Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MetaMT,a MetaLearning Method Leveraging Multiple Domain Data for Low Resource Machine Translation

Rumeng Li, Xun Wang, Hong Yu

TL;DR

The paper tackles low-resource, multi-domain machine translation by learning a domain-invariant word representation and a meta-learning policy that yields fast adaptation to new domains. It introduces MetaMT, which maps words into a shared semantic space via a base embedding $E^G$ and domain-specific transforms $A^i$, while optimizing a dual-parameter system $\theta_0$ (model) and $\theta_1$ (meta) to facilitate cross-domain transfer. Evaluated on seven English–Spanish datasets with a Transformer backbone, MetaMT achieves BLEU gains of about $1$–$2$ points over strong baselines and excels on a very small EHR corpus (3,020 sentences). The approach is language-agnostic, data-efficient, and extensible to other neural models, offering a practical path for robust MT in resource-limited, domain-diverse settings.

Abstract

Manipulating training data leads to robust neural models for MT.

MetaMT,a MetaLearning Method Leveraging Multiple Domain Data for Low Resource Machine Translation

TL;DR

The paper tackles low-resource, multi-domain machine translation by learning a domain-invariant word representation and a meta-learning policy that yields fast adaptation to new domains. It introduces MetaMT, which maps words into a shared semantic space via a base embedding and domain-specific transforms , while optimizing a dual-parameter system (model) and (meta) to facilitate cross-domain transfer. Evaluated on seven English–Spanish datasets with a Transformer backbone, MetaMT achieves BLEU gains of about – points over strong baselines and excels on a very small EHR corpus (3,020 sentences). The approach is language-agnostic, data-efficient, and extensible to other neural models, offering a practical path for robust MT in resource-limited, domain-diverse settings.

Abstract

Manipulating training data leads to robust neural models for MT.

Paper Structure

This paper contains 21 sections, 2 equations, 2 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. The Encoder-Decoder Model and NMT
  4. Domain adaptation for NMT
  5. Meta Learning
  6. Method
  7. Domain-Invariant Word Representation
  8. Encoding-Decoding
  9. Learning Policy
  10. Experiments & Analysis
  11. Data
  12. Implementations
  13. Baselines
  14. Transformer
  15. Fine Tuning
  16. ...and 6 more sections

Figures (2)

  • Figure 1: A graphical illustration of the parameter updating procedures in meta learning and fine tuning.
  • Figure 2: A graphical illustration of the proposed method. In training phase, source and target words are firstly projected to domain-invariant representational spaces and then are encoded/decoded. Parameters in the model are updated alternatively during training.