ASMR: Agentic Schema Generation for Ship Maintenance Report Writing¶

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

Ask a follow-up

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

TL;DR¶

ASMR is a two-agent pipeline that automatically induces structured schemas from historical ship maintenance narratives, enabling real-time writing assistance for personnel. A Field Generator Agent extracts and clusters semantic concepts using an LLM + K-means, and a Structural Optimizer Agent uses tabular Q-learning to prune the overcomplete candidate field space into compact, low-redundancy schemas. Preliminary results on ~500 forms per category demonstrate measurable improvements in coverage, compactness, and redundancy over raw-concept and candidate-schema baselines.

Problem¶

Ship maintenance reports are free-form text, leading to inconsistent structure across personnel, missed information, and poor downstream extractability. Manually designing schemas does not scale across large workforces or evolving operational requirements. No prior system automatically discovers per-report-type schemas from unstructured historical corpora.

Method¶

ASMR consists of two specialized agents operating in sequence:

Field Generator Agent — For each form type $t \in \mathcal{T}$ (identified via ESWBS codes), the agent uses few-shot GPT-4o Mini prompting to extract atomic semantic concepts $\mathcal{C}$ from sampled historical forms. It then embeds concepts and runs K-means across multiple values of $K$ (multi-granularity clustering), progressively increasing $K$ until no meaningful new abstractions emerge. Each cluster is abstracted into a candidate field $\phi \in \Phi$ via a contextual LLM prompt. The union across all granularities yields an overcomplete candidate field space $\Phi$. In parallel, the agent constructs a binary field–form coverage matrix $C(\phi_i, f_j)$ and a pairwise semantic redundancy matrix $M(\phi_i, \phi_j) \in [0,1]$.

Structural Optimizer Agent — Frames schema construction as a stochastic MDP over states $s = \langle\phi, t\rangle$. Actions are $\mathcal{A} = \{\text{ADD}, \text{MERGE}, \text{STOP}\}$. Transition probabilities for ADD are proportional to field occurrence frequency $\text{freq}(\phi_j, t)$; MERGE transitions are proportional to redundancy scores from $M$; both include additive noise to encourage exploration. The reward function is: $$R(s_t, s_{t+1}, a) = \alpha \cdot \Delta\text{Coverage} - \beta \cdot \Delta\text{Size} - \gamma \cdot \Delta\text{Redundancy}$$ where coverage, size, and redundancy are computed from $C$ and $M$ and incrementally updated per action. Optimization uses tabular TD Q-learning (Equation 13). After convergence, an LLM labels each finalized merged field group with a human-readable name and guide description. At inference, a separate FieldValue Extraction Tool maps unseen narratives to schema fields with confidence scores, triggering targeted follow-up questions for low-confidence fields.

Key Contributions¶

Formal definition of the automatic schema generation problem for heterogeneous operational report collections
ASMR framework: LLM-based concept extraction + adaptive multi-granularity K-means + RL-based field selection/merging in a two-agent pipeline
Six dataset-driven schema quality metrics (coverage, support, consistency, informativeness, redundancy, schema size) computed from the coverage matrix $C$ and redundancy matrix $M$
Schema population inference module with per-field confidence scores and human-AI collaborative guidance generation

Results¶

Evaluated on multiple ship maintenance form types (e.g., FUEL OIL TANKS, VOIDS AND COFFERDAMS, STORAGE COMPARTMENTS), each with ~500 historical forms
Three-way comparison: Raw Concepts → Candidate Schema (Field Generator output) → Optimized Schema (ASMR full pipeline)
Raw concepts yield highly fragmented, low-support, high-redundancy schemas; the Field Generator substantially improves coverage; ASMR further improves compactness and reduces redundancy (exact numbers truncated in provided text)
Running times: Field Generator concept extraction ~20 min, clustering ~5 min, field abstraction ~10 min, redundancy/coverage statistics ~35 min total; RL training ~8 sec; inference ~2 sec per form
Qualitative: RL agent correctly merges {Hazard, Safety, Security, System} into a single combined field for the COMPARTMENTS NOT OTHERWISE COVERED category

Limitations¶

Preliminary evaluation; final quantitative table (Table 4) is truncated — complete metric values not fully reported
Evaluated only on naval/ship domain; generalizability to other industries asserted but not empirically demonstrated
Field Generator dominates runtime (~35 min) due to serial LLM calls; does not scale obviously to very large corpora without batching strategies
Tabular Q-learning over a discrete field space may not scale gracefully as $|\Phi|$ grows large (exponential state space)
Coverage and redundancy matrices are precomputed from training forms; schema quality on out-of-distribution form types is not assessed
Human evaluation of schema utility by actual maintenance personnel is absent

Relevance to Harnesses / Meta-Harnesses¶

ASMR is a canonical two-agent harness: a generator agent feeds structured intermediate output (overcomplete field space + precomputed matrices) to an optimizer agent, which then runs a learned policy over that structure — a clean pipeline harness pattern where inter-agent communication is a typed artifact rather than free text. More distinctively, the generated schemas themselves act as a meta-harness for downstream inference: they define the structured interface through which a separate FieldValue Extraction Tool and question-generation module operate, making ASMR a harness that produces a harness configuration. This demonstrates a composable, schema-driven approach to multi-agent coordination that is domain-agnostic and may generalize to any setting where the structure of downstream AI assistance must be discovered from historical data rather than hand-engineered.