since it can be trivially considered by adding a UNION been helpful in supporting important applications, such as operator in all our applications. the monitoring of pandemic events [8], which naturally are modeled into knowledge graphs stored in graph databases. Formalizations of triggers for graph database and Property Graph have recently been proposed [8, 19]. Due to the rich- ness of the graph data model w.r.t. the relational model, triggers can also be leveraged to encode more complex (and interesting) events. For instance, since property graph items can be either nodes or relationships, triggers can include the creation and deletion of nodes/relationships as well as the setting and removal of their labels and properties. The Figure 3: Law 2021/179 repeals law 2000/174, by abrogating strong analogy between triggers, i.e., active rules, and de- its remaining in-force article (i.e., article 7). However, since law ductive rules, as defined in [20], has also enabled the use of 2000/174 is a legal foundation of law 2010/37, law 2021/179 also triggers for adopting reasoning approaches over data mod- undermines its validity since its foundation has been removed, elled into knowledge graphs stored as property graphs. raising a potential legislative void. Association rules (AR) mining was first introduced in [21, 22] and aim to uncover patterns or correlations within data. Classically, association rules take the form of 𝑋 → 𝑌 where pealed an older law by leveraging the graph schema, 𝑋 is referred to as the body of the rule and 𝑌 as its head. Both as discussed in Section 2. In such cases, we merge body and head represent predicates or conditions that can be REPEAL edges that connect the new law node with validated within the data, based on two key metrics: support the abrogated law nodes. and confidence. Association rules mining has been applied 2. By traversing the IS LEGAL BASIS OF edges, a sec- in many different domains: for instance, for highlighting ond reasoning rule monitors whether the REPEAL which items are frequently purchased together [23, 24], for of a law has created a legislative void in the legal detecting connections between symptoms and diseases in foundation of a third in-force law. For such cases, medicine [25], or to uncover patterns of user interest in it creates a warning edge informing about the po- social networks [26]. Like triggers, the flexibility and rich- tentially harmful situation (that we denote as an ness of the graph data model allow association rule mining UNDERMINES edge). to uncover more insightful patterns within the data. The A visual example of the reasoning process is available in MINE GRAPH RULE operator, introduced in [9], leverages Figure 3. all the capabilities of the graph query language to simplify The first rule can be encoded within the following trig- the definition and extraction of association rules from graph ger expressed in Cypher, a possible implementation of PG- databases. While the operator is not fully integrated into Triggers, as defined in [8]: the standard Cypher language, its syntax closely follows Cypher, ensuring clarity and readability. Furthermore, the UNWIND $createdRelationship AS newRel queries used to extract association rules can be implemented MATCH p=(l:Law)-[:HAS_ARTICLE]->(a:Article) in any graph database management system that supports <-[newRel:ABROGATES]-(a2:Article) GQL. <-[:HAS_ARTICLE]-(newLaw:Law) In the rest of this section, we present two practical appli- WHERE newRel.paragraph IS NULL WITH l, a, newLaw cations of such tools, demonstrating their contribution to MATCH (l)-[:HAS_ARTICLE]->(a)<-[r:ABROGATES]-() implementing more sophisticated monitoring activities in WHERE r.paragraph IS NULL the legislative domain. WITH l, l.numArt AS NUMART, COUNT(DISTINCT a) AS NUMREPEALS, newLaw 3.1. Warnings for Legal Basis Abrogation WHERE NUMREPEALS >= NUMART MERGE (newLaw)-[r:REPEAL]->(l) In the legislative context, reasoning through active rules can be used to create automatic monitoring systems that derive where the UNWIND clause captures, within the Neo4j warnings