Sustainability reports are often aligned with frameworks such as the Global Reporting Initiative (GRI) and the Sustainable Development Goals (SDGs), but large-scale, paragraph-level annotation remains a challenge. This paper introduces a fully automated pipeline that generates weak supervision by linking report paragraphs to GRI and SDG categories using structured content indices, official GRI-SDG mappings, and semantic similarity scoring. To mitigate the noise inherent in automatic annotation, we employ an instruction-tuned large language model (LLaMA 3.1) to filter assigned labels based on paragraph relevance. We evaluate the quality of our annotations through downstream SDG classification tasks on the OSDG Community Dataset, showing that LLM-based filtering aligns closely with human consensus and significantly improves model performance. Our results demonstrate that combining pruned, automatically annotated data with human-labeled examples leads to more accurate and robust SDG classification, supporting scalable, interpretable sustainability analysis.
This study investigates the persuasive and argumentative behaviors of two LLM-based chatbots, ChatGPT and Gemini, within the context of movie recommendation dialogues. Drawing on insights from argumentation-based dialogue and anthropomorphism research, we introduce a fine-grained annotation scheme to analyze chatbot strategies across dialogue phases. Through both linguistic analysis and user evaluation via ResQue and Godspeed questionnaires, we assess the systems' recommendation quality, perceived human-likeness, and strategic variation. Our findings reveal distinct conversational patterns: ChatGPT emphasizes affective engagement and trust-building, while Gemini adopts a more direct and efficiency-driven approach. These strategic differences are also reflected in the quality of the recommendation and the user perception. Gemini excels in recommendation quality and explanations, while ChatGPT performs better in emotional engagement, transparency, and user satisfaction.
Modern Large Language Models (LLMs) are commonly trained through a multi-stage pipeline encompassing pretraining and supervised finetuning. While recent studies have extensively investigated the benefits of continual pretraining on high-quality data, these efforts have focused primarily on English. In this work, we explore the effectiveness of various data mixtures in a continual pretraining setting to enhance performance on Italian-language tasks. Leveraging Minerva-7B, a fully opensource LLM pretrained on a corpus composed of 50% Italian, we define and evaluate three distinct data recipes-comprising mathematical, encyclopedic, and copyrighted content-spanning both Italian and English. We also investigate the effect of extending the model's context window during continual pretraining on its ability to handle long-context tasks. To support our evaluation, we introduce INDAQA, a new benchmark for narrative question answering in Italian. Our results reveal that both data composition and increased context length substantially improve performance, offering valuable insights into continual pretraining strategies for less represented languages within an open scientific framework.
This paper describes the inclusion of Sicilian in the Semantic Web through the development of new resources aligned with Linguistic Linked Open Data principles. More specifically, we model and publish the first Sicilian Lemma Bank and a bilingual Sicilian-Italian glossary extracted from the Sicilian Wiktionary (Wikizziunariu). These resources are formalized using the OntoLex-Lemon and LiLa (Linking Latin) ontologies with the aim of enabling cross-lingual interoperability. The glossary is also linked to the LiITA (Linking Italian) knowledge base. In addition, two preliminary experiments are reported: the first evaluates the translation capabilities of commercial Large Language Models (LLMs) from Sicilian into Italian; the second investigates bilingual lexicon induction through cross-lingual embedding alignment, with results indicating the challenges posed by low-resource dialects. This work aims to demonstrate the feasibility and importance of integrating under-resourced languages into broader Computational Linguistics and Semantic Web infrastructures.
While Automated Machine Learning (AutoML) systems have shown strong performance on structured data, their application to natural language processing (NLP) tasks remains limited by static, task-agnostic search spaces. In this work, we propose a context-aware extension of AutoPyTorch that dynamically adapts both the hyperparameter search space and neural architecture configuration based on corpus-level meta-features. Our approach extracts interpretable textual statistics-such as average sequence length, vocabulary richness, and class imbalance-to guide the configuration of key hyperparameters. We also introduce two adaptive neural backbones, whose structures are shaped by these meta-features to improve model expressiveness and generalization. Experiments on 20 diverse text classification datasets-including subsets of GLUE, selected Kaggle benchmarks, and private corpora-demonstrate consistent performance improvements over strong baselines, particularly on datasets with limited training samples or severe class imbalance. Our results highlight the effectiveness of integrating dataset-level insights into the AutoML search process for NLP.
