Leonardo Piano
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Leonardo Piano has a total of 31 co-authors. Here are the top co-authors by number of shared publications:
- Salvatore Carta
University of Cagliari
9 shared publications - Alessandro Giuliani
Environment and Health Department Istituto Superiore di Sanità Rome Italy
8 shared publications - Livio Pompianu
Department of Mathematics and Computer Science, University of Cagliari, Cagliari, Italy
7 shared publications - Alessandro Sebastian Podda
University of Cagliari
6 shared publications - Marco Manolo Manca
University of Cagliari
4 shared publications - Sandro Gabriele Tiddia
University of Cagliari
4 shared publications
Latest Publications
- Open Access
- 17 August 2026
- Published by Springer Nature in Neural Computing and Applications
- Vol. 38 (16) , 668
REMEDy: a dataset for rationale extraction and span-based moderation of dialogue prompts
The wide adoption of conversational AI systems necessitates urgent and interpretable safety moderation, especially given that Large Language Models (LLMs) continue to exhibit vulnerabilities despite alignment efforts, posing significant risks to individual users, organisations, and society. The ideal AI safety moderation system must be transparent and structurally interpretable. However, current moderation approaches typically rely on coarse classifications that offer limited interpretability and fail to capture the nuanced intent and contextual dependencies present in real-world user inputs. To advance moderation beyond these coarse labels, we present REMEDy, a novel dataset specifically built for extracting fine-grained rationales from user prompts. REMEDy features span-level annotations covering a broad taxonomy of safety-relevant categories, allowing for overlapping and nested textual spans to reflect complex prompt structures. Using REMEDy, we fine-tune multiple LLMs and evaluate their performance across two tasks: (i) rationale extraction, assessing their ability to accurately localise and classify harmful or ambiguous content; and (ii) prompt moderation, measuring improvements over state-of-the-art safety detectors. Our experiments demonstrate that REMEDy-trained models achieve competitive or superior moderation outcomes while simultaneously providing structured, human-readable rationales. REMEDy thus offers a valuable resource for developing safer, more transparent, and context-sensitive moderation systems.
- 17 August 2026
- Published by Springer Nature in Neural Computing and Applications
- Vol. 38 (16) , 668
Open AccessREMEDy: a dataset for rationale extraction and span-based moderation of dialogue prompts
The wide adoption of conversational AI systems necessitates urgent and interpretable safety moderation, especially given that Large Language Models (LLMs) continue to exhibit vulnerabilities despite alignment efforts, posing significant risks to individual users, organisations, and society. The ideal AI safety moderation system must be transparent and structurally interpretable. However, current moderation approaches typically rely on coarse classifications that offer limited interpretability and fail to capture the nuanced intent and contextual dependencies present in real-world user inputs. To advance moderation beyond these coarse labels, we present REMEDy, a novel dataset specifically built for extracting fine-grained rationales from user prompts. REMEDy features span-level annotations covering a broad taxonomy of safety-relevant categories, allowing for overlapping and nested textual spans to reflect complex prompt structures. Using REMEDy, we fine-tune multiple LLMs and evaluate their performance across two tasks: (i) rationale extraction, assessing their ability to accurately localise and classify harmful or ambiguous content; and (ii) prompt moderation, measuring improvements over state-of-the-art safety detectors. Our experiments demonstrate that REMEDy-trained models achieve competitive or superior moderation outcomes while simultaneously providing structured, human-readable rationales. REMEDy thus offers a valuable resource for developing safer, more transparent, and context-sensitive moderation systems.