for anomalous events based on graph patterns. ecosystem, the insertion of a new ABROGATES edges which We identified an important application of such a feature in derive from the insertion of a new law node. If a graph pat- the monitoring events of repealing important acts that are tern denoting an article abrogation without the paragraph legal basis to other pieces of legislation. Such abrogations property exists (denoting, according to our schema, a full might create legislative voids and hinder the applicability of abrogation), then the full abrogation is tested by monitoring third laws, which have their foundation in the repealed law. whether all its articles have been repealed. In such a sce- In other words, the content and rules stated in an act might nario, the trigger merges a new edge denoting an abrogation depend on legislation abrogated by the newly enacted law. (i.e., s REPEAL ) of the destination law node. In our schema, we modelled legislative dependencies The second rule can be implemented as: through IS LEGAL BASIS OF edges, which define the foun- dations of a given law. Thus, to implement the application, UNWIND $createdRelationships AS newRel we can use the graph structure to infer timely warnings MATCH (a)<-[:HAS_ARTICLE]-(newLaw:Law) about potentially harmful situations that must be moni- -[newRel:REPEAL]->(l:Law)-[:IS_LEGAL_BASIS_OF] tored. Such warnings take the form of new special edges ->(l2:Law)-[:HAS_ARTICLE]->(a2:Article) that can be visually inspected. In detail, we can define the WHERE NOT EXISTS (()-[:REPEAL]->(l2)) following deductive rules that generate warnings for the AND newLaw.id <> l2.id considered problem: AND NOT EXISTS ((a)-[:ABROGATES|CITES|AMENDS| INTRODUCES]->(a2)) 1. Whenever a new law containing abrogations is pub- MERGE (newLaw)-[:UNDERMINES]->(l2) lished, we compute whether the new law has re- where, here, the UNWIND clause captures the newly created Legislative Repealing Law Warnings Interventions REPEAL edges and checks whether it exists a non-abrogated 1949/264 1 4 law whose the repealed one is the legal basis of. Within 1954/615 1 0 the WHERE conditions, we exclude the cases in which the 1967/18 49 20 repealing law already contains any explicit “correction” to 1967/601 9 3 the destination law and its articles. If such a pattern exists, … … … we merge new derived relationships, i.e., UNDERMINES edges, which denote a potential legislative void since the newly Table 1 Sample of the warnings generated by a repealing law and the published law is deleting a law that is used as a legal founda- number of legislative interventions that have occurred, within tion in a third law. Thus, such edges require attention from one year after the publication of the repealing law, to third laws the legislator and any interested stakeholders in the domain having as foundation a law abrogated by the “repealing” law. of the law. The warnings might translate into corrective actions, i.e., new laws or articles that fix legislative voids that might have been generated. 4. Implementation and Results In this section, we execute the proposed triggers and associa- 3.2. Patterns of Government Attitude tion rules within the Italian legislation that we have already In a legislative graph proposed in Section 2, items represent modelled in a property graph [14], aiming to demonstrate laws, articles, and attachments. Such a schema can be easily the efficacy and utility of such tools in a real-world scenario. enriched by adding governments and topic nodes2 . The In particular, we show how warnings generated by our trig- former represents the governments under which laws have gers significantly correlate with ex-post interventions, i.e., been enacted; the latter connects each legislative node to correction measures that modify or add details about the the topics its content refers to. With these enrichments, we law signalled by the alert system based on triggers. Then, can exploit (graph) association rules to understand which we implement an AR in the Italian graph to analyze whether and if governments have the attitude to