CMU 11-731(MT&Seq2Seq) Advanced Topics: Multi-lingual Models

11731, Machine Translation, course notes

Pivot Translation

  • Result pivoting

    • use existing source-pivot and pivot-target systems (F is source language, G is pivoting language, E is target language)

      $$ \widehat G\;=\underset G{\;argmax}\;P(G\vert F) $$
      $$ \widehat E\;=\underset E{\;argmax}\;P(E\vert\widehat G) $$

  • Data pivoting

    • pseudoparallel data

  • Model pivoting

    • triangulation

Multi-lingual Training

  • Multi-task Learning Approaches

    • multiple sources to one target
    • one source to mutiple targets
    • many sources to many targets
      • zero shot learning: testing of models on examples that do not exist in their training data

  • Transfer Approaches

  • Ensembling Approaches
Share