To address the challenges posed by professional terminology

short texts

large data volumes

and mixed Chinese–English content in aircraft maintenance manuals

this paper proposes a knowledge extraction and knowledge graph construction method based on multi-head attention and full-token masking. First

we design the CoBiTex-FTM (Contextual bidirectional text encoder with full-token masking) model for named entity recognition

which enhances contextual modeling through multi-head attention and ensures label consistency via a whole-word constraint algorithm tailored for mixed-language scenarios. Second

we construct the BiHAM-FTM (Bidirectional LSTM multi-head attention with full-token masking) model to extract “entity–relation–entity” triples. Finally

an aircraft maintenance knowledge graph system is implemented using Neo4j for structured storage and visual representation of maintenance knowledge. To validate the approach

we build a domain-specific dataset and conduct comparative and ablation experiments. Experimental results show that CoBiTex-FTM achieves an F1 score of 95.16%

while BiHAM-FTM reaches 90.74%

demonstrating superior performance in complex

multilingual

and short-text environments.