{"paper":{"title":"What Makes Words Hard? Sakura at BEA 2026 Shared Task on Vocabulary Difficulty Prediction","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Spelling difficulty and test item construction often drive ratings in standard vocabulary difficulty lists beyond genuine word production demands.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Adam Nohejl, Hitomi Yanaka, Maria Angelica Riera Machin, Xuanxin Wu, Yi-Ning Chang, Yusuke Ide","submitted_at":"2026-05-14T01:57:35Z","abstract_excerpt":"We describe two types of models for vocabulary difficulty prediction: a high-accuracy black-box model, which achieved the top shared task result in the open track, and an explainable model, which outperforms a fine-tuned encoder baseline. As the black-box model, we fine-tuned an LLM using a soft-target loss function for effective application to the rating task, achieving r > 0.91. The explainable model provides insights into what impacts the difficulty of each item while maintaining a strong correlation (r > 0.77). We further analyze the results, demonstrating that the difficulty of items in t"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"The black-box model achieved r > 0.91 and topped the open track, while the explainable model reached r > 0.77 and showed that KVL item difficulty is affected by spelling difficulty or test item construction in addition to genuine production difficulty.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the shared task dataset and KVL lists provide a clean measure of genuine word production difficulty without significant confounding from test design or spelling factors that the models are capturing post-hoc.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Fine-tuned LLM with soft-target loss tops shared task on vocabulary difficulty prediction at r>0.91 while explainable model at r>0.77 shows spelling and item construction affect difficulty beyond word production.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Spelling difficulty and test item construction often drive ratings in standard vocabulary difficulty lists beyond genuine word production demands.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"9c8226b5302df29a661301952e5966be951009a708a70696529cc78c8e828497"},"source":{"id":"2605.14257","kind":"arxiv","version":1},"verdict":{"id":"c38e04d0-f0a0-4ba5-b797-8d89ebdb6037","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T02:56:02.140941Z","strongest_claim":"The black-box model achieved r > 0.91 and topped the open track, while the explainable model reached r > 0.77 and showed that KVL item difficulty is affected by spelling difficulty or test item construction in addition to genuine production difficulty.","one_line_summary":"Fine-tuned LLM with soft-target loss tops shared task on vocabulary difficulty prediction at r>0.91 while explainable model at r>0.77 shows spelling and item construction affect difficulty beyond word production.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the shared task dataset and KVL lists provide a clean measure of genuine word production difficulty without significant confounding from test design or spelling factors that the models are capturing post-hoc.","pith_extraction_headline":"Spelling difficulty and test item construction often drive ratings in standard vocabulary difficulty lists beyond genuine word production demands."},"references":{"count":35,"sample":[{"doi":"10.18653/v1/2022.tsar-1.28","year":2022,"title":"Dennis Aumiller and Michael Gertz. 2022. https://doi.org/10.18653/v1/2022.tsar-1.28 U ni HD at TSAR -2022 shared task: Is compute all we need for lexical simplification? In Proceedings of the Workshop","work_id":"cffb6528-85d4-4b3f-b7c2-0d9152671c69","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2007,"title":"BNC Consortium . 2007. https://llds.ling-phil.ox.ac.uk/llds/xmlui/handle/20.500.14106/2554 British National Corpus , XML edition . https://llds.ling-phil.ox.ac.uk/llds/xmlui/handle/20.500.14106/2554","work_id":"06b216ce-ad64-4654-a05b-4aaf8a2b0f03","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1017/s2041536212000013","year":2012,"title":"Annette Capel. 2012. https://doi.org/10.1017/S2041536212000013 Completing the English Vocabulary Profile : C1 and C2 vocabulary . English Profile Journal, 3:e1","work_id":"61e4b0ae-c6ac-4be6-8f36-9fc00f8ad5be","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1145/2939672.2939785","year":2016,"title":"In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp","work_id":"1c2af073-b162-4cd0-a5cd-d8aade23b9b5","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.18653/v1/2020.acl-main.747","year":2020,"title":"Proceedings of the Association for Computational Linguistics (ACL) , pages =","work_id":"bad774d3-20f4-421f-ba75-d1ef99f02a26","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":35,"snapshot_sha256":"fd40aa816827763ed96283ddbc2d24450ab44c51d9a6a4aecf67b036a09fd55b","internal_anchors":7},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}