profoundly change we can monitor and find patterns of the legislation related legislation referring to a topic for which its predecessor to procurement rules, which are under strong scrutiny in government had also previously implemented legislation. the daily Italian news. We discover that in recent years, Through this application, we can detect temporal graph many governments have had a reshaping attitude toward patterns that monitor the behaviour of governments and such legislation, with profound changes and discontinuity identify discontinuity evolution of the legislation. The min- among recent governments, which harm the stability of ing of this kind of pattern can be achieved by adopting the legislation, one the drivers of economic growth [29]. MINE GRAPH RULE operator, whose syntax was proposed in [9], and can be written as: 4.1. Relevance of Warnings MINE GRAPH RULE FrequentChangesOnTopic As discussed in Section 3.1, we consider as a warning the GROUPING ON (t:Topic) emergence of an UNDERMINES edge, derived via reasoning WHERE t.name CONTAINS '$topicName' with triggers. We implemented the triggers in a Neo4j DEFINING database, which we also used to store the graph of the Italian BODY AS (t)<-[:OF_TOPIC]-(a:Article) legislation. To implement triggers as reasoning rules, ca- <-[:HAS_ARTICLE]-(l:Law)<-[:CREATED_UNDER] pable of inferring novel knowledge through deductions, as -(g:Government)-[:SUCCEDED_BY] ->(g2.Government) required by this use case, we made use of a publicly available HEAD AS (t)<-[:OF_TOPIC]-(a2:Article) Neo4j plugin3 that, through the use of a trigger controller, <-[:ABROGATES]-(a3:Article)<-[:HAS_ARTICLE] manages the correct execution order of triggers that enables -(l2:Law)<-[:CREATED_UNDER]-(g3:Government) reasoning. WHERE g2 = g3 To test the significance of these warnings, we adopted IGNORE a, l, g2, a2, a3, l2 a historical analysis strategy: we replicated the (ordered) EXTRACTING RULES WITH SUPPORT > 0.2 insertion of each law within the graph to let the Neo4j AND CONFIDENCE > 0.6 triggers activate based on the legislative situation at each timestamp. which extracts whether a certain topic of interest, regulated First, we computed the number of warnings generated for under a certain government, is also significantly reshaped by each law, which we depict in Figure 4. Then, we correlated its direct successor in government. Therefore, the output of such values with the number of ex-post interventions that such a rule is statistically significant patterns of disconti- the legislator had to make within the year after the publica- nuity that allow us to monitor a possible important shift tion of the law to fix or repair a legislative void. We count of regulation from one government to another regarding a interventions as the distinct number of references to the certain topic. We highlight how such output is more high- “undermining” law that occurs in the following 365 days af- level than looking for modifications and abrogations that ter the official publication. Thus, for each law that activates a government made to the legislation implemented by its the triggers, we get two values: the number of warnings predecessor. In fact, the complexity of the legislative system generated and the number of interventions. In Table 1, we hinders the result of such a query since these shifts could present a portion of the two values that we computed for also be captured indirectly through third laws, which share laws that have been signalled as warnings. the same (or similar) topic. Finally, we compute the Pearson correlation coefficient 2 to measure the linear relationship between the two datasets. Multiple approaches can be used to derive topics from textual docu- ments, either semi-supervised ones [27] or state-of-the-art LLM-based 3 topic extraction tools [28] The implementation is available at [30] Figure 4: Warnings generated by our trigger-based reasoning system, aiming to monitor the abrogations of relevant laws. Each bar represents the number of warnings generated whenever a new law is published, i.e., the number of UNDERMINES