Denoising Neural Machine Translation Training with Trusted Data and Online Data Selection

TL;DR

Proposes a denoising NMT training method using trusted data and online data selection, achieving significant BLEU improvements on noisy datasets.

cs.CL 🔴 Advanced 2018-09-01 41 views
Wei Wang Taro Watanabe Macduff Hughes Tetsuji Nakagawa Ciprian Chelba
neural machine translation denoising online data selection trusted data machine learning

Key Findings

Methodology

The method combines small-scale trusted data with dynamic online data selection. Key steps include: 1) fine-tuning an initial model on trusted data; 2) defining a noise scoring function; 3) dynamically selecting low-noise batches during training.

Key Results

  • On the Paracrawl dataset, BLEU improved by 3.5 (from 28.2 to 31.7), outperforming static filtering methods.
  • On WMT 2014, the method showed cross-domain improvements across test sets (e.g., newstest2014 and patent), with an average BLEU increase of 2.1.
  • Ablation studies revealed that the noise scoring function and dynamic selection contributed 70% and 20% of the total gains, respectively.

Significance

This study addresses NMT's sensitivity to noisy data, offering a novel approach to efficiently utilize large-scale low-quality datasets. The method enhances translation quality and has broad applicability to other noise-prone machine learning tasks.

Technical Contribution

Introduces a noise scoring function based on trusted data and integrates dynamic data selection into denoising training. Unlike static filtering, this approach dynamically adjusts data selection during training, improving model robustness.

Novelty

First to apply dynamic scheduling from domain adaptation to denoising, introducing a model-comparison-based noise definition that handles partially noisy data effectively.

Limitations

  • The method relies on small-scale trusted data; low-quality trusted data may reduce effectiveness.
  • The computational cost of online data selection may be high for extremely large datasets.
  • Generalization to different language pairs remains underexplored.

Future Work

Future work could explore more efficient noise scoring methods, investigate denoising without trusted data, and validate the approach on more language pairs.

AI Executive Summary

Neural Machine Translation (NMT) often struggles with noisy training data, especially from large-scale web crawls. While domain data selection has been well-studied, denoising remains underexplored. This study proposes a novel denoising method combining trusted data and online data selection, significantly improving NMT performance on noisy datasets.

The approach involves three key steps: 1) fine-tuning an initial model on small-scale trusted data to create a denoised model; 2) defining a noise scoring function based on model comparisons; 3) dynamically selecting low-noise batches during training. Experiments on Paracrawl and WMT 2014 datasets demonstrate substantial BLEU improvements, with gains of 3.5 and 2.1 points, respectively.

Despite relying on trusted data and incurring computational overhead, this method's innovation and effectiveness provide a valuable reference for future denoising research. Future directions include optimizing noise scoring and exploring denoising in scenarios without trusted data.

Deep Analysis

Background

NMT has become the dominant paradigm due to its superior translation quality. However, it is highly sensitive to noisy training data, which is common in large-scale web-crawled datasets. Previous research has focused on domain data selection, such as cross-entropy difference methods, but denoising remains underexplored.

Core Problem

The challenge is to effectively remove noise from large-scale datasets while retaining useful training data. Static filtering methods fail to adapt dynamically to training progress and often discard partially useful samples.

Innovation

Key innovations include: 1) a noise scoring function based on trusted data and model comparison; 2) dynamic data selection integrated into training to progressively reduce noise; 3) a lightweight implementation to minimize computational overhead.

Methodology

  • �� Fine-tune an initial model on trusted data to create a denoised model.
  • �� Define a noise scoring function based on the difference between noisy and denoised model predictions.
  • �� Dynamically select low-noise batches during training.
  • �� Use an exponential decay strategy to gradually reduce the selection ratio, ensuring higher-quality data over time.

Experiments

Experiments were conducted on Paracrawl (300M pairs) and WMT 2014 (36M pairs). Trusted data was sourced from newstest 2010-2011, with newstest 2014 and patent used as test sets. The design included baseline comparisons, ablation studies, and cross-domain evaluations.

Results

On Paracrawl, BLEU improved from 28.2 to 31.7. On WMT 2014, average BLEU gains of 2.1 were observed across test sets. Ablation studies showed the noise scoring function contributed 70% of the improvements.

Applications

The method is applicable to scenarios requiring large-scale noisy data processing, such as low-resource language translation, cross-domain model training, and general machine learning data cleaning.

Limitations & Outlook

The method depends on trusted data, incurs computational overhead, and requires further validation on diverse language pairs. Future work could optimize implementation and explore denoising without trusted data.

Plain Language Accessible to non-experts

Imagine you're sifting through a pile of sand to find gold. Traditional methods might discard anything that looks like dirt, but this risks losing small gold particles. This method acts like a smart sieve that adjusts its mesh size over time, gradually filtering out dirt while keeping valuable gold. By the end, you're left with a cleaner, more valuable pile of gold without wasting resources.

ELI14 Explained like you're 14

Think of it like a video game where you have to complete quests, but some quests are fake and give no rewards. This method is like a smart assistant that helps you pick the real quests worth doing. After each quest, it learns and gets better at spotting the good ones. By the end, you finish faster and get all the rewards!

Glossary

Trusted Data

A small, high-quality dataset used to guide denoising. It helps fine-tune the initial model.

Used to define the noise scoring function and improve model robustness.

Noise Scoring Function

A function that quantifies the noise level of training data based on model comparisons.

Used for dynamic data selection during training.

Dynamic Data Selection

A strategy to dynamically adjust training batches based on noise levels.

Ensures progressively cleaner data during training.

BLEU Score

A metric for evaluating translation quality; higher scores indicate better accuracy.

Used to measure model performance on test sets.

Paracrawl Dataset

A large-scale web-crawled parallel dataset with significant noise.

Used to test the method's effectiveness on noisy data.

Open Questions Unanswered questions from this research

  • 1 How can denoising be achieved without trusted data?
  • 2 What is the applicability of the noise scoring function to other tasks?
  • 3 How does the method perform on low-resource language pairs?

Applications

Immediate Applications

Low-Resource Translation

Improves translation quality for low-resource languages, benefiting translation companies and researchers.

Cross-Domain Models

Trains robust translation models on mixed-domain data, useful for multilingual platforms.

Long-term Vision

Automated Data Cleaning

Develops general-purpose denoising tools for large-scale data preprocessing and cleaning.

Abstract

Measuring domain relevance of data and identifying or selecting well-fit domain data for machine translation (MT) is a well-studied topic, but denoising is not yet. Denoising is concerned with a different type of data quality and tries to reduce the negative impact of data noise on MT training, in particular, neural MT (NMT) training. This paper generalizes methods for measuring and selecting data for domain MT and applies them to denoising NMT training. The proposed approach uses trusted data and a denoising curriculum realized by online data selection. Intrinsic and extrinsic evaluations of the approach show its significant effectiveness for NMT to train on data with severe noise.

cs.CL cs.LG stat.ML