The Ministry of Justice performs administrative functions in both the civil and criminal fields. The judiciary is a complex of structures and institutions aimed at the administration of justice, overseen by individual judges. The primary activities of the Ministry of Justice and the judges (collectively referred to as Justice) concern criminal and civil proceedings. The criminal proceeding includes preliminary investigations, activities of cognition in the three levels of judgment in the criminal process, and the execution of penalties or alternative activities in juvenile and community justice. Similarly, the civil proceeding consists of a cognition phase, which includes three levels of judgment, and an execution phase. The cognition phase of the civil proceeding has long been subject to automation within the On-Line Civil Trial (in Italian abbreviated as PCT) information system. Consequently, the digitization of structured data and documents in the civil proceedings files is significantly more advanced than in the preliminary investigations and criminal proceedings files. The digital file of a civil proceeding consists of acts and documents, and, for concluded proceedings, the judgment. An act of the civil proceeding is a documentary artifact related to a file, whose content and form are prescribed by regulations. A document is any artifact (text, audio recording, image, video, etc.) related to the file and attached to acts. The progress of the civil proceeding is represented in terms of states and events. In the first phase of the civil proceeding and, to a greater extent, in the preliminary investigative phase of the criminal proceeding, numerous documentary sources of evidence are acquired, including telephone records, credit card traces, inspections, transcribed telephone interceptions, and many others. The primary activities of preliminary investigations and cognition of the criminal proceedings have only recently become the subject of automation. The execution phase of criminal proceedings is characterized by greater automation compared to civil proceedings, with the Judiciary Record (in Italian "Casellario Giudiziale") and databases of the Department of Penitentiary Administration and the Department of Juvenile and Community Justice being the main realizations. The Datalake project was initiated in 2019 by the DGSIA, which entrusted the CRUI and, subsequently, the Consorzio Interuniversitario Nazionale per l'Informatica (CINI), with a renewed line of research over the years, investigating the adoption of technologies based on natural language processing (NLP), knowledge graphs, machine learning, and, recently, generative AI, which can be most useful in carrying out the primary processes of civil and criminal cognition and execution. In 2022, Justice included Datalake among the projects funded with PNRR funds, whose implementation was entrusted through a tender to a group of three companies: Almaviva, Almawave, and Accenture.
The Datalake project focused on the preliminary investigation phase of criminal proceedings, criminal enforcement, and the cognition phase of civil proceedings. The functionalities developed originate from a user requirements elicitation activity, which for the criminal procedure involves Prosecutors and the Judicial Police, and for the civil procedure involves Judges. The requirements were collected in the preliminary investigation phase, through the sharing of Proof of Concepts on defined procedures, and subsequently through consultations on ongoing procedures. In the civil domain, interviews were conducted with Judges of the Court of Appeal of Milan, who will soon experiment with a first set of functionalities developed by the supplier Almawave. The outcome of the experiment will result in a new version of the system that can be adopted in all Courts of Appeal. The user needs are briefly described below.
• Predictive models of the expected durations and variability of the proceedings based on their characteristics. • Assessment of the expected complexity of the proceedings for the distribution of workload among judges, identification of "bottlenecks", identification of signals and events that significantly impact the duration of the proceedings, and analysis of the impact of changes in laws, regulations, and practices. • Descriptive statistics on structured data and judgments. • Correlation analysis between salient characteristics and outcomes in civil proceedings for uniformity purposes (so-called "tabulation"). • Comparative analysis of trial durations in different sections and districts.
The Proof of Concept developed and the functional specifications produced within the Datalake project concern the following macro-functionalities: Preparation, Semantic Enrichment and Knowledge Integration, Semantic Search and Analysis, Knowledge Base Management, and Quality Control. The following are the detailed functionalities.
1. Document pre-processing (removal of special characters, correction of accented letters, removal of headers, removal of stamps, punctuation management) 2. OCR and generation of interpretable documents.
preamble, case description, and decision. 4. Classification of texts within the files.
Semantic enrichment is performed by extracting information from documents, especially named entities and terms, and persisting the result of this extraction process into semantic annotations. This process is in use in legal AI to a large extent. The peculiar characteristic of the proposed approach lies in the effort to consolidate the knowledge extracted by linking different mentions that refer to the same entities (exploiting background knowledge bases like Wikipedia and clustering mentions of the entities -of course, the majority -that are not present in Wikipedia) [1,2,3]. The impact of this approach is particularly noticeable during document search (see functionality F3).
Civil Trials. Various NLP techniques have been applied to extract, link, and consolidate entity mentions from judgments and produce semantic annotations that associate the extracted entities with specific token sequences in the judgments. In particular, the current pipeline combines the following techniques [ Criminal Trials -Preliminary Investigations. For documents related to preliminary investigations, a very similar pipeline was applied for entity extraction and subsequent document annotation, a similar semantic search paradigm. A first discussion of the application of entitycentric approaches to manage documents in preliminary investigations can be found in [5]. However, other functionalities and techniques were applied such as:
• Extraction of graph representations from instant messaging applications (IMA) data, e.g., What-sApp dumps, and storage in a graph DB (Neo4J); messages can be queried using a structured language that supports graph-based data analysis. • Content enrichment with speech-to-text technology; OpenAI's Whisper was used to transcribe audio messages and make these contents searchable. All messages and chats are analyzed using small adaptations of the NLP pipeline described earlier, supporting semantic search powered by entity-based annotations. • Semantic enrichment and specialization of entity annotation ontologies relative to specific taxonomy (is-overlapping, is-within, ordering). Other developed functionalities include domain concept extraction, text summarization, and georeferencing of spatial entities. For all functionalities, accuracy analyses were conducted based on scientific methodologies.
