The first step of establishing a Many-to-Many dataset is to select 100 languages for which there already exist high-quality, annotated datasets that can be used for model evaluation.

                 
       Therefore, We cover a diversity of scripts and resource levels (as shown in Table 1) to have high coverage of languages worldwide. (widely spoken language)

Language selection

Supervised translation systems rely on large quantities of parallel sentences

Hence, we use mining technique

Covering the Language Matrix

We use publicly available evaluation datasets to evaluate the performance of all of our models. 

To cover our set of 100 languages and 2200 direction

Evaluation according to existing models

The evaluation of a many-to-many multilingual translation model can include a human evaluation study to assess translation quality. This study can focus on translation directions that do not involve English, particularly between languages commonly spoken in the same region (e.g., Japanese-Chinese, Hindi-Tamil, and Russian-Ukrainian).

Human Evaluation of English-Free Translation Directions

1) Language selection \n 
2) Evaluation
3) Covering the Language Matrix by Mining

University of Michigan - Ann Arbor

1) Mechanism
2) Data collection
3) Methods

English-Centric Multilingual Machine Translation

How to Build a Many-to-Many Parallel Dataset for 100 Languages

In a Many-to-Many translation setting, the amount of parallel data grows quadratically with the number of languages, so a standard-capacity neural network underfits rapidly. To supply enough capacity, model size is scaled far beyond typical bilingual systems: models over 50 times larger than current bilingual models are trained using model parallelism, building on scaling results from Kaplan et al. (2020) and Arora et al. (2018).

Learn Before

Related

Learn After