Event extraction is the task of automatically identifying and extracting structured information about events from unstructured text. Despite Italian being a well-resourced language, it still lacks annotated datasets specifically designed for fine-grained event extraction. To address this gap, we propose a novel methodology for the generation of synthetic data suitable for fine-grained event extraction tasks. This work is motivated by the high cost and limited scalability of manual annotation. We introduce a controlled synthetic data generation pipeline that strictly adheres to a target annotation schema, providing a scalable alternative to extensive human labeling. The key methodological innovation is a two-phase, document-level generation framework that leverages Large Language Models, ensures structural consistency and mitigates generation biases, enabling the creation of high-quality datasets for complex event extraction scenarios. Using this methodology, we release SYNTH-ITA, the first collection of four medium-scale synthetic datasets for fine-grained Italian event extraction, generated from 10,000 structured crime scenarios each. Experiments conducted on event argument extraction using a QA formulation demonstrate that fine-tuning models on SYNTH-ITA leads to better or comparable performances to models fine-tuned on 200 manually annotated real news articles (+14% improvement with ELECTRA, -0.4% with BERT). Conversely, NER-based models for event argument extraction trained on synthetic data exhibit an 18% performance drop compared to those trained on manually annotated articles. • First medium-scale synthetic datasets for Italian event extraction. • A defined methodology and pipeline to ensure structured annotations and text alignment. • Expert validation and bias mitigation enhance data quality and fairness. • Synthetic data boosts Italian QA performance. • Public release of dataset/code supports Italian NLP tasks and reproducibility.

Event extraction from unstructured text is a critical task in natural language processing, often requiring substantial annotated data. This study presents an approach to document-level event extraction applied to Italian crime news, utilizing large language models (LLMs) with minimal labeled data. Our method leverages zero-shot prompting and in-context learning to effectively extract relevant event information. We address three key challenges: (1) identifying text spans corresponding to event entities, (2) associating related spans dispersed throughout the text with the same entity, and (3) formatting the extracted data into a structured JSON. The findings are promising: LLMs achieve an F1-score of approximately 60% for detecting event-related text spans, demonstrating their potential even in resource-constrained settings. This work represents a significant advancement in utilizing LLMs for tasks traditionally dependent on extensive data, showing that meaningful results are achievable with minimal data annotation. Additionally, the proposed approach outperforms several baselines, confirming its robustness and adaptability to various event extraction scenarios. • Novel use of LLMs for event extraction from Italian news with minimal annotated data. • In-context learning outperforms zero-shot prompting for identifying event spans. • Mixtral achieves top performance in event extraction from Italian crime news. • LLMs outperform QA models, proving robust in data-scarce environments. • In-context learning performance depends on the quality of selected examples.

Extracting events from news stories as the aim of several Natural Language Processing (NLP) applications (e.g., question answering, news recommendation, news summarization) is not a trivial task, due to the complexity of natural language and the fact that news reporting is characterized by journalistic style and norms. Those aspects entail scattering an event description over several sentences within one document (or more documents), applying a mechanism of gradual specification of event-related information. This implies a widespread use of co-reference relations among the textual elements, conveying non-linear temporal information. In addition to this, despite the achievement of state-of-the-art results in several tasks, high-quality training datasets for non-English languages are rarely available.

Several studies have shown that the use of embeddings improves outcomes in many Natural Language Processing (NLP) activities, including text categorization.This paper focuses on how word embeddings can be used on newspaper articles related to crimes.The scope is the categorization of the news articles based on the type of crime they report.We compare different Word2Vec models and methods to obtain word embeddings.Then, we exploit both supervised and unsupervised Machine Learning categorization algorithms.Experiments were conducted on an Italian dataset of 15,361 crime news articles showing very promising results.