{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:ERBMSKCVHF5ETVX2DJW5NHOQD3","short_pith_number":"pith:ERBMSKCV","schema_version":"1.0","canonical_sha256":"2442c92855397a49d6fa1a6dd69dd01ed208fffb1703e0dd7cb1e5486fb175bb","source":{"kind":"arxiv","id":"2605.13436","version":1},"attestation_state":"computed","paper":{"title":"Pretraining Language Models with Subword Regularization: An Empirical Study of BPE Dropout in Low-Resource NLP","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Stochastic tokenization during both pretraining and fine-tuning yields the best results in low-resource NLP tasks.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Marcel Dunaiski, Ruan Visser, Trienko Grobler","submitted_at":"2026-05-13T12:31:04Z","abstract_excerpt":"Subword regularization methods such as BPE dropout are typically applied only during fine-tuning, while pretraining is usually done with deterministic tokenization. This creates a potential segmentation mismatch between pretraining and fine-tuning. We investigate whether applying BPE dropout during pretraining improves downstream performance in low-resource NLP. We train monolingual and bilingual BERT models on downsampled subsets of English, German, French, Spanish, Kiswahili, and isiXhosa, and evaluate them on XNLI, PAWS-X, PAN-X, and MasakhaNER 2.0. Across tasks, the best results are typica"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.13436","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-05-13T12:31:04Z","cross_cats_sorted":["cs.LG"],"title_canon_sha256":"fe3fc4f0fe3e5638c42b357229d5dddc318e3b413ba0ca2cab87bff14dadd35a","abstract_canon_sha256":"95cf7dcc087fa1067e2a6e12016390653396469c8ef71be08c5541b0a847e0e6"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T02:44:47.101528Z","signature_b64":"JD4J6P/l1VAlJ8JdqQ9o3dr6/hWGrRFHDIHRijlT/JHPhmMHyyzTfHeAbkTpZOi55mb4VR8wwC/HqNAjN+26CA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"2442c92855397a49d6fa1a6dd69dd01ed208fffb1703e0dd7cb1e5486fb175bb","last_reissued_at":"2026-05-18T02:44:47.101101Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T02:44:47.101101Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Pretraining Language Models with Subword Regularization: An Empirical Study of BPE Dropout in Low-Resource NLP","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Stochastic tokenization during both pretraining and fine-tuning yields the best results in low-resource NLP tasks.","cross_cats":["cs.LG"],"primary_cat":"cs.CL","authors_text":"Marcel Dunaiski, Ruan Visser, Trienko Grobler","submitted_at":"2026-05-13T12:31:04Z","abstract_excerpt":"Subword regularization methods such as BPE dropout are typically applied only during fine-tuning, while pretraining is usually done with deterministic tokenization. This creates a potential segmentation mismatch between pretraining and fine-tuning. We investigate whether applying BPE dropout during pretraining improves downstream performance in low-resource NLP. We train monolingual and bilingual BERT models on downsampled subsets of English, German, French, Spanish, Kiswahili, and isiXhosa, and evaluate them on XNLI, PAWS-X, PAN-X, and MasakhaNER 2.0. Across tasks, the best results are typica"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across tasks, the best results are typically obtained when stochastic tokenization is applied during both pretraining and fine-tuning, whereas applying BPE dropout only during fine-tuning can underperform deterministic tokenization in smaller-data settings.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the downsampled subsets of high-resource languages and the chosen evaluation tasks sufficiently represent truly low-resource scenarios, and that the modest morphological alignment gains explain the performance benefits.