Transliteration Model

The core architecture typically employs an encoder-decoder framework with recurrent or transformer layers. The encoder processes the source-script character sequence into a fixed-length context vector, while the decoder autoregressively generates the target-script character sequence. Modern implementations favor transformer-based models for their ability to capture long-range dependencies and parallelize training.

• Input: "Mikhail" (Cyrillic: Михаил)
• Output: "Mikhail" (Latin script)
• Key mechanism: Attention layers align source characters with target characters

Transliteration models operate on phonetic principles, not semantic translation. The model learns to map graphemes (written characters) to phonemes (sound units) and then to target-script graphemes. This differs fundamentally from translation, which maps meaning.

• Phonetic approach: Preserves pronunciation across scripts
• Orthographic approach: Direct character-to-character mapping
• Hybrid systems: Combine both for languages with irregular orthographies like English

In cross-lingual entity linking, transliteration models generate candidate strings that are used to query a target-language knowledge base. When a mention like "Москва" appears in Russian text, the model produces "Moskva," enabling retrieval of the canonical entity identifier for Moscow from Wikidata.

• Pipeline role: Bridges the script gap before candidate retrieval
• Integration: Feeds into dense passage retrieval or surface form dictionaries
• Critical for: Low-resource languages lacking direct cross-lingual mappings

Attention mechanisms are critical for handling variable-length character sequences and non-linear alignments. The model learns to focus on specific source characters when generating each target character, accommodating scripts where character order or grouping differs.

• Bahdanau attention: Additive attention for soft alignment
• Luong attention: Multiplicative attention for faster computation
• Multi-head self-attention: Captures multiple alignment patterns simultaneously, essential for scripts with complex ligatures like Devanagari

Models are trained on parallel corpora of name pairs in source and target scripts, often derived from Wikipedia interlanguage links or governmental gazetteers. Evaluation uses string similarity metrics rather than semantic accuracy.

• Primary metric: Phoneme Error Rate (PER) — measures sound-level accuracy
• Secondary metrics: BLEU, Character Error Rate (CER), and Top-1 Accuracy in entity retrieval
• Data augmentation: Back-transliteration and rule-based generation expand limited training sets

A single source name can have multiple valid transliterations due to competing romanization standards. The model must be conditioned on the target scheme or output a ranked list of candidates.

• ISO 9 vs. BGN/PCGN for Cyrillic
• Hepburn vs. Kunrei-shiki for Japanese
• Solution: Add a control token specifying the target standard, or use multi-task learning to output all valid forms simultaneously