scribe/research/papers.json

{
  "version": 1,
  "snapshot_date": "2026-07-13",
  "usage_scope": "research_only",
  "promotion_status": "blocked_research_only",
  "license_note": "The license field covers the paper PDF only. Dataset, model, audio, transcript, and source-content rights require separate review before use.",
  "papers": [
    {
      "title": "ViMedCSS: A Vietnamese Medical Code-Switching Speech Dataset & Benchmark",
      "filename": "vimedcss-2026.pdf",
      "url": "https://arxiv.org/pdf/2602.12911v2",
      "research_question": "How should Vietnamese ASR recognize English medical terms embedded in Vietnamese speech?",
      "dataset": "ViMedCSS: 34 hours and 16,576 Vietnamese-English medical code-switching utterances across five medical topics.",
      "method": "Benchmark Vietnamese-specialized and multilingual ASR models, then compare fine-tuning strategies and code-switch-aware metrics.",
      "strongest_result": "Vietnamese-specialized models were stronger on surrounding Vietnamese while multilingual pretraining better recognized English insertions; combining both gave the best overall/code-switched balance.",
      "limitations": "Every utterance contains an English medical term, so prevalence differs from routine outpatient speech; source-audio and dictionary rights require independent review.",
      "status": "Accepted at LREC 2026; arXiv version 2.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "Overall WER can hide clinically important code-switched term failures, so model selection must prioritize hard-split medical-term accuracy.",
      "carepath_experiment": "Evaluate Gipformer and a PhoWhisper/Whisper LoRA candidate on test and hard splits using WER, CS-WER/PIER, medical-term recall, number/unit accuracy, latency, and VRAM.",
      "sha256": "95715571ded3bbcaa08ce23be7e2e0dcfa91df0369e89130b39f34f8f2e8369f"
    },
    {
      "title": "Contrastive Training with LLM-generated Near-Misses for Robust Code-Switching Speech Recognition",
      "filename": "near-misses-2026.pdf",
      "url": "https://arxiv.org/pdf/2606.06985v2",
      "research_question": "Can contrastive learning make ASR distinguish the correct code-switched term from acoustically plausible near misses?",
      "dataset": "CS-FLEURS Mandarin-English and ViMedCSS Vietnamese-English, with real N-best hypotheses expanded near code-switch points.",
      "method": "Detect points of interest, generate LLM near misses, filter them acoustically, phonemically, and textually, then train Whisper-small LoRA with anchored cross-entropy and contrastive ranking.",
      "strongest_result": "On ViMedCSS, WER/PIER improved from 24.72/21.95 with standard LoRA to 21.87/18.74, with over two absolute points of improvement in general and code-switch-aware error rates.",
      "limitations": "Requires genuine N-best hypotheses and added filtering/training complexity; results cover two code-switch pairs and Whisper-small, not the current Gipformer serving path.",
      "status": "Accepted at Interspeech 2026; arXiv version 2.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "A direct ASR adapter may preserve Vietnamese context while learning dangerous English medical near misses better than post-correction alone.",
      "carepath_experiment": "Run near-miss contrastive training only after plain ASR LoRA improves WER and CS-WER by at least 5% relative; use genuine beam outputs and the paper's three filters.",
      "sha256": "40c59e9faaa9279da523d633d076f3075ed01bb9759e6a2ab6f738c4b212e6d7"
    },
    {
      "title": "Failing Forward: Improving Generative Error Correction for ASR with Synthetic Data and Retrieval Augmentation",
      "filename": "darag-2025.pdf",
      "url": "https://aclanthology.org/2025.findings-acl.125.pdf",
      "research_question": "How can generative ASR correction generalize to unseen errors, domains, and named entities?",
      "dataset": "Multiple English in-domain and out-of-domain ASR benchmarks with synthetic error pairs and named-entity datastores.",
      "method": "DARAG augments GEC training with generated ASR errors and retrieves likely named entities into the correction prompt.",
      "strongest_result": "Reported 8%-30% relative WER improvement in-domain and 10%-33% out-of-domain over baselines.",
      "limitations": "The evidence is English and relies on simulated errors; it does not establish Vietnamese medical safety or a gain for single-best Gipformer transcripts.",
