Reality Series, Episode 6: Have LLMs mastered the art of translation?

How good are LLMs, like GPT, PaLM and Claude at translation and localization? And, is it possible to declare if one is better than the others?


LLMs have caused a major shift in the realm of translation. In this episode, we aim to demystify the advantages and drawbacks of employing LLMs in localization. We offer insights on making informed decisions about the utilization of LLMs in the translation/localization process, and explore the language capabilities of LLMs.

This discussion is led by Mei Zheng and Valerie Dehant.

Myth Busting: Have LLMs Mastered the Art of Translation?

In episode six of Smartling’s ‘Reality Series’, Mei Zheng, Senior Data Scientist, and Valérie Dehant, Senior Director of Language Services, tackled a few common myths related to LLMs and Machine Translation (MT), and whether Large Language Models (LLMs) have truly mastered the art of translation.

Busting some myths about LLMs and translation

Myth #1: LLMs are better than Machine Translation

The first myth was whether LLMs outperform MT in the realm of translation. Offering her insights, Mei dispelled the myth, asserting that while LLMs have commendable general language understanding, their skills are not particularly made for translation. They will not outdo MT systems (at this time, anyway), but they do add value to translations by contributing to grammatical precision and enhancing fluency of language.

Myth #2: LLMs can translate all language pairs

The second myth challenged the all-embracing capabilities of LLMs, holding that they can translate across all language pairs. Strongly contradicting this belief, Mei pointed out that the majority of LLMs are restricted in their multilingual abilities. Mei specifically emphasized the importance of examining model cards, a tool for transparency in large language and machine learning models to help understand which languages it supports, and to test them for specific translation use-cases.

Myth #3: LLMs caneplace humans

Lastly, the session addressed the common assertion that linguists could be replaced by LLMs. While underscoring the strengths of LLMs, Valérie painted a clear picture of why human linguists remain indispensable. LLMs, even with their pattern recognition skills, can overlook nuances which are consciously discerned by human linguists. Also, LLMs have a tendency to hallucinate information–providing ‘translations’ for phrases that don’t exist in the source text!

As the roles of linguists evolve alongside generative AI, they transform into co-pilots supervising and evaluating the quality of MT suggestions. The expertise of human linguists remains crucial for maintaining the quality and accuracy of translations.

A further look into LLMs and translation

While clarifying the myths and realities around LLMs, our speakers dove deeper into the current dynamics of translating in the era of LLMs and MT systems. Mei further stated that the challenge of evaluating translation quality goes beyond fluency, involving complexities such as HT (the results after human editing), and semantic similarity. The challenge of LLMs providing incorrect translations due to multiple meanings in the target language was also highlighted.

Illustrating the brighter side of AI, Valérie explained how, with increased efficiency, linguistic teams can translate more content within the same time frame, without needing to increase manpower. When asked about training LLMs, Mei spoke about the fine-tuning procedure that involves adapting the top layer of the model while keeping the foundational parameters unchanged. The data preparation for LLMs broadly aligns with MT, however, LLMs might require specific prompts to tailor translations better.

When comparing LLM translations to professional ones, Mei confirmed that despite all progress, human input continues to be needed. The industry is actively figuring out which parts require human intervention and which do not.

The verdict

Episode 6 concluded that although AI and LLMs have made a sizable impact on translation, human intervention is indispensable for securing quality and accuracy in translations. Human translation is here to stay for the foreseeable future. While LLMs are making strides, they have not yet taken the throne in the realm of translation. They are, however, playing a vital complementary role in amplifying the capabilities of translation and shaping the future of this industry. Ultimately, machine translation and large language models can complement each other to deliver improved translation services. However, remember, the human touch remains irreplaceable.