{"id":21561,"date":"2025-06-04T15:41:05","date_gmt":"2025-06-04T13:41:05","guid":{"rendered":"https:\/\/www.ilc.cnr.it\/four-cnr-ilc-papers-accepted-at-acl-2025\/"},"modified":"2025-06-04T16:04:58","modified_gmt":"2025-06-04T14:04:58","slug":"four-cnr-ilc-papers-accepted-at-acl-2025","status":"publish","type":"post","link":"https:\/\/www.ilc.cnr.it\/en\/four-cnr-ilc-papers-accepted-at-acl-2025\/","title":{"rendered":"Four CNR-ILC papers accepted at ACL 2025"},"content":{"rendered":"\n<p>Four papers of the <a href=\"http:\/\/www.italianlp.it\/\" target=\"_blank\" rel=\"noreferrer noopener\">ItaliaNLP Lab<\/a> of the <strong>Cnr-Istituto di Linguistica Computazionale \u201cAntonio Zampolli\u201d <\/strong>(<strong>CNR-ILC<\/strong>) have been accepted at <a href=\"https:\/\/2025.aclweb.org\/\" target=\"_blank\" rel=\"noreferrer noopener\">ACL 2025<\/a>, the 63rd Annual Meeting of the <a href=\"https:\/\/www.aclweb.org\/portal\/\" target=\"_blank\" rel=\"noreferrer noopener\">Association for Computational Linguistics (ACL)<\/a>.<\/p>\n\n\n\n<div style=\"height:30px\" aria-hidden=\"true\" class=\"wp-block-spacer\"><\/div>\n\n\n\n<p><strong>&#8220;From Human Reading to NLM Understanding: Evaluating the Role of Eye-Tracking Data in Encoder-Based Models&#8221; (Luca Dini, Lucia Domenichelli, Dominique Brunato, Felice Dell\u2019Orletta)<\/strong><br>This work, accepted at the main conference, explores how integrating human eye-tracking data into language models affects task performance, attention mechanisms, and representation space &#8211; all together for the first time.<\/p>\n\n\n\n<p>Key insights:  <\/p>\n\n\n\n<ul>\n<li>ET signals make model attention more aligned with human gaze, even after downstream fine-tuning<\/li>\n\n\n\n<li>they compress the model\u2019s representation space &#8211; more efficient, yet still strong on tasks <\/li>\n\n\n\n<li>full fine-tuning strategies are the most robust across all dimensions, while partial tuning maximizes attention alignment  <\/li>\n<\/ul>\n\n\n\n<p>This study highlights the potential of cognitive signals to build more interpretable, efficient, and human-aligned AI systems.<\/p>\n\n\n\n<div style=\"height:30px\" aria-hidden=\"true\" class=\"wp-block-spacer\"><\/div>\n\n\n\n<p><strong>&#8220;Evaluating Lexical Proficiency in Neural Language Models&#8221; (Cristiano Ciaccio, Alessio Miaschi, Felice Dell&#8217;Orletta)<\/strong><br>This study, accepted at the main conference, proposes a novel and unified framework to evaluate lexical proficiency in Transformer-based language models, testing their ability to generate, define, and use words across a range of lexical categories: commonly lexicalized words, recent neologisms and nonce words with an emphasis on the creative aspects of this last category. <br><br>Key contributions: <\/p>\n\n\n\n<ul>\n<li>proposal of a novel framework to assess lexical abilities across diverse tasks and word types.<\/li>\n\n\n\n<li>developement of a new lexical resource for Italian, with curated definitions and usage examples. <\/li>\n\n\n\n<li>evaluation on how model size, multilinguality and linguistic features affect lexical generalization. <\/li>\n\n\n\n<li>a human evaluation based on the Optimal Innovation Hypothesis to assess the plausibility and creativity of generated nonce words<\/li>\n<\/ul>\n\n\n\n<p>The findings show that Transformer-based models can handle lexical composition and meaning inference to some extent, effectively producing and interpreting plausible lexical innovations, although with a substantial drop in performance compared to standard lexical items. <\/p>\n\n\n\n<p>Code &amp; dataset: <a href=\"https:\/\/github.com\/snizio\/Lexical-Proficiency\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/snizio\/Lexical-Proficiency<\/a> <\/p>\n\n\n\n<div style=\"height:30px\" aria-hidden=\"true\" class=\"wp-block-spacer\"><\/div>\n\n\n\n<p><strong>&#8220;Beyond the Spelling Miracle: Investigating Substring Awareness in Character-Blind Language Models&#8221; (Cristiano Ciaccio, Marta Sartor, Alessio Miaschi, Felice Dell\u2019Orletta)<\/strong>. <br>Language Models (LMs) typically operate on subword tokens and lack explicit access to characters. Despite so, they show a limited, but surprising, ability to recognize spelling-level patterns &#8211; a phenomenon known as the Spelling Miracle. <\/p>\n\n\n\n<p>This work, accepted at the Findings of ACL, takes a systematic look at when, where, and how such character-level awareness emerges. The study proposes a controlled binary task &#8211; no prompting, no probing &#8211; asking