Language

• Study low-latency simultaneous translation between Japanese and Korean, which have similar word order
• Use BS-SiMT, which learns from data when to wait before generating a translation
• Show that some delay is necessary to maintain accuracy even when the source and target languages have similar word order
• Next step: develop more advanced translation strategies, including paraphrasing, to reduce latency

• Use large language models to translate descriptions of monuments to poems from the Man'yōshū
• Improve translation by providing readings for difficult personal and place names that frequently occur in the descriptions
• Use a large language model to identify difficult words and manually annotate their readings
• Next step: automatically create and expand reading-information resources for difficult words

• Identify key named entities in text, including proper names and numerical expressions
• Automatically create additional training data by replacing named entities with others of the same type
• Use large language models to generate effective replacement data
• Next step: improve data augmentation by accounting for writing style and context

• Automatically correct grammatical errors made by beginning learners of Japanese
• Analyze accuracy in correcting inflectional-ending errors for na-adjectives (adjectival nouns)
• Find that correction accuracy tends to decrease for words written in hiragana
• Next step: evaluate Japanese grammatical error correction using large language models

• Identify language in social media posts that is associated with mental health difficulties
• Use large language models to extract keywords that may be characteristic of people experiencing mental distress
• Analyze the extracted expressions in a large collection of social media posts
• Next steps: extract longer expressions and infer the meaning of entire posts

• Use large language models to help people revise their writing
• Ask a large language model to identify points for improvement in Japanese research paper abstracts
• Evaluate three types of prompts using nearly 100 abstracts
• Next step: extend the approach to types of writing other than research papers