relationships that are merged within the graph. The idea is that a significant positive correlation coefficient Head Body Support Confidence implies that the higher the number of warnings generated, Government Predecessor the higher the probability of necessary intervention to fix I Meloni I Draghi 0.57 0.80 the legislative void, giving our warning system also some II Conte I Conte 0.29 0.66 predictive power to predict harmful legislation. Since the I Conte I Gentiloni 0.57 1.00 distributions of warnings and interventions are not nor- I Gentiloni I Renzi 0.57 0.80 mally distributed, we adopted a bootstrap approach [31, 32] I Renzi I Letta 0.28 1.00 to estimate the sampling distribution of the correlation co- I Letta I Monti 0.28 0.66 efficient and to test whether it is significantly different from I Monti IV Berlusconi 0.57 0.80 zero. We observed a Pearson coefficient of 0.394 and a p- IV Berlusconi II Prodi 0.57 1 II Prodi III Berlusconi 0.43 0.5 value of 0.02, implying a significant positive relationship II Berlusconi II Amato 0.29 1 between the two series, which can be interpreted as a higher II D’Alema I D’Alema 0.29 0.33 probability of intervention whenever more warnings are I D’Alema I Prodi 0.29 0.5 generated, i.e., a “central” law is repealed and requires more I Prodi I Dini 0.29 0.5 interventions. I Ciampi I Amato 0.29 1 VII Andreotti VI Andreotti 0.43 0.75 4.2. Evolution of Procurement Regulation III Moro II Moro 0.43 1 As a representative example, we identified regulation re- Table 2 garding “procurement” as an interesting topic to analyze Output of the MINE GRAPH RULE operator used to discover pat- through association rules. The theme is under scrutiny by terns of changes in procurement regulation. Results have been ordered by the temporal dimension, i.e., starting from the most the general public for its long-standing episodes of corrup- recent pair of adjacent governments, and depict a scenario in tion in the country, as also studied in dedicated economic which the government in the body column enacted relevant leg- research [33]. Through the running of our association rule islation about “procurement” which was profoundly reshaped by tool for this topic, we aim to monitor the behaviour of gov- the (successor) government in the head column. ernments in continuously modifying legislation in this area. We implemented the MINE GRAPH RULE operator in Neo4j4 a previous work (see [9]). The output of the AR motivations and benefits of adopting Property Graphs as run for the procurement legislation is presented in Table 2. a powerful solution for comprehensive legislative knowl- Considering the number of topics in the graph, setting the edge management. We demonstrated how adopting a graph support and confidence thresholds to 0.2 effectively identi- database and our schema can seamlessly handle the tem- fies strong rules. Out of the 69 governments of the Italian poral aspects and evolution of legislative acts, delegating Republic, 16 have significantly reshaped the topics their to queries the task of retrieving the correct information, predecessor had worked on. An alarming trend is that this such as computing the in-force text, instead of storing mul- ”practice” has intensified in recent years. Notably, 15 out tiple versions of the same act. Then, we looked at recent of the 16 governments identified began after 1990. Another tools recently adopted and developed by the Property Graph concerning behavior highlighted by the extracted rules is community, i.e., triggers and association rules, to offer in- that governments succeeding those of a different political novative and powerful applications that can be useful in alignment are more likely to repeal pieces of procurement conducting advanced legislative knowledge management. legislation, suggesting a dismantling of the predecessors’ To this aim, we demonstrated how triggers can be used to work. For instance, one of the rules with the highest sup- build a simple but effective warning system that monitors port and confidence values reveals that the fourth Berlusconi whether