AI Glossary
The complete dictionary of Artificial Intelligence
Cross-Lingual Transfer
Learning technique where knowledge acquired by a model on a source language is transferred to improve performance on a target language, without requiring additional training data in the latter. This approach exploits universal linguistic similarities to generalize learning across different languages.
Zero-Shot Cross-Lingual Transfer
Ability of a model to perform tasks in a target language without any training examples in that language, relying exclusively on knowledge acquired during training on source languages. This method maximizes transfer efficiency by eliminating the need for multilingual annotated data.
Few-Shot Cross-Lingual Transfer
Variant of cross-lingual transfer where the model only requires very few annotated examples in the target language to effectively adapt transferred knowledge. This approach combines transfer efficiency with rapid adaptation to the linguistic specificities of the target language.
Multilingual Embeddings
Dense vector representations designed to encode words or phrases from multiple languages in a common semantic space, enabling cross-lingual comparisons and transfers. These embeddings capture semantic relationships regardless of the original language of the text.
Language-Agnostic Representations
Internal representations of a model that capture semantic concepts without being tied to a specific language, thus facilitating knowledge transfer between languages. These abstract representations allow the model to generalize beyond superficial linguistic particularities.
Cross-Lingual Fine-Tuning
Process of adapting a multilingual pre-trained model to a specific task using data from one or more source languages, then direct application to target languages without retraining. This technique optimizes transfer while minimizing the need for annotated data.
Multilingual Pre-training
Initial training phase where a model learns linguistic representations from vast text corpora in multiple languages simultaneously, establishing the foundation for cross-lingual transfer. This approach creates inherently multilingual models capable of understanding and generating text in different languages.
Cross-Lingual Model Adaptation
Systematic process of adjusting model parameters to optimize performance on specific target languages, while preserving knowledge acquired on source languages. Adaptation aims to reduce the performance gap between source and target languages.
Universal Language Model
Model architecture designed to understand and effectively process a large number of different languages through multilingual attention mechanisms and shared representations. These models aim to create universal linguistic understanding that transcends the barriers of individual languages.
Cross-Lingual Knowledge Distillation
Technique where a large multilingual teacher model transfers its knowledge to a more compact student model, preserving cross-lingual capabilities while reducing computational complexity. This method enables effective deployment of multilingual models on limited resources.
Multilingual Tokenization
Strategy for segmenting text into meaningful units that works consistently across multiple languages, often using shared vocabularies such as multilingual Byte-Pair Encodings. Effective tokenization is crucial for the success of cross-lingual transfer.
Cross-Lingual Alignment
Process of semantically aligning representation spaces between different languages, ensuring that similar concepts occupy similar positions in the shared vector space. Alignment is fundamental to enabling meaningful comparisons and transfers between languages.
Transferability Assessment
Systematic evaluation of a model's ability to effectively transfer its knowledge from a source language to a target language, often measuring performance degradation. This evaluation guides decisions on optimal transfer strategies for specific language pairs.
Cross-Lingual Domain Adaptation
Simultaneous adaptation of a model to both new languages and new topic domains, combining the challenges of linguistic transfer and domain adaptation. This approach is essential for real-world applications requiring multilingual expertise in specialized sectors.
Multilingual Transformer
Architecture based on attention mechanisms specifically designed to efficiently process multiple languages in a unified model, using shared parameters and multilingual embeddings. These transformers form the basis of modern cross-lingual transfer models.