Pangeanic at AMTA 2022: What Did We Learn?

Written by Marisol Letelier | 10/05/22

Every year, Pangeanic participates in numerous events and conferences in the field of natural language processing. This year, our team traveled to Orlando to attend the AMTA 2022 event.

AMTA 2022

This was the 15th conference of the Association for Machine Translation in the Americas held in Orlando, Florida, as well as online, from September 12 to 16, 2022.

The conference covered a variety of topics, demonstrations of the latest MT providers, tutorials for beginners and more experienced professionals, and workshops.

Others interested in MT and computational linguistics were able to connect with academic and industry mentors. They were given the opportunity to learn about what to expect from this technology that holds so many possibilities for the future.

Keynote speakers at this year's event included Marco Trombetti, CEO of Translated, Angela Fan, research scientist at Meta AI Research, and Dr. Alex Waibel, professor of computer science at Carnegie University and Karlsruhe Institute of Technology.

 

Key points of the conference

Nikita Teslenko and Carmen Grau attended on behalf of Pangeanic to present the latest advances in our language technology. They also seized the occasion to surround themselves with industry professionals and observe where MT and natural language processing stand, what advances and new developments are being proposed, and where innovation is leading us.

They both took away the following key points:

  1. The arrival of neural machine translation has opened the door to constant improvements in the field. Fascinating innovation was presented at the conference, such as new human evaluation metrics, dataset augmentation, or data selection and cleaning methods. However, we are still in the "Transformer" era.

     

  2. Dynamic adaptive MT is an innovative idea in which translations are modified on the spot, without the need for an expensive and time-consuming model retraining process. In addition, translators do not have to correct the same MT error twice, which is a frustrating and demotivating task. These days, some companies already claim to have this feature, but its definition remains unclear.

     

  3. Multilingual translation models seem to have a greater impact on translation results. Training with domain-specific data in languages that are similar to each other seems to improve in-domain translation in all of the languages, which is quite impressive!

     

  4. The pros and cons of large multilingual neural machine translation models were also discussed. These big models are being used and adapted to work with low-resource languages, and are yielding good results. They can significantly reduce post-editing costs. However, due to their size, different strategies need to be adopted to reduce computational costs. In this regard, reducing the carbon footprint and making neural machine translation sustainable is a major concern for the MT industry.

     

  5. The three keynote lectures were brilliant and impressed the audience. Marco Trombetti gave an inspiring talk on how far the MT industry has come and what remains to be done to face the new challenges of removing language barriers. Professor Alexander Waibel presented his extensive experience in the development of technologies for bridging the language gap. Finally, Angela Fan presented Meta's "No Language Left Behind (NLLB)" project, with models capable of providing high quality translations between 200 low-resource languages.

     

  6. Last but not least, as we all know and have seen at the conference, large amounts of quality data are what will make the difference in the near future of MT. In different talks, some text generation and data augmentation ideas were presented and proposed. However, much work remains to be done in this area to make it worthwhile and useful to the industry.

At Pangeanic, we like to keep up to date and aware of everything that is to come. You never know where new opportunities will be found, and it is at events like these that we can learn the most.

As a language technology and natural language processing company, our team spends every day researching and developing our technologies to offer the best services adapted to our customers' needs. Attending these events allows us to continue growing, and we think it is important that the members of our team participate.