legislative voids are created. Finally, we used asso- government (right-wing) reshaped procurement legislation ciation rules to discover significant patterns in the evolution previously modified or enacted by the second Prodi govern- of the Italian legislation referred to procurement, discover- ment (left-wing). ing an increasing trend of instability in such legislation. In future work, we will further explore the utility of these tools for knowledge discovery, refining their applications in the 5. Conclusions legislative domain with the help of experts and identifying new domains that could benefit from these advancements. In this paper, we built upon recent efforts in modelling legislative systems within graph databases, describing the 4 The implementation is available at [34] References formation Processing and Management 62 (2025) 104082. URL: https://www.sciencedirect.com/ [1] C. Lupo, F. Vitali, E. Francesconi, M. Palmirani, science/article/pii/S030645732500024X. doi:https: R. Winkels, E. de Maat, A. Boer, P. Mascellani, et al., //doi.org/10.1016/j.ipm.2025.104082 . General xml format (s) for legal sources., Technical [12] OASIS, Akoma ntoso version 1.0 becomes report, IST-2004-027655 ESTRELLA European project an oasis standard, 2018. Available online at: for Standardised Transparent Representations in order https://www.oasis-open.org/news/announcements/ to Extend Legal Accessibility (2007). akoma-ntoso-version-1-0-becomes-an-oasis-standard/, [2] R. Hoekstra, J. Breuker, M. Di Bello, A. Boer, et al., last accessed on 01.05.2024. The lkif core ontology of basic legal concepts., CEUR [13] F. Vitali, M. Palmirani, et al., Akoma ntoso: flexibility Workshop Proceedings 321 (2007) 43–63. 2nd Work- and customization to meet different legal traditions, shop on Legal Ontologies and Artificial Intelligence BALISAGE SERIES ON MARKUP TECHNOLOGIES Techniques, LOAIT 2007 ; Conference date: 04-06-2007 24 (2019) 1–9. Through 04-06-2007. [14] A. Colombo, Knowledge graph of the italian leg- [3] T. Athan, H. Boley, G. Governatori, M. Palmirani, islation, 2024. URL: https://zenodo.org/doi/10.5281/ A. Paschke, A. Wyner, Oasis legalruleml, in: Pro- zenodo.13798158. doi:10.5281/ZENODO.13798158 . ceedings of the Fourteenth International Conference [15] K. P. Eswaran, Aspects of a trigger subsystem in an on Artificial Intelligence and Law, ICAIL ’13, Associa- integrated database system, in: Proceedings of the 2nd tion for Computing Machinery, New York, NY, USA, International Conference on Software Engineering, 2013, p. 3–12. ICSE ’76, IEEE Computer Society Press, Washington, [4] G. Barabucci, L. Cervone, M. Palmirani, S. Peroni, DC, USA, 1976, p. 243–250. F. Vitali, Multi-layer markup and ontological struc- [16] J. Widom, S. Ceri, Active database systems: Triggers tures in akoma ntoso, in: International Workshop on and rules for advanced database processing, Morgan AI Approaches to the Complexity of Legal Systems, Kaufmann, 1995. Springer, Springer, Beijing, China, 2009, pp. 133–149. [17] J. Melton, A. R. Simon, SQL: 1999: understanding rela- doi:10.1007/978- 3- 642- 16524- 5_9 . tional language components, Morgan Kaufmann Pub- [5] W. Linder, Swiss legislation in the era of globali- lishers Inc., San Francisco, CA, USA, 2001. sation: A quantitative assessment of federal legis- [18] S. Schmitz, M. Schluetter, U. Epple, Automation of lation (1983-2007), Swiss Political Science Review automation — definition, components and challenges, 20 (2014) 223–231. doi:https://doi.org/10.1111/ in: 2009 IEEE Conference on Emerging Technologies spsr.12100 . and Factory Automation, 2009, pp. 1–7. doi:10.1109/ [6] A. Bonifati, D. Braga, A. Campi, S. Ceri, Active xquery, ETFA.2009.5347197 . in: Proceedings 18th International Conference on Data [19] L. Bellomarini, A. Bernasconi, S. Ceri, A. Gagliardi, Engineering, 2002, pp. 403–412. doi:10.1109/ICDE. D. Magnanimi, D. Martinenghi, et al., Towards a stan- 2002.994754 . dard for triggers in property graphs, in: CEUR WORK- [7] R. Angles, A. Bonifati, S. Dumbrava, G. Fletcher, SHOP PROCEEDINGS, volume 3741, CEUR-WS, 