Robust severity classification and LLM-based information extraction for noisy maritime distress communications. IEEE Access resubmission (Manuscript ID Access-2026-13028). Author: Tomer Atia, HIT - Holon Institute of Technology.
submission/ -> the current, mutually-consistent submission package (start here)
archive/ -> all historical / superseded artifacts and pre-edit backups
| File | Role |
|---|---|
SeaAlert_v10.docx |
Current manuscript. |
SupplementaryMaterials.docx |
Supplementary S1-S4 (S2 recomputed from the gold data). |
SeaAlert_Response_to_Reviewers_v6.docx |
Response to reviewers (updated). |
humen_binary_validation.xlsx |
Gold-source validation data behind Section S2. |
SUBMISSION_README.md |
Package contents, supplementary map, and highlight legend. |
Edits/additions from this revision pass are highlighted in red in the submission docs
(see submission/SUBMISSION_README.md for the legend).
| File | Why archived |
|---|---|
SeaAlert_v9.docx |
Previous manuscript version (superseded by v10). |
SeaAlert_Tasks1_2_Report.docx |
Internal validation/correlation report (Pearson framing). |
SeaAlert_Supplementary_S2_standalone.docx |
Early standalone S2 draft (superseded by the S2 in SupplementaryMaterials.docx). |
SeaAlert_Response_to_Reviewers_v6_root.docx |
Earlier response copy. |
S2_DATA_PROVENANCE_NOTE.md |
Record of the S2 number reconciliation that led to the xlsx-based recompute. |
pre-edit-backups/ |
Pre-edit snapshots captured before each in-place change. |
A two-stage maritime decision-support pipeline on a controlled synthetic benchmark:
- Severity classification (Distress / Urgency / Safety / Routine), comparing a rule-based GMDSS keyword spotter, Logistic Regression, and RoBERTa under simulated VHF / ASR degradation.
- Structured information extraction of seven operational fields from noisy ASR transcripts using GPT-4o-mini.
Section S2 triangulates a second model family (Claude Sonnet) and human adjudication to show that low recovery of alphanumeric identifiers reflects objective ASR information loss rather than an extractor or reference artifact; all S2 numbers trace to submission/humen_binary_validation.xlsx.