models whether a substring appears in a word. Using the MorphoLex database, it evaluate models from the Pythia family across: <\/p>\n\n\n\n<ol>\n<li>substring position and length<\/li>\n\n\n\n<li>morphemic vs. non-morphemic substrings (prefixes, suffixes, roots)<\/li>\n\n\n\n<li>pre-training checkpoints, including from scratch<\/li>\n<\/ol>\n\n\n\n<p>Key findings:<\/p>\n\n\n\n<ul>\n<li>larger models develop more robust substring awareness<\/li>\n\n\n\n<li>morphemes &#8211; especially suffixes and roots &#8211; are recognized better than meaningless substrings<\/li>\n\n\n\n<li>awareness emerges early for suffixes and roots, later for non-morphemic units, especially in the middle of words<\/li>\n\n\n\n<li>linguistic features like productivity, word frequence and tokenization shape this ability. <\/li>\n<\/ul>\n\n\n\n<p>This work opens up new directions for analyzing character knowledge in LMs providing both conceptual and empirical grounding for the Spelling Miracle. <\/p>\n\n\n\n<div style=\"height:30px\" aria-hidden=\"true\" class=\"wp-block-spacer\"><\/div>\n\n\n\n<p><strong>&#8220;Stress-testing Machine Generated Text Detection: Shifting Language Models Writing Style to Fool Detectors&#8221;<\/strong> <strong>(Andrea Pedrotti, Michele Papucci, Cristiano Ciaccio, Alessio Miaschi, Giovanni Puccetti, Felice Dell\u2019Orletta, Andrea Esuli)<\/strong><br>As LLMs become increasingly capable of producing human-like text, the task of detecting Machine-Generated Text (MGT) is more important &#8211; and more difficult -than ever. <br><br>This work, accepted at the Findings of ACL, shows just how fragile current state-of-the-art MGT detectors really are. The authors fine-tune LLMs using Direct Preference Optimization (DPO) to subtly shift the style of synthetic text toward human-written text (HWT), creating adversarial examples that significantly reduce detection accuracy. The result? Detection performance drops dramatically, even with minor stylistic alignment. <br><br>Other contributions: <\/p>\n\n\n\n<ul>\n<li>analysis of the linguistic &#8220;shortcuts&#8221; used by detectors (and how easily they can be bypassed)<\/li>\n\n\n\n<li>comparison of how humans vs. machines detect MGTs &#8211; and how little they overlap<\/li>\n\n\n\n<li>release code, models, and data to help the community build more robust and realistic benchmarks.<\/li>\n<\/ul>\n\n\n\n<p>This work underscores the urgent need to move beyond shallow heuristics in detection &#8211; and to build systems that generalize across domains and stylistic shifts. <\/p>\n\n\n\n<p>The preprint of the paper is available at the following link: <a href=\"https:\/\/arxiv.org\/pdf\/2505.24523\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/arxiv.org\/pdf\/2505.24523<\/a> <\/p>\n","protected":false},"excerpt":{"rendered":"<p>Four papers of the ItaliaNLP Lab of the Cnr-Istituto di Linguistica Computazionale \u201cAntonio Zampolli\u201d (CNR-ILC) have been accepted at ACL&hellip;<\/p>\n","protected":false},"author":3,"featured_media":21545,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"inline_featured_image":false,"footnotes":"","_jetpack_memberships_contains_paid_content":false},"categories":[8],"tags":[],"tag-sottositi":[47],"acf":[],"fimg_url":"https:\/\/www.ilc.cnr.it\/wp-content\/uploads\/2025\/06\/ACL_2025.png","jetpack_sharing_enabled":true,"jetpack_featured_media_url":"https:\/\/www.ilc.cnr.it\/wp-content\/uploads\/2025\/06\/ACL_2025.png","publishpress_future_action":{"enabled":false,"date":"2026-04-23 22:56:15","action":"change-status","newStatus":"draft","terms":[],"taxonomy":"category"},"publishpress_future_workflow_manual_trigger":{"enabledWorkflows":[]},"_links":{"self":[{"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/posts\/21561"}],"collection":[{"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/comments?post=21561"}],"version-history":[{"count":3,"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/posts\/21561\/revisions"}],"predecessor-version":[{"id":21567,"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/posts\/21561\/revisions\/21567"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/media\/21545"}],"wp:attachment":[{"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/media?parent=21561"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/categories?post=21561"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/tags?post=21561"},{"taxonomy":"tag-sottositi","embeddable":true,"href":"https:\/\/www.ilc.cnr.it\/en\/wp-json\/wp\/v2\/tag-sottositi?post=21561"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}