Year: 2026
Type: conference-paper
Abstract: We present our submission to Task 1 of the Cruciverb-IT challenge at EVALITA 2026, which focuses on answering clues extracted from Italian crosswords. The task is framed as a constrained question-answering problem, to generate a ranked list of candidate solutions for each clue, given the expected answer length. To address this problem, we explore three complementary approaches: a lexical retrieval model based on BM25, a dense semantic retrieval model that relies on sentence-level embeddings, and a hybrid retrieve-and-rerank architecture that combines the outputs of the two retrieval strategies using a large language model as a judge. Experimental results on the official evaluation benchmark show that dense semantic retrieval achieves the highest overall performance, whereas both lexical and hybrid methods consistently outperform the official task baseline. These findings highlight the effectiveness of embedding-based representations for Italian crossword-clue answering and provide insights into the role of hybrid retrieval and reranking strategies in this task.
Source: PORTO@Iris
Institution: Politecnico di Torino
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7218089687