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Stochastic tokenization with BPE dropout during both pretraining and fine-tuning outperforms deterministic tokenization or fine-tuning-only dropout on low-resource NLP tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Stochastic tokenization during both pretraining and fine-tuning yields the best results in low-resource NLP tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c79f1748cd71246fa7fa1d7261e3aa5ce914459802ef982c38de64bce6e8276d"},"source":{"id":"2605.13436","kind":"arxiv","version":1},"verdict":{"id":"aab51e3e-c161-4e4d-8984-7eb1b006def7","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:25:56.908331Z","strongest_claim":"Across tasks, the best results are typically obtained when stochastic tokenization is applied during both pretraining and fine-tuning, whereas applying BPE dropout only during fine-tuning can underperform deterministic tokenization in smaller-data settings.","one_line_summary":"Stochastic tokenization with BPE dropout during both pretraining and fine-tuning outperforms deterministic tokenization or fine-tuning-only dropout on low-resource NLP tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the downsampled subsets of high-resource languages and the chosen evaluation tasks sufficiently represent truly low-resource scenarios, and that the modest morphological alignment gains explain the performance benefits.","pith_extraction_headline":"Stochastic tokenization during both pretraining and fine-tuning yields the best results in low-resource NLP tasks."},"references":{"count":34,"sample":[{"doi":"","year":2022,"title":"[Adelaniet al., 2022 ] David Ifeoluwa Adelani, Graham Neubig, Sebastian Ruder, Shruti Rijhwani, Michael Beuk- man, Chester Palen-Michel, Constantine Lignos, Je- sujoba O. Alabi, Shamsuddeen H. Muhamma","work_id":"67c4ec17-bbd8-4240-bd97-3ed5e3b53551","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"[Arnett and Bergen, 2025] Catherine Arnett and Ben- jamin K","work_id":"04ee1166-7fe1-4f96-9b36-381954561a32","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Morphynet: A large multilin- gual database of derivational and inflectional morphol- ogy","work_id":"f48a1b4e-0a78-4d41-889d-14a5f4d316dc","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"BPE-knockout: Pruning pre-existing BPE tokenisers with backwards-compatible morpho- logical semi-supervision","work_id":"12aa29ab-7de1-4364-a9c4-963a01b0bb96","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"[Cognettaet al., 2024 ] Marco Cognetta, Vil ´em Zouhar, and Naoaki Okazaki","work_id":"a9e124dd-e469-477d-94f1-f97b809339de","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":34,"snapshot_sha256":"7a67ea21a534b828b28cca2517bca675cafe6d48ba93e209a7f06cd6148f3bbc","internal_anchors":1},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.13436","created_at":"2026-05-18T02:44:47.101168+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.13436v1","created_at":"2026-05-18T02:44:47.101168+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13436","created_at":"2026-05-18T02:44:47.101168+00:00"},{"alias_kind":"pith_short_12","alias_value":"ERBMSKCVHF5E","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"ERBMSKCVHF5ETVX2","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"ERBMSKCV","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/ERBMSKCVHF5ETVX2DJW5NHOQD3","json":"https://pith.science/pith/ERBMSKCVHF5ETVX2DJW5NHOQD3.json","graph_json":"https://pith.science/api/pith-number/ERBMSKCVHF5ETVX2DJW5NHOQD3/graph.json","events_json":"https://pith.science/api/pith-number/ERBMSKCVHF5ETVX2DJW5NHOQD3/events.json","paper":"https://pith.science/paper/ERBMSKCV"},"agent_actions":{"view_html":"https://pith.science/pith/ERBMSKCVHF5ETVX2DJW5NHOQD3","download_json":"https://pith.science/pith/ERBMSKCVHF5ETVX2DJW5NHOQD3.json","view_paper":"https://pith.science/paper/ERBMSKCV","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.13436&json=true","fetch_graph":"https://pith.science/api/pith-number/ERBMSKCVHF5ETVX2DJW5NHOQD3/graph.json","fetch_events":"https://pith.science/api/pith-number/ERBMSKCVHF5ETVX2DJW5NHOQD3/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/ERBMSKCVHF5ETVX2DJW5NHOQD3/action/timestamp_anchor","attest_storage":"https://pith.science/pith/ERBMSKCVHF5ETVX2DJW5NHOQD3/action/storage_attestation","attest_author":"https://pith.science/pith/ERBMSKCVHF5ETVX2DJW5NHOQD3/action/author_attestation","sign_citation":"https://pith.science/pith/ERBMSKCVHF5ETVX2DJW5NHOQD3/action/citation_signature","submit_replication":"https://pith.science/pith/ERBMSKCVHF5ETVX2DJW5NHOQD3/action/replication_record"}},"created_at":"2026-05-18T02:44:47.101168+00:00","updated_at":"2026-05-18T02:44:47.101168+00:00"}