Skip to content

ASMR: Agentic Schema Generation for Ship Maintenance Report Writing

🕒 Published (v1): 2026-07-09 07:25 UTC · Source: Arxiv · link

Why this paper was selected

Agentic schema discovery from unstructured reports; generalizable zero-shot schema generation approach

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

ASMR is a two-agent framework that automatically discovers structured schemas from collections of unstructured ship maintenance reports, replacing ad-hoc free-text with guided, field-level information requirements. A Field Generator Agent extracts semantic concepts and clusters them into candidate fields; a Structural Optimizer Agent uses tabular Q-learning to prune and merge those fields into compact, low-redundancy schemas. Preliminary results on ~500-form datasets show the pipeline produces higher-coverage, lower-redundancy schemas than either raw concepts or unoptimized candidate fields.

Problem

Ship maintenance personnel write free-form narrative reports across heterogeneous form categories (fuel oil tanks, void spaces, storage compartments, etc.), resulting in inconsistent structure, missing critical information, and poor interoperability across units. There is no principled method to automatically infer what structured fields a given form category should contain from historical examples alone, without manual schema design by domain experts.

Method

ASMR is a sequential two-agent pipeline:

Field Generator Agent (LLM-based): 1. Groups historical forms by ESWBS category codes. 2. Uses few-shot, instruction-guided prompting with GPT-4o Mini to extract atomic semantic concepts (e.g., corrosion, water intrusion, tank top) from narrative text. 3. Embeds concepts and runs K-means at multiple values of \(K\) (adaptive multi-granularity clustering) to form coherent concept groups. 4. Abstracts each cluster into a named candidate field (e.g., \(\{\text{Ripped, Cracked, Loose}\} \to \text{Condition}\)), building an overcomplete candidate field space \(\Phi\). 5. Precomputes a binary field–form coverage matrix \(C(\phi_i, f_j) \in \{0,1\}\) and a field–field semantic redundancy matrix \(M(\phi_i, \phi_j) \in [0,1]\) using LLM-assisted semantic comparison.

Structural Optimizer Agent (RL-based): Models schema construction as a stochastic MDP: - State: \(s = \langle \phi, t \rangle\) (current candidate field, form type) - Actions: \(\mathcal{A} = \{\text{ADD, MERGE, STOP}\}\) - Transition: ADD probability \(\propto \text{freq}(\phi_j, t) + \text{noise}\); MERGE probability \(\propto M(\phi_j, \phi_k, t) + \text{noise}\) - Reward: \(R(s_t, s_{t+1}, a) = \alpha \cdot \Delta\text{Coverage} - \beta \cdot \Delta\text{Size} - \gamma \cdot \Delta\text{Redundancy}\)

Optimization uses tabular TD Q-learning: $\(Q(s_t, a_t) \leftarrow Q(s_t, a_t) + \alpha\left[r_t + \gamma \max_{a'} Q(s_{t+1}, a') - Q(s_t, a_t)\right]\)$

After convergence, an LLM assigns human-readable labels and guide descriptions to finalized field groups. At inference, a Field-Value Extraction Tool populates schema fields from unseen narratives and returns per-field confidence scores to drive human-AI collaborative review.

Key Contributions

  • Formal definition of the automatic schema generation problem for operational/maintenance reporting: given \(\mathcal{H}\) and form types \(\mathcal{T}\), discover optimal \(S_{t_j}\) for each \(t_j\).
  • ASMR framework combining LLM-based semantic concept extraction, adaptive multi-granularity K-means clustering, and RL-based field selection/merging/pruning.
  • Six dataset-driven schema quality metrics: coverage, support, consistency, informativeness (lift-based), redundancy, and schema size — enabling quantitative comparison without human annotation.
  • Schema population inference module: field-value extraction with confidence scores feeding a human-AI collaborative reporting workflow.
  • Demonstrated downstream benefits including predictive maintenance support, anomaly detection readiness, and cross-unit interoperability.

Results

Results compare three schema representations: Raw Concepts → Candidate Schema (Field Generator) → Optimized Schema (ASMR):

  • Raw concepts produce highly fragmented schemas with low statistical support, limited coverage, and weak consistency.
  • Candidate schemas substantially improve coverage over raw concepts.
  • ASMR optimized schemas achieve the best trade-off: highest coverage and informativeness with lowest redundancy and smallest schema size (text truncated before final table, but directional conclusions are stated).
  • Qualitative example: for Compartments Not Otherwise Covered, the optimizer merges \(\{\text{Hazard} + \text{Safety} + \text{Security} + \text{System}\}\) into a single consolidated field.
  • Inference timing: ~20 min concept extraction, ~5 min clustering, ~10 min field abstraction, ~35 min total Field Generator; Structural Optimizer RL training ~8 sec, schema finalization ~2 sec, inference ~2 sec per form.
  • Dataset: ~500 historical forms per category; evaluated on form types including FUEL OIL TANKS, STORAGE COMPARTMENTS, VOIDS AND COFFERDAMS, WASTE TANKS, LIVING QUARTERS COMPARTMENTS.
  • No external baselines (e.g., PALIMPZEST, ELEET, DocETL, TabAgent) are directly benchmarked — the paper notes these address complementary problems, not the same schema-generation task.

Limitations

  • Evaluation is preliminary (workshop paper); full ablations and statistical significance tests are absent.
  • Final metric table (Table 4) is truncated in the provided text; specific numbers for ASMR vs. baselines cannot be verified.
  • No human expert evaluation of schema quality; all metrics are dataset-driven and may not capture operational utility.
  • Tabular Q-learning scales poorly as the candidate field space \(|\Phi|\) grows; deep RL or LLM-guided policy search not explored.
  • Expensive LLM calls dominate cost (~35 min), limiting scalability to very large form corpora without batching/caching strategies.
  • Evaluated only on ship maintenance forms; generalization to healthcare, aviation, or logistics is claimed but not demonstrated.
  • Hyperparameters (\(\alpha\), \(\beta\), \(\gamma\), clustering granularity, \(K\) range) selected by grid search — sensitivity analysis is not reported.

Relevance to Agentic AI / LLM Agents

ASMR is a concrete instantiation of the tool-augmented, multi-agent pipeline paradigm: one LLM agent handles semantic extraction and delegates structured data (the coverage/redundancy matrices) to a downstream RL optimization agent, cleanly separating expensive reasoning from lightweight sequential decision-making. The schema finalization and field-value extraction stages further demonstrate how LLM agents can serve as both schema discoverers and structured-data extractors in the same system. For researchers tracking agentic AI, this paper offers a worked example of embedding classical RL (tabular Q-learning) inside an agentic loop to enforce hard structural constraints (compactness, non-redundancy) that pure LLM generation cannot reliably satisfy. The human-AI collaborative reporting workflow — where agent-generated schemas drive real-time clarification questions — also advances the line of work on agents that actively scaffold human information production rather than merely consuming or retrieving it.