Do Multilingual Language Models think better in English?!
Most large language model (LLM)-based chatbots are trained on data from dozens of languages, but English is still the dominant language, as most of the web data is written in this language. Because of this, multilingual LLMs have much better understanding and generation capabilities in English than the other languages. However, we still need LLMs to perform well in other languages to ensure accurate and reliable results in multilingual contexts.
One solution which has been around for a while is to detect the input language, translate it to English, let the LLM process the question and generate the answer in English. Finally we can translate the response back to the desired language of the user. Although this process works well, in practice, it means you will need a few extra tools, such as a language detector and a machine translator. These additional tools can add to the complexity of the project.
In an experiment conducted by the University of the Basque Country, researchers confirmed that multilingual LLMs perform better in English than other languages seen during training. Their study shows that using multilingual LLMs to translate the input into English and perform the intended task over the translated input works better than using the original non-English input. This work shows that letting the model translate the input by itself can achieve almost the same performance as using an external translation system. This opens up opportunities for end-to-end multilingual chatbots and other generative AI models without relying on external translation systems.
Lingua Custodia’s VERTO NLP financial document processing platform allows you to easily translate documents from several languages to English. Our NLP translation service detects both the source language and document type which helps to optimise the translation quality. Our generative AI document analyser service then allows you to process questions in English, and again the responses can then quickly be translated back to the necessary language.