      "status": "Peer-reviewed, Findings of ACL 2025.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "Gipformer error pairs plus retrieval from CarePath's canonical term store can reduce unseen Vietnamese medical-term errors without replacing runtime ASR.",
      "carepath_experiment": "Compare Gipformer alone, GEC with real Gipformer pairs and lexical retrieval, and the same GEC data plus clean/phonetic augmentation; keep serving input single-best.",
      "sha256": "706ca285a95e1afefec79126cdb2d35789baddb27277c2be6f3c1e5da22185ae"
    },
    {
      "title": "PiDA: Phonetically-Informed Data Augmentation for Robust Vietnamese Speech Translation",
      "filename": "pida-2026.pdf",
      "url": "https://arxiv.org/pdf/2606.12911v1",
      "research_question": "Which Vietnamese ASR substitutions harm downstream translation, and can phonetic text corruption improve robustness?",
      "dataset": "Vietnamese-English FLEURS ASR and speech-translation examples with categorized substitution errors.",
      "method": "Model substitution causes, build phonetic word embeddings, and fine-tune on a 1:1 mixture of clean text and phonetically corrupted ASR-like text.",
      "strongest_result": "PiDA improved translation of erroneous ASR output by up to 2.04 BLEU over standard fine-tuning while slightly improving clean-text performance.",
      "limitations": "The target task is Vietnamese-English translation rather than ASR correction or clinical notes; BLEU does not measure medication, dosage, or terminology safety.",
      "status": "Accepted at Interspeech 2026; arXiv version 1.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "Text-only Vietnamese phonetic corruptions can improve GEC cheaply before CarePath invests in TTS-generated speech.",
      "carepath_experiment": "Add vowel, tone, consonant, and code-switch-boundary corruptions to real Gipformer error pairs in a 1:1 clean/corrupt mixture and measure hard-split PIER plus clean WER.",
      "sha256": "b893629a67caacf5107e18e8a5304a15a00971303120b3a149afe7caace8031f"
    },
    {
      "title": "VietMed: A Dataset and Benchmark for Automatic Speech Recognition of Vietnamese in the Medical Domain",
      "filename": "vietmed-2024.pdf",
      "url": "https://aclanthology.org/2024.lrec-main.1509.pdf",
      "research_question": "Can Vietnamese medical ASR generalize across accents, speaker roles, conditions, and ICD-10 disease groups?",
      "dataset": "16 hours labeled medical speech, 1,000 hours unlabeled medical speech, and 1,200 hours unlabeled general Vietnamese speech.",
      "method": "Pretrain Vietnamese wav2vec/XLSR models at scale and fine-tune/benchmark them for medical ASR.",
      "strongest_result": "XLSR-53-Viet reduced test WER from 51.8% to 29.6% versus XLSR-53, a relative reduction above 40%.",
      "limitations": "Only 16 labeled medical hours are available, and the recording/task distribution differs from CarePath consultations; dataset rights must be checked separately.",
      "status": "Peer-reviewed, LREC-COLING 2024.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "An untouched VietMed test slice will expose accent and disease-domain regressions hidden by ViMedCSS code-switch evaluation.",
      "carepath_experiment": "Use VietMed only as out-of-domain evaluation for Gipformer, direct ASR LoRA, and the selected correction path; report accent and disease-group slices when metadata permits.",
      "sha256": "f9d31022cedc6679bb088486014ae686a87c460c8126940653efcf8ca171fd93"
    },
    {
      "title": "Improving Vietnamese-English Medical Machine Translation",
      "filename": "medev-2024.pdf",
      "url": "https://aclanthology.org/2024.lrec-main.784.pdf",
      "research_question": "What data and model adaptation improve Vietnamese-English medical machine translation?",
      "dataset": "MedEV, approximately 360,000 Vietnamese-English parallel medical sentence pairs.",
      "method": "Compare commercial translation systems and bilingual/multilingual sequence-to-sequence models, including direction-specific fine-tuning.",
      "strongest_result": "Direction-specific fine-tuning of vinai-translate performed best in both translation directions in the reported experiments.",
      "limitations": "Parallel sentences are not doctor-patient dialogue or dialogue-to-SOAP supervision; translation metrics cannot establish clinical-note factuality.",
      "status": "Peer-reviewed, LREC-COLING 2024.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "MedEV can help validate bilingual terminology alignment without being misused as SOAP training data.",
      "carepath_experiment": "Use MedEV only for Vietnamese-English terminology checks against the canonical CarePath term store and silver-corpus quality reports.",
