We disentangle multilingual sentence embeddings into language-dependent and language-agnostic components, leveraging the latter to improve cross-lingual similarity estimation. Previous studies on this approach have trained disentanglers by combining intra-component constraints, which either align or disalign language-dependent embeddings or language-agnostic embeddings, with inter-component constraints across both embeddings. However, when and how these constraints are effective remains unclear. Our experiments on sentence similarity estimation and machine translation quality estimation revealed that while intra-component constraints and the combination of both constraints are effective for encoder-based multilingual sentence embeddings, inter-component constraints are effective for decoder-based ones. Furthermore, our detailed analysis revealed distinct roles: intra-component constraints improve uniformity within the embedding space, while inter-component constraints enhance cross-lingual alignment between parallel sentences.

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.

We disentangle multilingual sentence embeddings into language-dependent and language-agnostic components, leveraging the latter to improve cross-lingual similarity estimation.Previous studies focused on encoder-based approaches that use only the input sentence; in contrast, this study examines the effectiveness of disentanglement methods across a broader range of sentence embeddings, including decoder-based approaches and those that utilize prompts.Experimental results demonstrate that embedding disentanglement is effective for a wide variety of sentence embeddings.