2024, A. Green, J. Hidders, B. Li, L. Libkin, V. Marsault, pp. 70–79. W. Martens, F. Murlak, S. Plantikow, O. Savkovic, [20] S. Ceri, J. Widom, Deriving incremental produc- M. Schmidt, J. Sequeda, S. Staworko, D. Tomaszuk, tion rules for deductive data, Information Systems H. Voigt, D. Vrgoc, M. Wu, D. Zivkovic, Pg-schema: 19 (1994) 467–490. URL: https://www.sciencedirect. Schemas for property graphs, Proc. ACM Manag. com/science/article/pii/0306437994900027. doi:https: Data 1 (2023). URL: https://doi.org/10.1145/3589778. //doi.org/10.1016/0306- 4379(94)90002- 7 . doi:10.1145/3589778 . [21] R. Agrawal, T. Imieliński, A. Swami, Mining associa- [8] S. Ceri, A. Bernasconi, A. Gagliardi, D. Martinenghi, tion rules between sets of items in large databases, in: L. Bellomarini, D. Magnanimi, Pg-triggers: Trig- Proceedings of the 1993 ACM SIGMOD international gers for property graphs, in: Companion of the conference on Management of data, 1993, pp. 207–216. 2024 International Conference on Management of [22] R. Srikant, R. Agrawal, Mining generalized association Data, SIGMOD/PODS ’24, Association for Comput- rules, Future generation computer systems 13 (1997) ing Machinery, New York, NY, USA, 2024, p. 373–385. 161–180. URL: https://doi.org/10.1145/3626246.3653386. doi:10. [23] S. W. Changchien, T.-C. Lu, Mining association rules 1145/3626246.3653386 . procedure to support on-line recommendation by cus- [9] F. Cambria, F. Invernici, A. Bernasconi, S. Ceri, Mine tomers and products fragmentation, Expert systems graph rule: A new cypher-like operator for mining with applications 20 (2001) 325–335. association rules on property graphs, arXiv preprint [24] T. Osadchiy, I. Poliakov, P. Olivier, M. Rowland, E. Fos- arXiv:2406.19106 (2024). ter, Recommender system based on pairwise associ- [10] A. Colombo, Leveraging knowledge graphs and llms ation rules, Expert Systems with Applications 115 to support and monitor legislative systems, in: Pro- (2019) 535–542. ceedings of the 33rd ACM International Conference on [25] M. Tandan, Y. Acharya, S. Pokharel, M. Timilsina, Dis- Information and Knowledge Management, CIKM ’24, covering symptom patterns of covid-19 patients using Association for Computing Machinery, New York, NY, association rule mining, Computers in biology and USA, 2024, p. 5443–5446. URL: https://doi.org/10.1145/ medicine 131 (2021) 104249. 3627673.3680268. doi:10.1145/3627673.3680268 . [26] H. Si, J. Zhou, Z. Chen, J. Wan, N. N. Xiong, W. Zhang, [11] A. Colombo, A. Bernasconi, S. Ceri, An llm- A. V. Vasilakos, Association rules mining among in- assisted etl pipeline to build a high-quality terests and applications for users on social networks, knowledge graph of the italian legislation, In- IEEE Access 7 (2019) 116014–116026. [27] A. Bernasconi, F. Invernici, S. Ceri, et al., Tetys: To- wards the next-generation open-source web topic ex- plorer, in: CEUR WORKSHOP PROCEEDINGS, vol- ume 3692, CEUR-WS, 2024, pp. 26–33. [28] R. Y. Maragheh, C. Fang, C. C. Irugu, P. Parikh, J. Cho, J. Xu, S. Sukumar, M. Patel, E. Korpeoglu, S. Kumar, et al., Llm-take: theme-aware keyword extraction using large language models, in: 2023 IEEE Interna- tional Conference on Big Data (BigData), IEEE, 2023, pp. 4318–4324. [29] K. W. Dam, The law-growth nexus: The rule of law and economic development, Rowman & Littlefield, 2007. [30] D. Magnanimi, Neo4j APOC light-weight reasoner, https://github.com/davidemagnanimi/apoc, 2023. [31] B. Efron, Bootstrap Methods: Another Look at the Jackknife, Springer New York, New York, NY, 1992, pp. 569–593. URL: https://doi.org/10.1007/978-1-4612-4380-9_41. doi:10.1007/978- 1- 4612- 4380- 9_41 . [32] J. L. Rasmussen, Estimating correlation coefficients: Bootstrap and parametric approaches., Psychological Bulletin 101 (1987) 136. [33] F. Decarolis, R. Fisman, P. Pinotti, S. Vannutelli, Rules, discretion, and corruption in procurement: Evidence from italian government contracting, Technical Re- port, National Bureau of Economic Research, 2020. [34] F. Invernici, F. Cambria, Neo4j Mine Graph Rule Opera- tor, https://github.com/FrInve/mine_graph_rule, 2024.