Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Retrieving Examples from Memory for Retrieval Augmented Neural Machine Translation: A Systematic Comparison

Maxime Bouthors, Josep Crego, Francois Yvon

TL;DR

The effect of varying retrieval methods for several translation architectures, based on a standard autoregressive model, an edit-based model, and a large language model with in-context learning are studied.

Abstract

Retrieval-Augmented Neural Machine Translation (RAMT) architectures retrieve examples from memory to guide the generation process. While most works in this trend explore new ways to exploit the retrieved examples, the upstream retrieval step is mostly unexplored. In this paper, we study the effect of varying retrieval methods for several translation architectures, to better understand the interplay between these two processes. We conduct experiments in two language pairs in a multi-domain setting and consider several downstream architectures based on a standard autoregressive model, an edit-based model, and a large language model with in-context learning. Our experiments show that the choice of the retrieval technique impacts the translation scores, with variance across architectures. We also discuss the effects of increasing the number and diversity of examples, which are mostly positive across the board.

Retrieving Examples from Memory for Retrieval Augmented Neural Machine Translation: A Systematic Comparison

TL;DR

The effect of varying retrieval methods for several translation architectures, based on a standard autoregressive model, an edit-based model, and a large language model with in-context learning are studied.

Abstract

Retrieval-Augmented Neural Machine Translation (RAMT) architectures retrieve examples from memory to guide the generation process. While most works in this trend explore new ways to exploit the retrieved examples, the upstream retrieval step is mostly unexplored. In this paper, we study the effect of varying retrieval methods for several translation architectures, to better understand the interplay between these two processes. We conduct experiments in two language pairs in a multi-domain setting and consider several downstream architectures based on a standard autoregressive model, an edit-based model, and a large language model with in-context learning. Our experiments show that the choice of the retrieval technique impacts the translation scores, with variance across architectures. We also discuss the effects of increasing the number and diversity of examples, which are mostly positive across the board.
Paper Structure (42 sections, 8 equations, 2 figures, 15 tables)

This paper contains 42 sections, 8 equations, 2 figures, 15 tables.

Table of Contents

  1. Introduction
  2. Retrieval in NMT architectures
  3. Retrieval Pipeline
  4. Measuring Retrieval Quality
  5. Smoothed Longest Common Subsequence
  6. Controlling Diversity: Contrastive Retrieval
  7. Integrating TMs in Translation
  8. Data and Metrics
  9. Data
  10. Metrics
  11. Implementation and Parameters
  12. Computational issues
  13. Experiments
  14. Comparing Retrieval Techniques
  15. Interactions Between Retrieval and Translation
  16. ...and 27 more sections

Figures (2)

  • Figure 1: High-level overview of the retrieval pipeline in fuzzy-matching.
  • Figure 2: Illustration of the variability across some retrieval settings and their respective coverage for a source sentence from EMEA. For each setting, we represent the source-side $3$ best-ranked sentences.