- Open Access
- 10 April 2026
- Published by Springer Nature in Journal of Intelligent Manufacturing
KG4CUT: an ontology to facilitate cutting tool selection and interoperability
Selecting cutting tools for milling is a critical and complex task that directly affects product quality, cost, and operational efficiency. The growing diversity of tools and vendor-specific catalogues makes this process especially challenging, particularly for less experienced operators. In this paper, we present , an application ontology aligned with W3C Semantic Web standards and FAIR principles, designed to standardize and integrate cutting tool information across providers. To demonstrate its practical utility, we populated a knowledge graph using an automated pipeline that extracts structured data from real-world PDF catalogues. This graph serves as both a proof of concept and a functional basis for intelligent tool recommendation and cutting parameter retrieval, based on material properties, operation types, and geometric constraints. Evaluation with domain experts showed improved retrieval efficiency and reduced selection errors. thus supports the digitalization of machining knowledge and enables faster, more accurate process planning in industrial settings.
- 10 April 2026
- Published by Springer Nature in Journal of Intelligent Manufacturing
Open AccessKG4CUT: an ontology to facilitate cutting tool selection and interoperability
Selecting cutting tools for milling is a critical and complex task that directly affects product quality, cost, and operational efficiency. The growing diversity of tools and vendor-specific catalogues makes this process especially challenging, particularly for less experienced operators. In this paper, we present , an application ontology aligned with W3C Semantic Web standards and FAIR principles, designed to standardize and integrate cutting tool information across providers. To demonstrate its practical utility, we populated a knowledge graph using an automated pipeline that extracts structured data from real-world PDF catalogues. This graph serves as both a proof of concept and a functional basis for intelligent tool recommendation and cutting parameter retrieval, based on material properties, operation types, and geometric constraints. Evaluation with domain experts showed improved retrieval efficiency and reduced selection errors. thus supports the digitalization of machining knowledge and enables faster, more accurate process planning in industrial settings.
- 1 April 2026
- Published by Springer Nature
- p. 500-516
Small but Dangerous: Evaluating and Mitigating Jailbreak Vulnerabilities in Small Language Models
- 1 April 2026
- Published by Springer Nature
- p. 500-516
Small but Dangerous: Evaluating and Mitigating Jailbreak Vulnerabilities in Small Language Models
- 3 February 2026
- Published by Springer Nature in Neural Computing and Applications
- Vol. 38 (3) , 33
Are LLMs adequate SPARQL query generators? Investigating zero-shot NL-to-SPARQL translation
- 3 February 2026
- Published by Springer Nature in Neural Computing and Applications
- Vol. 38 (3) , 33
Are LLMs adequate SPARQL query generators? Investigating zero-shot NL-to-SPARQL translation
- 13 January 2026
- Published by Springer Nature
- p. 271-283
Preserving Sardinian Linguistic Diversity: An Automatic Language Identifier Approach
- 13 January 2026
- Published by Springer Nature
- p. 271-283
Preserving Sardinian Linguistic Diversity: An Automatic Language Identifier Approach
- Open Access
- 29 August 2025
- Published by Springer Nature in Journal of Intelligent Information Systems
- Vol. 63 (6) , 2147-2175
LLIMONIIE: Large Language Instructed Model for Open Named Italian Information Extraction
The exponential growth of unstructured documents generated daily underscores the urgent need to develop technologies to structure information effectively. Traditional Information Extraction (IE) models enable the transformation of textual data into structured formats (e.g., semantic triplets), facilitating efficient searches and uncovering hidden data insights. However, they require predefined ontologies and, often, extensive human efforts. On the other hand, Open IE tools extract information without any input knowledge, but they are limited in capturing entire and in-depth contexts. Furthermore, the state of the art presents a substantial discrepancy between the efforts carried out in English-centric methods and those in low-resource languages, such as Italian. Our study aims to address the aforementioned key challenges. To this end, we first define Open Named Information Extraction (ONIE), an approach that generalizes IE across diverse domains without requiring input ontologies and captures complex relationships. Then, we develop LLIMONIIE (Large Language Instructed Model for Open Named Italian Information Extraction), a novel end-to-end generative information extraction framework that leverages the capabilities of Large Language Models (LLMs) to perform ONIE from Italian documents, able to extract Named Entities and Open Relations uniformly. Furthermore, we devise an innovative dataset generation methodology to support our research. Finally, we release the code and dataset, contributing to the scientific community and the development of low-resource languages. Experiments demonstrate the potential of our proposal, achieving competitive results compared to the actual state of the art of Italian IE.