      "sha256": "5c1ff2bae97f24a957be125d39fe09f0722ae9e37f2d293ccd573f54557bf4d4"
    },
    {
      "title": "SpecialtyScribe: Enhancing SOAP note Scribing for Medical Specialties using LLM's",
      "filename": "specialtyscribe-2025.pdf",
      "url": "https://aclanthology.org/2025.cl4health-1.4.pdf",
      "research_question": "Can a modular, smaller-model pipeline generate specialty-specific SOAP notes competitively?",
      "dataset": "In-house oncology clinician-patient transcripts and specialty notes.",
      "method": "Information extraction, context retrieval for transcript grounding, and a note writer trained for specialty preferences.",
      "strongest_result": "In-house models beat similarly sized open models by more than 12% on ROUGE and matched leading closed models while using under 1% of their parameter count.",
      "limitations": "Evidence is in-house, oncology-specific, ROUGE-heavy, and based on hardware/data unavailable to CarePath; it does not validate general Vietnamese outpatient notes.",
      "status": "Peer-reviewed, CL4Health 2025.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "Fact extraction before note writing can constrain a small SOAP model to transcript-supported content.",
      "carepath_experiment": "Use one Qwen base and one SOAP adapter in two sequential passes: typed fact extraction with source spans, then grounded generation into CarePath's existing schema.",
      "sha256": "fb7bca90a0a196f18a3ffa39cc49b15e83dcdf9bcb14ccd0c4323f93034ab226"
    },
    {
      "title": "Generating medically-accurate summaries of patient-provider dialogue: A multi-stage approach using large language models",
      "filename": "multi-stage-summarization-2023.pdf",
      "url": "https://aclanthology.org/2023.clinicalnlp-1.26.pdf",
      "research_question": "Does decomposing medical dialogue summarization improve clinical accuracy over a single zero-shot prompt?",
      "dataset": "Patient-provider dialogues with reference summaries and medical entity/affirmation annotations.",
      "method": "Identify medical entities and affirmations, dynamically retrieve relevant few-shot examples, then summarize with GPT-3 and medical-correctness metrics.",
      "strongest_result": "Human review and medical-correctness metrics favored the multi-stage method over zero-shot single-prompt summarization.",
      "limitations": "Uses proprietary GPT-3 and English data; the reported evaluation is not a production safety proof and does not address Vietnamese terminology.",
      "status": "Peer-reviewed, ClinicalNLP 2023.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "Explicit facts, negation, and uncertainty should be extracted before SOAP generation to make unsupported statements detectable.",
      "carepath_experiment": "Require every generated critical SOAP fact to map to an extracted fact and exact dialogue span; fail export on unsupported medication, dose, diagnosis, number, or negation.",
      "sha256": "7eefe9035a477fc9990ad8b7b41403314071f43bb24c975e600a616496d33436"
    },
    {
      "title": "An Empirical Study of Clinical Note Generation from Doctor-Patient Encounters",
      "filename": "mts-dialog-2023.pdf",
      "url": "https://aclanthology.org/2023.eacl-main.168.pdf",
      "research_question": "How well do summarization models, augmentation, and guided generation produce clinical-note sections from short medical dialogues?",
      "dataset": "MTS-Dialog: 1,700 short doctor-patient conversations with section headers and corresponding note summaries.",
      "method": "Benchmark language models, data augmentation, guided summarization, automatic metrics, and expert factual evaluation.",
      "strongest_result": "Released a public 1.7K-pair benchmark plus manual fact precision/recall/F1, hallucination, and omission annotations for 400 generated summaries.",
      "limitations": "English, mostly short section-level dialogues rather than full encounters; only 79 examples are SOAP-category notes, and translated examples remain silver data.",
      "status": "Peer-reviewed, EACL 2023.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "MTS-Dialog can seed a small research-only Vietnamese corpus and factual evaluation, but cannot establish Vietnamese clinical readiness.",
      "carepath_experiment": "Translate/adapt training split examples with deterministic teacher settings, retain source provenance, reject fact-misaligned pairs, and keep validation/test splits untouched.",
      "sha256": "aa01b0ccb2615dedcb4c79ea905c0d17b85b8033dc94feab50d0b9153851b691"
