Research on teaching human morality to AI has evolved to enhance alignment with human commonsense based on descriptive ethics. However, Language Models (LMs) based on human commonsense are limited in comprehending various moral perspectives and making moral judgments in unusual situations. To tackle these challenges, we propose Moral Instruction Fine-Tuning (MIST) to align LMs with diverse ethical principles. We demonstrate that MIST enables diverse ethical reasoning and significantly enhances zero-shot generalization in unseen moral tasks.

Word embeddings have demonstrated high performance in Natural Language Processing (NLP) tasks, but they face challenges in handling Out-Of-Vocabulary (OOV) words. OOV words that are not included in the pre-trained vocabulary, leading to degraded model performance. To deal with this problem, we combine word formation with contextual information. In this study, we propose a novel approach using Large Language Models (LLMs) to generate high-quality representations for OOV words. Our method uses LLMs to generate contextually appropriate replacement words for OOV words in a sentence. And make positive sample pairs and train the word representations of these samples to be closer to each other. This approach effectively combines morphological features and contextual information. It also has the advantage of maintaining context-agnostic performance for downstream tasks. This approach achieves high performance across diverse intrinsic and extrinsic tasks. This research is expected to improve performance on various NLP tasks including OOV, such as typos and neologisms.