EAMT Conference Summary

EAMT Conference: Life after ChatGPT

We attended the annual conference of the European Association for Machine Translation, which took place in Tampere, Finland this June. The event brings together industry professionals, academic researchers, and innovators from around the world to celebrate and explore the advancements in the field of machine translation (MT). Unsurprisingly, the topic of LLMs (large language models) was on everyone’s mind.

LLMs use cases for translation

While many of us remain in a state of anticipation of what comes next after the release of ChatGPT, more seasoned conference attendees see a lot of similarities with the neural MT revolution that happened around 2015-2016. We are not quite there yet, but the general prediction is that LLMs will eventually take over.

Here are a few most interesting use cases where we already see promising results.

1. LLMs as automatic translators

Unlike most existing MT systems, LLMs can translate chunks of text that include more than one sentence, which means taking into account wider context and producing more fluent and consistent translations. In addition, they can be used for few-shot translation: you can adapt the translation to your style by including some correct translation examples in your prompt. This characteristic of LLMs means that they can overcome some of the limitations of neural MT.

2.    MT evaluation

It looks like GPT models are quite good at choosing the best MT engine out of several candidates based on their translations only (without the correct reference translations). They may soon be widely used for MT quality estimation.

3.    Improving source text for MT

Another promising application consists in using LLMs to make source text more “MT-friendly”: normalizing the spelling, simplifying, and making it more consistent, which results in better MT output.

Many participants from academia and companies seemed quite convinced that it’s only a matter of time that LLMs will become better at translation than specialized MT systems, and will also be successfully used for translation quality evaluation. However, as we progress and the overall quality of automatic translation improves, the risks of error will not disappear and the last word must still belong to human experts.

Public Impact and Regulations

A panel discussion on the topic of generative AI covered its societal, ethical, and political aspects. As the business interest of the companies behind the GenAI revolution is not necessarily fully aligned with the public interest, regulation is needed to ensure better quality and a great future for this technology and us as its users. Participants called for the creation of policies that would increase the number of open-source non-commercial models and ensure that the power of AI is not only concentrated in the hands of the big tech. Our regulations should create a financial incentive for tech companies to do the right thing, as opposed to banning or restricting the development of this technology. We cannot stop the progress, but we can decide the direction it’s going to take.

The question of public trust is also an important one. We must correctly set the public’s expectations on what these systems can and cannot do well. If they are not set, the mistrust towards the AI community will turn the public against it.

Beyond LLMs

Of course, it was not all about LLMs. “Traditional” neural MT is not going anywhere and there were a lot of excellent presentations on many related topics. These are only a few of them.

Speech translation

Spoken language translation has always been one of the north stars for the MT community, and one of the most challenging problems to solve. In his brilliant keynote talk, the head of MT at Zoom showed that many challenges remain, but we are getting closer and closer to real-time, good-quality translations of our online meetings.

Ethical and social aspects of MT

It is encouraging to see a lot of research that is taking a more socially conscious direction. Several works this year were dedicated to such topics as an automatic translation of sign languages, mitigating gender bias in MT (these two topics even had dedicated post-conference workshops), and expanding support for low-resource languages.

Terminology compliance

Last but not least, several sessions touched on the topic of ensuring terminology compliance, including the one by TransPerfect’s research team. In our experiments, we compared different methods for enforcing terminology in neural MT and investigated the effect they have on the overall quality of the translation. “Seamlessly” instructing MT engines to produce a correct translation of terms is not a trivial problem, especially for languages with rich morphology, and this work helped us ensure we use the best solution while minimizing any quality risks.

If you are curious to know more, you can read the full proceedings of the conference, or contact aiteam@transperfect.com.