    },
    {
      "title": "ACI-BENCH: a Novel Ambient Clinical Intelligence Dataset for Benchmarking Automatic Visit Note Generation",
      "filename": "aci-bench-2023.pdf",
      "url": "https://arxiv.org/pdf/2306.02022v1",
      "research_question": "How should automatic full visit-note generation be benchmarked on realistic doctor-patient conversations?",
      "dataset": "ACI-BENCH: 207 full doctor-patient conversations and notes assembled from existing challenge and synthetic collections.",
      "method": "Release common train/validation/test splits and benchmark extractive and abstractive note-generation approaches.",
      "strongest_result": "Established a public full-dialogue/full-note benchmark and showed that model rankings and quality differ across note-generation approaches and sections.",
      "limitations": "Small, English, and substantially synthetic; it is a benchmark rather than evidence that generated notes are safe for clinical use.",
      "status": "Peer-reviewed in Scientific Data 2023; arXiv version 1 snapshot.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "Longer full-encounter examples complement short MTS-Dialog sections when teaching structure and testing omissions.",
      "carepath_experiment": "Translate/adapt only training examples into research-only Vietnamese silver data; reserve validation/test examples for factual and section-completeness evaluation.",
      "sha256": "1e4468f56b7601fc1f05c67a3ec463dfa37f5438d147e3e42ad70106a4b78d0b"
    },
    {
      "title": "An Investigation of Evaluation Methods in Automatic Medical Note Generation",
      "filename": "medical-note-evaluation-2023.pdf",
      "url": "https://aclanthology.org/2023.findings-acl.161.pdf",
      "research_question": "Which automatic metrics track expert judgments of generated clinical notes?",
      "dataset": "Seven domain-expert-annotated medical note datasets with system/reference facts and critical-error judgments.",
      "method": "Compare graph, domain-weighted, fine-tuned, and ensemble metrics against factual correctness, hallucination, omission, and manual ratings.",
      "strongest_result": "No single metric behaved consistently across all note datasets; an aggregate subset correlated most stably with human judgments.",
      "limitations": "Metric correlations are dataset-dependent and cannot replace clinician review; the study does not cover Vietnamese notes.",
      "status": "Peer-reviewed, Findings of ACL 2023.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "ROUGE alone will reward fluent but unsafe SOAP notes, so export must be gated by factual precision, hallucination, omission, negation, medication, dose, and number checks.",
      "carepath_experiment": "Compare CKey, base Qwen, and the adapter with schema validity and critical-fact gates; report ROUGE only as a diagnostic secondary metric.",
      "sha256": "437e497217060a5b5d055b8391626688710420b3f93df73d9bdaf84c962d9324"
    },
    {
      "title": "From Feedback to Checklists: Grounded Evaluation of AI-Generated Clinical Notes",
      "filename": "clinical-note-checklists-2025.pdf",
      "url": "https://aclanthology.org/2025.emnlp-industry.104.pdf",
      "research_question": "Can real clinician feedback be distilled into interpretable, scalable note-quality checklists?",
      "dataset": "Deidentified feedback and ratings from more than 21,000 encounters in a deployed ambient-scribe system.",
      "method": "Convert feedback into section-aware yes/no checklist items and test LLM evaluators against ratings and quality-degrading perturbations.",
      "strongest_result": "Feedback-derived checklists improved coverage, diversity, predictive power for human ratings, and robustness to controlled note degradations over a baseline checklist.",
      "limitations": "The data and checklist-generation tooling are private, physician preferences are context-dependent, and LLM checklist scores are not clinical validation.",
      "status": "Peer-reviewed, EMNLP Industry Track 2025.",
      "license": "CC BY 4.0 for the paper only.",
      "carepath_hypothesis": "A short Vietnamese checklist grounded in CarePath failure cases can make owner review more consistent and interpretable.",
      "carepath_experiment": "Use deterministic checklist questions for unsupported facts, omissions, medication/dose/number fidelity, negation, section placement, and required clinician review; keep LLM judging informational.",
      "sha256": "3cccf2486bc008b75a798ea7f6977edeaecc3c267cc9f3855876cf8873e7096a"
    }
  ]
}