Shogo Tsujimoto

Author directory

2026

Patent claims are legal documents with distinctive formats and complex syntactic dependencies. Translating these claims requires not only accurately understanding the complex syntactic structure of the source claims but also ensuring the translation follows the structural conventions of claims in the target language. In this study, we leverage reasoning models to disentangle the complex syntactic structures of source claims. Furthermore, to ensure the translation reflects the structural requirements in the target language, we incorporate FastKASSIM—a syntactic similarity metric—into the reward of reinforcement learning, alongside conventional metrics such as BLEU and COMET. Experimental results confirm that our model, trained on Qwen3-8B, outperforms Qwen3-235B in syntactic similarity while simultaneously achieving comparable translation quality, despite the latter having approximately 30 times more parameters. Automatic and human evaluations show that models trained with rewards combining FastKASSIM and translation quality metrics such as COMET or COMETKiwi achieve strong performance, suggesting that syntactic rewards help improve patent-specific structural conformity while maintaining translation quality.

2025

Recent studies on semantic frame induction have demonstrated that the emergence of pre-trained language models (PLMs) has led to more accurate results.However, most existing studies evaluate the performance using frame resources such as FrameNet, which may not accurately reflect real-world language usage.In this study, we conduct semantic frame induction using the Colossal Clean Crawled Corpus (C4) and assess the applicability of existing frame induction methods to real-world data.Our experimental results demonstrate that existing frame induction methods are effective on real-world data and that frames corresponding to novel concepts can be induced.