- 29 August 2025
- Published by Springer Nature in Journal of Intelligent Information Systems
- Vol. 63 (6) , 2147-2175
Open AccessLLIMONIIE: Large Language Instructed Model for Open Named Italian Information Extraction
The exponential growth of unstructured documents generated daily underscores the urgent need to develop technologies to structure information effectively. Traditional Information Extraction (IE) models enable the transformation of textual data into structured formats (e.g., semantic triplets), facilitating efficient searches and uncovering hidden data insights. However, they require predefined ontologies and, often, extensive human efforts. On the other hand, Open IE tools extract information without any input knowledge, but they are limited in capturing entire and in-depth contexts. Furthermore, the state of the art presents a substantial discrepancy between the efforts carried out in English-centric methods and those in low-resource languages, such as Italian. Our study aims to address the aforementioned key challenges. To this end, we first define Open Named Information Extraction (ONIE), an approach that generalizes IE across diverse domains without requiring input ontologies and captures complex relationships. Then, we develop LLIMONIIE (Large Language Instructed Model for Open Named Italian Information Extraction), a novel end-to-end generative information extraction framework that leverages the capabilities of Large Language Models (LLMs) to perform ONIE from Italian documents, able to extract Named Entities and Open Relations uniformly. Furthermore, we devise an innovative dataset generation methodology to support our research. Finally, we release the code and dataset, contributing to the scientific community and the development of low-resource languages. Experiments demonstrate the potential of our proposal, achieving competitive results compared to the actual state of the art of Italian IE.
- Open Access
Evaluation of LLMs on Long-tail Entity Linking in Historical Documents
Entity Linking (EL) plays a crucial role in Natural Language Processing (NLP) applications, enabling the disambiguation of entity mentions by linking them to their corresponding entries in a reference knowledge base (KB). Thanks to their deep contextual understanding capabilities, LLMs offer a new perspective to tackle EL, promising better results than traditional methods. Despite the impressive generalization capabilities of LLMs, linking less popular, long-tail entities remains challenging as these entities are often underrepresented in training data and knowledge bases. Furthermore, the long-tail EL task is an understudied problem, and limited studies address it with LLMs. In the present work, we assess the performance of two popular LLMs, GPT and LLama3, in a long-tail entity linking scenario. Using MHERCL v0.1, a manually annotated benchmark of sentences from domain-specific historical texts, we quantitatively compare the performance of LLMs in identifying and linking entities to their corresponding Wikidata entries against that of ReLiK, a state-of-the-art Entity Linking and Relation Extraction framework. Our preliminary experiments reveal that LLMs perform encouragingly well in long-tail EL, indicating that this technology can be a valuable adjunct in filling the gap between head and long-tail EL.
Open AccessEvaluation of LLMs on Long-tail Entity Linking in Historical Documents
Entity Linking (EL) plays a crucial role in Natural Language Processing (NLP) applications, enabling the disambiguation of entity mentions by linking them to their corresponding entries in a reference knowledge base (KB). Thanks to their deep contextual understanding capabilities, LLMs offer a new perspective to tackle EL, promising better results than traditional methods. Despite the impressive generalization capabilities of LLMs, linking less popular, long-tail entities remains challenging as these entities are often underrepresented in training data and knowledge bases. Furthermore, the long-tail EL task is an understudied problem, and limited studies address it with LLMs. In the present work, we assess the performance of two popular LLMs, GPT and LLama3, in a long-tail entity linking scenario. Using MHERCL v0.1, a manually annotated benchmark of sentences from domain-specific historical texts, we quantitatively compare the performance of LLMs in identifying and linking entities to their corresponding Wikidata entries against that of ReLiK, a state-of-the-art Entity Linking and Relation Extraction framework. Our preliminary experiments reveal that LLMs perform encouragingly well in long-tail EL, indicating that this technology can be a valuable adjunct in filling the gap between head and long-tail EL.
