This study aims to reveal how accurately Large Language Models (LLMs) can deal with a speaker’s actual utterances and their true feelings behind them in Japanese dialogue. Speakers use not only private thoughts which express one’s true feelings and intentions, but also public statements which convey their intentions while considering the interlocutor’s feelings and social status. While public statements help to maintain interpersonal relationships, they can obscure the speaker’s true intention, potentially leading to misunderstandings. We extended existing Japanese dialogue corpora by annotating public statements and private thoughts responses for each dialogue in the corpora, and then evaluated LLMs’ ability to classify and generate between these two types of expressions. The results of the classification task revealed that the current LLMs do not understand those expressions at all, and that training with our corpus can significantly improve the recognition performance. Furthermore, the results of the generation task demonstrated that generating private thoughts is more difficult than generating public statements, according to both automatic and human evaluations. We release our corpus, which contains 7,964 human-annotated dialogues.

Controlling text difficulty is crucial for adapting texts to individual language learners. Existing controllable text simplification methods condition generation on discrete proficiency-level tokens, treating these levels as independent categories. As a result, they fail to capture the ordinal structure of language proficiency and cannot provide fine-grained control beyond predefined levels. To address this limitation, we propose a continuous difficulty control method that adjusts simplification difficulty by editing the norm of a conditioning embedding derived from the input sentence semantics. The embedding is incorporated into the transformer decoder through attention-based conditioning, enabling continuous control over output difficulty while modeling ordinal relationships across levels. Unlike token-based approaches, our method supports interpolation between proficiency levels and flexible control over edit magnitude. Experiments on controllable text simplification benchmarks demonstrate improved difficulty controllability, simplification quality, and sentence difficulty estimation performance.

We introduce the EhiMeNLP submission, which won the TSAR 2025 Shared Task on Readability-Controlled Text Simplification. Our system employed a two-step strategy of candidate generation and reranking. For candidate generation, we simplified the given text into more readable versions by combining multiple large language models with prompts. Then, for reranking, we selected the best candidate by readability-based filtering and ranking based on semantic similarity to the original text.