Learning Complementary Action Pairs from Automotive Repair Manuals

ACL ARR 2026 January Submission6409 Authors

Keywords: Complementary action relations, Contrastive representation learning, Generative modeling, Industrial technical documentation

Abstract:

Learning complementary action pairs is a fundamental prerequisite for procedural understanding in automotive repair.
Repair manuals describe mechanical procedures as short, imperative instructions that implicitly encode inverse action relations, such as assembly--disassembly or tighten--loosen, without explicit semantic links.
Recovering these relations is challenging due to highly repetitive documentation, concise phrasing, and a strong dependence on contextual cues.
Although expert-defined heuristics can resolve a subset of complementary action pairs, such rules generalize poorly to linguistic variation and achieve only partial coverage in practice.
We study the problem of learning complementary action pairs from real-world automotive repair manuals.
We construct a dataset of complementary repair step pairs using automatic alignment followed by manual verification, and investigate learning-based formulations based on contrastive representation learning and generative modeling.
Experimental results show that learned models capture complementary action relations beyond surface-level similarity and generalize more robustly than rule-based approaches under realistic industrial constraints.

Paper Type: Long

Research Area: Semantics: Lexical, Sentence-level Semantics, Textual Inference and Other areas

Research Area Keywords: Semantics: Lexical and Sentence-Level, Machine Learning for NLP, NLP Applications

Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models, Data resources

Languages Studied: German

Submission Number: 6409