- Open Access
- 10 September 2024
- Published by Springer Nature in Research Square
LLIMONIIE: Large Language Instructed Model for Open Named Italian Information Extraction
The exponential growth of unstructured documents generated daily underscores the urgent need to develop technologies to structure information effectively. Traditional Information Extraction (IE) models permit transforming textual data to structured formats (e.g., semantic triplets), enabling efficient searches and uncovering hidden data insights. However, they require predefined ontologies and, often, extensive human efforts. On the other hand, Open IE tools extract information without any input knowledge, but they are limited in capturing entire and in-depth contexts. Furthermore, the state of the art presents a substantial discrepancy between the efforts carried out in English-centric methods and other low-resource languages. Our study addresses the aforementioned key challenges: the need for a high-performative human effort agnostic model and overcoming the lack of Italian IE resources. To this end, we introduce Open Named Information Extraction (ONIE), a novel approach to generalize IE across diverse domains without requiring input ontologies and capturing complex relationships. In this work, we propose LLIMONIIE, a novel end-to-end framework that leverages the capabilities of Large Language Models (LLMs) to perform ONIE from entire documents, able to uniformly extract Named Entities and Open Relations. Furthermore, we devised an innovative dataset generation methodology to support our research on Italian IE. The code and dataset are publicly available, contributing to the scientific community and development of low-resource languages. Experiments demonstrate the potential of LLIMONIIE, achieving competitive results compared to the actual Italian information extraction state of the art.
- 10 September 2024
- Published by Springer Nature in Research Square
Open AccessLLIMONIIE: Large Language Instructed Model for Open Named Italian Information Extraction
The exponential growth of unstructured documents generated daily underscores the urgent need to develop technologies to structure information effectively. Traditional Information Extraction (IE) models permit transforming textual data to structured formats (e.g., semantic triplets), enabling efficient searches and uncovering hidden data insights. However, they require predefined ontologies and, often, extensive human efforts. On the other hand, Open IE tools extract information without any input knowledge, but they are limited in capturing entire and in-depth contexts. Furthermore, the state of the art presents a substantial discrepancy between the efforts carried out in English-centric methods and other low-resource languages. Our study addresses the aforementioned key challenges: the need for a high-performative human effort agnostic model and overcoming the lack of Italian IE resources. To this end, we introduce Open Named Information Extraction (ONIE), a novel approach to generalize IE across diverse domains without requiring input ontologies and capturing complex relationships. In this work, we propose LLIMONIIE, a novel end-to-end framework that leverages the capabilities of Large Language Models (LLMs) to perform ONIE from entire documents, able to uniformly extract Named Entities and Open Relations. Furthermore, we devised an innovative dataset generation methodology to support our research on Italian IE. The code and dataset are publicly available, contributing to the scientific community and development of low-resource languages. Experiments demonstrate the potential of LLIMONIIE, achieving competitive results compared to the actual Italian information extraction state of the art.
- Open Access
- 27 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 460-466
Towards Knowledge Graph Refinement: Misdirected Triple Identification
In the current digital transformation scenario, Knowledge Graphs (KGs) represent an across-the-board instrument for representing knowledge in a structured form. Such tools allow to effectively enhance the performance of Artificial Intelligence models in manifold contexts, such as reasoning or information retrieval. Nevertheless, the effectiveness of KGs is often affected by the incorrect directionality of some of their edges, due in most cases to human error or the inefficiency of automatic and semi-automatic graph creation methods. This paper proposes a classification-based approach to identify misdirected triples within a KG, aiming to support and assist humans in creating graph refinement. Triples are the main component of KGs, and they model the connection between nodes with a form. Our proposal allows us to refine a KG by devising a classification-based approach for recognizing whether the subjects and objects are not compliant with the logic directionality of the corresponding predicate, meaning that they should be switched (e.g., the triple should be inverted as ). We compare traditional machine learning techniques with cutting-edge advanced methods, including pre-trained language models and large language models. Extensive experiments have been performed across several datasets, confirming the effectiveness of our proposal.
