Cross-Lingual Embedding Mapper

Build a module that aligns two pretrained embedding spaces using bilingual seed pairs, translates source-language tokens into a target language, and reports top-k accuracy on held-out pairs. Tests use small in-memory embeddings (under 100 tokens) and do not require external downloads.

Capabilities

Train mapping from seed dictionary

• Training with at least ten bilingual pairs learns a linear transform; pairs missing from either vocabulary are skipped, but training still succeeds when five or more usable pairs remain and enables translation/accuracy calls, while translating before fitting raises a descriptive error. @test

Translate source tokens

• Translating ["dog", "unknown"] returns up to topn ordered target candidates per token, preserving input order and yielding an empty list for out-of-vocabulary items. @test

Evaluate held-out dictionary

• Given held-out bilingual pairs and a k value, computing accuracy returns the share of evaluable sources whose gold target appears within the top k translations; pairs with missing vocabulary entries are ignored, and no evaluable pairs returns 0.0. @test

Implementation

@generates

API

from typing import Iterable, List, Tuple, Dict, Any

class CrossLingualMapper:
    def __init__(self, source_vectors: Any, target_vectors: Any, seed_pairs: Iterable[Tuple[str, str]]): ...
    def fit(self, report_skipped: bool = True) -> None: ...
    def translate(self, source_words: Iterable[str], topn: int = 3) -> Dict[str, List[Tuple[str, float]]]: ...
    def accuracy(self, evaluation_pairs: Iterable[Tuple[str, str]], k: int = 1) -> float: ...

Dependencies { .dependencies }

gensim { .dependency }

Provides cross-lingual mapping utilities for aligning embedding spaces and translating vectors.