Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Transliteration
Process of converting text from one script to another based on phonetic correspondence, essential for multilingual systems handling languages with different alphabets such as Arabic or Cyrillic.
Multilingual Models (mBERT, XLM-R)
Transformers pre-trained on large parallel or concatenated corpora in many languages, capable of understanding and generating text in over 100 languages without language-specific training.
Language Anchor
Architecture strategy using a pivot language (typically English) as a common semantic representation space to facilitate information transfer between multiple languages in a dialogue system.
Multilingual Processing Pipeline
Sequential architecture integrating specialized modules (language detection, adapted tokenization, specific models) to manage different stages of natural language processing in a multilingual context.
Multilingual Evaluation (BLEU, chrF)
Evaluation metrics adapted to multilingual contexts measuring the quality of translations or generated responses, such as chrF++ which is more sensitive to morphological characteristics of different languages.
Cultural Diversity Management
Set of techniques enabling a dialogue system to adapt its responses not only linguistically but also culturally, taking into account social norms, references, and specific cultural contexts.
Weakly Supervised Multilingual Language Models
Training approaches combining large amounts of unlabeled data in multiple languages with minimal annotations to build robust systems where supervised data is scarce.
Multilingual Semantic Space Alignment
Technique of projecting vector spaces of different languages into a common semantic space, allowing direct comparisons of meaning between words and phrases from different languages.
Multilingual Textual Normalization
Process of standardizing text including the management of script variations (simplified/traditional), diacritics, capitalization, and language-specific conventions for consistent processing.
Hybrid Multilingual Dialogue System
Architecture combining rule-based approaches (for low-resource languages) and neural models (for well-resourced languages) in a unified framework to optimize overall performance.
Interlingual Transfer
Translation paradigm where source text is first converted into an intermediate language-independent semantic representation, then generated in the target language, facilitating multilingual translations.
Back-Translation
Data augmentation technique where synthetic translations are generated by translating target monolingual data to the source language, then back to the target to create artificial parallel pairs.