For the entity extraction pipeline, some results are reported in [1,2]. As examples of accuracy measured for relation and feature extraction capability, we report accuracy for the Relationship "against", 83.5%, and for the extraction of the maintenance payment in favor of children in separation cases, 77.52%.
Common Search Functionalities for Preliminary Investigations and Civil Proceedings. Search functionalities are inspired by the well-known faceted and semantic search paradigms, with additional and more experimental Question Answering (QA) capabilities based on the Retrieval Augmented Generation (RAG) paradigm. Based on the semantic enrichment functionalities shown in the previous point, the entities that appear in the filters during the search phase can refer to mentions present in different documents; moreover, when a user explores, for example, a judgment, they can find all mentions of an entity throughout the document, a feature that can become particularly relevant for long judgments or other documents. The conceptual architecture for semantic search is shown in Fig 1 . The components are:
• Keyword search: Allows simple keyword searches.
This module can be useful as a starting point for the search, before activating the faceted search. The above functionalities have all been demonstrated using DAVE, a prototype open-source application for semantic search developed in the context of this and the PON Next Generation UPP1 project. A video demonstrating the proposed combination of semantic and conversational search on judgments of criminal trials published online is available at https://www.youtube.com/watch? v=XG7RsI3t-2Q. However, the data enrichment process developed in the project supports also other forms of search, such as Advanced search. This functionally supports advanced searches by combining various filters on document attributes. This module is included in many search applications on structured or semi-structured data, to complement the modules based on Keyword search and Faceted search; typically, the function of this module is to construct precise queries based on structured descriptions of documents.
Analysis Functionalities for Governance Activities -Civil. The semantic organization of documents obtained through semantic enrichment and integration functionalities enabled by the Entity registry allows for multiple statistics and correlations on structured data linked to annotated documents, e.g., the number of documents involving natural legal entities, the number and average value of minors involved in divorce decisions, correlation for tabulation purposes of the compensation value and related features in non-pecuniary damage cases. Further analyses concern survival curves of processes and explanatory variables of temporal duration and process complexity. Several analysis functionalities were developed within the PON Next Generation UPP project and other CRUI-funded projects. The following research based on the SICID system registers for the PCT was conducted (see [6,7]):
• Variant Analysis: Clusters of proceedings with the same structure and sequence of states and their evaluation for monitoring purposes. In particular, the factors that have the greatest impact on the duration of the processes were analyzed. For this activity, the process mining tool Apromore 2 was used. • Identification of Critical Events: The impact of specific events on the duration of a process execution is evaluated to identify events systematically associated with anomalous situations. Both the phases and the total duration of the proceedings were examined. • Predictive Approaches for Alerts: Predictors were constructed from sequences of states or events in the registers, based on machine learning techniques with LSTM neural networks, to predict the residual duration of processes and states during their course.
A management control dashboard was created for the Court of Cassation. The adopted solution was to create a dashboard directly fed by the underlying database of the Court's SIC register, with data updated four times a day. All data were identified for:
• Feeding the variables and indicators identified as necessary to describe the file path in the various phases and to calculate indices such as the Disposition Time and the turnover index; • Building the historical series of such data from January 2019.
Analysis Functionalities for Preliminary Investigations. Relational knowledge analysis with visualization (e.g., selection of clusters of nodes with certain properties) and anomaly detection.
Functionalities for Penal Execution. Integration for the social analysis of data relating to liberty restrictions/alternative penalties experienced by detainees during their lives.
1. Manual of Pseudonymization Trial Policies: Different types of pseudonymization are considered, and various types of data and document processing where pseudonymization is relevant (e.g., publication, linking databases, etc.) are identified, along with the properties that must be respected in each case. A general method is provided that can be followed for the different types of data processing relevant to the Datalake project. [8]). For accuracy and fairness, the manual aligns with policy documents issued by the EU (see [9]).
In the functionalities of the Datalake, the following ontologies are used:
• The following additional ontologies are represented in the form of two-level taxonomies: i) Top ontology of preliminary investigations, ii) Top ontology of the civil trial, iii) Domain ontologies of the civil process: banking, labor, non-patrimonial damage from privacy violation, judicial separation, iv) Ontology for penal-cognition procedure: victim-perpetrator relationship. The top ontology of Justice procedures and penal execution were produced through reverse engineering from logical schemas. The ontologies for the victim-perpetrator relationship and non-patrimonial damage were produced by domain experts. The ontologies for banking and labor were produced from lexicons built through the analysis of judgments.
The developed functionalities adopt a service architecture for deployment. The multi-node macro functional architecture is shown in Fig.
The semantic document system described in the work is potentially useful for all Public Administrations (PAs).
A project to disseminate the system should involve two phases: an initial phase of parameterization, concerning the organizational structure, ontologies, and primary and governance processes, and a second phase of customization (see Fig. 3). Such a project requires strong governance, aligning with the strategic path of digital transformation of the country, currently being implemented in the National Strategic Hub. Including services for a semantic document system for the PA in the service architecture of the Hub would require the production of a common top ontology for the PA and high-level modeling of primary and governance processes, with subsequent customization by the individual PAs.