- 27 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 460-466
Open AccessTowards Knowledge Graph Refinement: Misdirected Triple Identification
In the current digital transformation scenario, Knowledge Graphs (KGs) represent an across-the-board instrument for representing knowledge in a structured form. Such tools allow to effectively enhance the performance of Artificial Intelligence models in manifold contexts, such as reasoning or information retrieval. Nevertheless, the effectiveness of KGs is often affected by the incorrect directionality of some of their edges, due in most cases to human error or the inefficiency of automatic and semi-automatic graph creation methods. This paper proposes a classification-based approach to identify misdirected triples within a KG, aiming to support and assist humans in creating graph refinement. Triples are the main component of KGs, and they model the connection between nodes with a form. Our proposal allows us to refine a KG by devising a classification-based approach for recognizing whether the subjects and objects are not compliant with the logic directionality of the corresponding predicate, meaning that they should be switched (e.g., the triple should be inverted as ). We compare traditional machine learning techniques with cutting-edge advanced methods, including pre-trained language models and large language models. Extensive experiments have been performed across several datasets, confirming the effectiveness of our proposal.
- Open Access
- 27 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 467-473
Towards Zero-shot Knowledge Graph building: Automated Schema Inference
In the current Digital Transformation scenario, Knowledge Graphs are essential for comprehending, representing, and exploiting complex information in a structured form. The main paradigm in automatically generating proper Knowledge Graphs relies on predefined schemas or ontologies. Such schemas are typically manually constructed, requiring an intensive human effort, and are often sensitive to information loss due to negligence, incomplete analysis, or human subjectivity or inclination. Limiting human bias and the resulting information loss in creating proper Knowledge Graphs is paramount, particularly for user modeling in various sectors, such as education or healthcare. To this end, we propose a novel approach to automatically generating a proper entity schema. The devised methodology combines the language understanding capabilities of LLM with classical machine learning methods such as clustering to properly build an entity schema from a set of documents. This solution eliminates the need for human intervention and fosters a more efficient and comprehensive knowledge representation. The assessment of our proposal concerns adopting a state-of-the-art entity extraction model (UniNER) to estimate the relevance of the extracted entities based on the generated schema. Results confirm the potential of our approach, as we observed a negligible difference between the topic similarity score obtained with the ground truth and with the automatically generated schema (less than 1% on average on three different datasets). Such an outcome confirms that the proposed approach may be valuable in automatically creating an entity schema from a set of documents.
- 27 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 467-473
Open AccessTowards Zero-shot Knowledge Graph building: Automated Schema Inference
In the current Digital Transformation scenario, Knowledge Graphs are essential for comprehending, representing, and exploiting complex information in a structured form. The main paradigm in automatically generating proper Knowledge Graphs relies on predefined schemas or ontologies. Such schemas are typically manually constructed, requiring an intensive human effort, and are often sensitive to information loss due to negligence, incomplete analysis, or human subjectivity or inclination. Limiting human bias and the resulting information loss in creating proper Knowledge Graphs is paramount, particularly for user modeling in various sectors, such as education or healthcare. To this end, we propose a novel approach to automatically generating a proper entity schema. The devised methodology combines the language understanding capabilities of LLM with classical machine learning methods such as clustering to properly build an entity schema from a set of documents. This solution eliminates the need for human intervention and fosters a more efficient and comprehensive knowledge representation. The assessment of our proposal concerns adopting a state-of-the-art entity extraction model (UniNER) to estimate the relevance of the extracted entities based on the generated schema. Results confirm the potential of our approach, as we observed a negligible difference between the topic similarity score obtained with the ground truth and with the automatically generated schema (less than 1% on average on three different datasets). Such an outcome confirms that the proposed approach may be valuable in automatically creating an entity schema from a set of documents.