A high performance open-source library for NLP Transformer model training and inferencing.

Hello fellow readers! In this post, I would like to share a recent advance in the field of Machine Translation. Specifically, I will be presenting the paper Neural Machine Translation with Monolingual Translation Memory by Cai et al, which received one of the six distinguished paper awards from ACL 2021.

Self-training is a very prevalent semi-supervised method. Its key idea is to augment the original labeled dataset with unlabeled data paired with the model's prediction (i.e. the pseudo-parallel data). Self-training has been widely used in classification tasks. However, will it work on sequence generation tasks (e.g. machine translation)? If so, how does it work? This blog introduces a work [1] which investigates these questions and gives the answers.

Motivation

While the advance in deep learning has dramatically improved the machine translation quality, there is little development in the evaluation of machine translation models. The most widely-used metrics like BLEU [Papineni et al., 2002] and METEOR [Lavie and Denkowski, 2009] simply match the n-gram between the hypothesis text and reference text, which is too rigid without considering the variance in ground-truth translations and fail to differentiate the current highest performance machine translation models. They also cannot be accurately correlated with human judgment for a piece of text.

Speech translation (ST) has increasing demand in our daily life and work. Applications like travel assistant, simultaneous conference translation and movie subtitling can highly reduce translation costs. Building a ST system that can understand and directly translate acoustic speech signals into text in a target language is challenging. For example, people do not always premeditate what they are going to say. Not like text translation, ST lacks completed organization sometimes. Another part is that the parallel corpus for ST is not enough, compared to the MT task. Especially, most ST methods are limited by the amount of parallel corpus.

​Upon its emergence, the Transformer Neural Networks [1] dominates the sequence-to-sequence tasks. It even outperforms the Google Neural Machine Translation model in specific tasks. Specifically, the multi-head attention mechanism that depends on element-wise dot-product is deemed as one of the critical building blocks to get things to work. But is it really that important?

Introduction

A typical neural machine translation (NMT) system needs to support the translation among various languages, that is, a multilingual(many-to-many) NMT system rather than only support the translation between two languages. However, to support the multilingual translation is still a challenge. One direct idea is to use a separate model to translation one language to another language, which is very easy to implement but brings high costs: To support the translation among N languages, we need train N(N-1)/2 separate models. Such a method does not allow the sharing of information across languages, which can result in overparameterization and sub-optimal in performance. We denote this method as per-language NMT

Can one build a neural machine translation model without parallel data?

In general, neural machine translation (NMT) requires a large amount of parallel data (e.g., EN->CN). However, it is not easy to collect enough high-quality parallelly-paired sentences for training the translation model. On the other hand, we can capture enormous plain text from Wikipedia or news articles for each specific language. In this paper, MGNMT tries to make good use of non-parallel data and boost the performance of NMT.

Machine translation has helped people daily life, and is also an important research topic especially in computer science community. It expands from one language translate to another language, speech translate to text, etc. Today, I'm going to talk about a paper "Generative Imagination Elevates Machine Translation". I'll cover the background, challenge and motivation behind this paper. Then I'll go through some technical details of this paper as well as some in-depth analysis of their experimental settings and results. Finally, we will discuss about the potential extension of this work. Hopefully, this would give you a better understanding of this area, and point out to a promising research direction.