{"paper":{"title":"Retrieval-Based Multi-Label Legal Annotation: Extensible, Data-Efficient and Hallucination-Free","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Retrieval in a frozen embedding space assigns multiple legal labels to documents with competitive accuracy, strong data efficiency, and no risk of hallucinating outside the taxonomy.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Jaromir Savelka, Kevin Ashley, Li Zhang","submitted_at":"2026-05-16T02:40:01Z","abstract_excerpt":"Multi-label legal annotation requires assigning multiple labels from large, evolving taxonomies to long, fact-intensive documents, often under limited supervision. Parametric encoders typically require task-specific training and retraining when the label set changes, while prompting generative large language models becomes costly and degrades as the label space grows. We cast legal annotation as retrieval: we embed documents and label descriptions with a frozen retrieval model and predict labels via k-nearest neighbors in the embedding space, enabling updates by re-embedding and re-indexing ra"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across three legal datasets, retrieval achieves competitive accuracy and strong data efficiency; on Eurlex, Qwen-8B retrieval improves Macro-F1 from 40.41 (GPT-5.2, zero-shot) to 49.12 while reducing estimated compute by 20-30 times compared to fine-tuning, and with N=100 training samples nearly doubles Micro-F1 over hierarchical Legal-BERT on ECtHR-A.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That similarity in the frozen retrieval embedding space reliably indicates label applicability for long, fact-intensive legal documents without any task-specific adaptation or fine-tuning of the embedder.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Retrieval with frozen embeddings and k-NN delivers competitive accuracy, high data efficiency, and zero hallucinations on legal multi-label annotation across ECtHR and Eurlex datasets.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Retrieval in a frozen embedding space assigns multiple legal labels to documents with competitive accuracy, strong data efficiency, and no risk of hallucinating outside the taxonomy.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"96e5454bcaab74bab3b9de4b178399406b7b5b0ef35e6dd96d3281f315c27b63"},"source":{"id":"2605.16767","kind":"arxiv","version":1},"verdict":{"id":"227b96a6-abef-4248-820a-6289953f6b99","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T21:36:03.815453Z","strongest_claim":"Across three legal datasets, retrieval achieves competitive accuracy and strong data efficiency; on Eurlex, Qwen-8B retrieval improves Macro-F1 from 40.41 (GPT-5.2, zero-shot) to 49.12 while reducing estimated compute by 20-30 times compared to fine-tuning, and with N=100 training samples nearly doubles Micro-F1 over hierarchical Legal-BERT on ECtHR-A.","one_line_summary":"Retrieval with frozen embeddings and k-NN delivers competitive accuracy, high data efficiency, and zero hallucinations on legal multi-label annotation across ECtHR and Eurlex datasets.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That similarity in the frozen retrieval embedding space reliably indicates label applicability for long, fact-intensive legal documents without any task-specific adaptation or fine-tuning of the embedder.","pith_extraction_headline":"Retrieval in a frozen embedding space assigns multiple legal labels to documents with competitive accuracy, strong data efficiency, and no risk of hallucinating outside the taxonomy."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.16767/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T22:01:19.752116Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T21:40:53.803060Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"claim_evidence","ran_at":"2026-05-19T19:01:56.313657Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.447215Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"ac59bdd6cdaf9b16ed0b83840e62121c909debf5499a70985354a4cbb7d7f088"},"references":{"count":37,"sample":[{"doi":"","year":2017,"title":"K. D. Ashley, Artificial intelligence and legal analytics: new tools for law practice in the digital age, Cambridge University Press, 2017","work_id":"3089d55e-0ac5-456f-b65f-959d4d33bdc6","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"I. Chalkidis, A. Jana, D. Hartung, M. Bommarito, I. Androutsopoulos, D. Katz, N. Aletras, Lexglue: A benchmark dataset for legal language understanding in english, in: Proceedings of the 60th Annual M","work_id":"75699e84-9918-4637-89dc-b7c5b24780e0","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2016,"title":"N. Aletras, D. Tsarapatsanis, D. Preoţiuc-Pietro, V. Lampos, Predicting judicial decisions of the european court of human rights: A natural language processing perspective, PeerJ computer science 2 (2","work_id":"9b115abc-17c2-4e6b-b006-1d1612db283b","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"I. Chalkidis, E. Fergadiotis, P. Malakasiotis, I. Androutsopoulos, Large-scale multi-label text classification on eu legislation, in: Proceedings of the 57th annual meeting of the association for comp","work_id":"176e2bf5-e663-43f5-a87d-b2075a589cf9","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"W.-C. Chang, H.-F. Yu, K. Zhong, Y. Yang, I. S. Dhillon, Taming pretrained transformers for extreme multi-label text classification, in: Proceedings of the 26th ACM SIGKDD international conference on ","work_id":"974911e0-572f-46ef-89c9-17ae79cd2c26","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":37,"snapshot_sha256":"9957652cd2a0410ea7124e55c8423ea97a85b29b4afd76e3c13f7f41a5072342","internal_anchors":8},"formal_canon":{"evidence_count":1,"snapshot_sha256":"f6289cbd5237cf977d30235174ce61b5fdbc4f023ba7600fe38e29f3de68d4d6"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}