This study proposes a simple, one-stage approach to coordinate structure extraction using multilingual Large Language Models (LLMs) with Translation between Augmented Natural Languages (TANL) to develop an error detection system for coordinate structure translation. Unlike conventional multi-component methods such as CoRec, our method employs an end-to-end Transformer decoder (LLM) trained via Continual Pre-Traning (CPT) and/or Supervised Fine-Tuning (SFT) on English and Japanese datasets obtained from parsed treebanks that includes coordinate structures. We evaluated the proposed models on 100 English and Japanese patent claims manually annotated with coordinate structure tags. The proposed method using open-weight models such as Llama-3.2-8B or gemma-3-4b-it significantly outperformed GPT-5 and CoRec by approximately 0.02-0.03 in F1 score for the English task. The proposed method using open-weight models such as llama-3-youko-8b and Llama-3-swallow-8B-0.1v significantly outperformed GPT-5 by approximately 0.02-0.05 in F1 score for the Japanese task. In addition, models using both English and Japanese training data significantly outperform those using monolingual training data only.

This paper proposes an LLM-based method for cross-lingual alignment of coordinate structures in Japanese–English and Chinese–English patent claims. The approach has two components. First, we construct task-specific alignment datasets by translating Penn Treebank-3 English sentences into Japanese and Chinese and projecting coordinate-structure annotations using syntactic parse trees, word alignments, and the GHKM algorithm. Second, we formulate alignment as a TANL-based text generation task in which LLMs generate the target sentence with the corresponding span tagged. Although the training data are derived from a general-domain corpus, evaluation is conducted on manually annotated real patent claims. Across four translation directions, the proposed approach consistently outperforms SpanAlign, an encoder-based alignment baseline built on SQuAD-style cross-lingual span prediction and ILP-based alignment. Controlled analyses show that the TANL formulation improves performance independently of supervised fine-tuning and that automatically generated demonstrations achieve performance close to manually constructed examples in the tested 3-shot setting. Supervised fine-tuning with the automatically constructed datasets further improves open-weight LLMs, with the best fine-tuned model in each direction achieving an F1 score above 0.8.