AI 용어집
인공지능 완전 사전
Cross-Lingual Zero-Shot Learning
Learning paradigm where a model performs tasks in a target language without any specific training examples, by transferring knowledge from one or more source languages.
Shared Multilingual Embeddings
Vector representations of linguistic units (words, phrases) in a common semantic space where equivalents from different languages are positioned close spatially.
Target Language Projection
Technique for mapping vector representations from a source language to the vector space of a target language without requiring direct parallel data.
Vector Space Alignment
Mathematical process of geometrically aligning the embedding spaces of different languages to enable the transfer of semantic knowledge between them.
Cross-Domain Adaptation
Ability of a model to generalize its performance from a specific domain in a source language to a different domain in a target language without training data.
Multilingual Transformer Models
Neural architectures pre-trained simultaneously on multiple languages, capable of performing zero-shot transfers between language pairs not seen during training.
Zero-Shot Paraphrasing
Method where the model generates descriptions or paraphrases in the source language to solve complex tasks in the target language without direct examples.
Zero-Shot Text Classification
Task of categorizing text in a target language using only class descriptions in a source language, without labeled target data.
Zero-Shot Evaluation
Specific measurement protocols to evaluate model performance on languages or tasks not seen during training, quantifying transfer capability.
Cross-Lingual Generalization
A model's ability to maintain its performance when applied to languages structurally or typologically different from those used for training.
Few-Shot Fine-Tuning
Adaptation technique using a minimal number of examples in the target language to significantly improve zero-shot performance while preserving transfer efficiency.
Zero-Shot Translation
The ability of multilingual models to translate between language pairs not directly trained together, using pivot languages or shared representations.
Multilingual Prompt Engineering
Optimized design of instructions and examples in a source language to effectively guide task execution in unseen target languages.