Which one should be preferred when? LLM or Machine Translation or both?
Today, we will discuss the advantages and disadvantages of LLM or Machine Translation services.
- History of Machine Translation Evolution
- Top 10 Criteria to Select LLM or MT
- Final Decision: LLM or Machine Translation?
In most comparisons between MT and LLM, it is not accurate to compare early examples of MT systems. Many comparisons are designed to highlight the Generative AI and LLM trend or to attract organic traffic by leveraging this trend.
Let’s make this comparison in a more in-depth, objective, and fair manner.
History of Machine Translation Evolution
When examining the history of machine translation systems, we should start by noting that the techniques used in their development have evolved over time. Therefore, to make the most accurate comparison, we need to rely on the latest generation of neural machine translation systems (NMTs).
- 1940s – 1950s: The idea of being able to translate using a machine
- 1960s – 1980s: Rule-Based Machine Translation (RBMT)
- 1980s – 2000s: Statistical Machine Translation (SMT)
- 2010s – Present: Neural Machine Translation (NMT)
You can find dozens of technical articles about the technologies behind MT systems and LLMs. However, we will focus on which technology a SaaS company should prefer and when. When making decisions, you don’t need to know all the details of the technology used. However, understanding some basic principles will help you make decisions more easily.
Top 10 Criteria to Select LLM or Machine Translation
1. Quality
NMTs are significantly more successful than LLMs in producing quality translation results. To achieve high translation quality with LLMs, prompt engineering and more contextual information are necessary. Moreover, in terms of supported language pairs, LLMs generally support fewer languages compared to NMTs. Obtaining quality translations in low-resource languages supported by NMTs is not straightforward. However, the zero-shot translation feature of LLMs gives them a slight advantage in low-resource languages. Nevertheless, NMTs typically offer quality translations in commonly used languages.
To learn more, you can visit the Linguistic Quality Assurance: How to measure translation quality of AI translation? article.
2. Data Security
LLMs store context for use in their subsequent outputs, whereas NMTs do not retain any context for future translations. Consequently, data security concerns are more pronounced with LLMs. However, if the data is already public or will be made public, it might be wiser not to overlook the productivity and creativity advantages offered by LLMs.
3. Consistency
Since NMTs do not store context, they may struggle to produce consistent translation results. Their ability to understand context is limited to the data provided during each translation, making it challenging to choose the right terms, especially in very short translations. In contrast, LLMs, which store context, have a higher ability to produce more consistent results and maintain term consistency. This makes brand-level translations theoretically more achievable. However, determining the amount and quality of context required to reach this level of quality can be difficult in practice.
It is also worth noting that technologies such as glossary and translation memory have been developed to achieve consistency in NMT translations. When used correctly, these technologies can provide consistent translations that surpass the capabilities of LLMs.
4. Fluency and Grammar
Fluency is one of the evaluation criteria for translation quality. LLMs can produce much more fluent translations than NMTs. Additionally, LLMs are generally more capable of generating grammatically appropriate translations. LLMs can also be consulted for general grammatical corrections.
5. Creativity and Productivity
LLMs excel in the ability to provide creative or alternative translations, making them particularly useful in tasks such as post-editing or proofreading to enhance productivity. In terms of linguistic quality assurance, NMTs lack this capability.
6. Technical Maintenance
Starting to work with an LLM may require more development skills, especially during the technical integration process. LLMs are not yet at a sufficient level of technical maturity, so you may need to make frequent revisions to your integrations. NMTs, on the other hand, are much more mature and technically stable. After integration, changes are rarely needed.
7. Accessibility
Many NMTs can be used free of charge up to a certain limit, and some do not even require creating an account. While not all LLMs offer this, some do provide a free trial option. However, access to NMTs and the ability to experiment quickly with them is generally greater than with LLMs.
8. Speed
The translation speed of NMTs is generally higher than that of LLMs. However, this difference is not a significant differentiating factor. If the context provided to the LLM is extensive, you are more likely to experience slowdowns in its speed. Additionally, it should be noted that while the sole purpose of NMTs is translation, LLMs are designed for dozens of different tasks besides translation. This versatility may cause them to spend relatively more time interpreting tasks.
9. Cost
Although LLMs do not apply to all NMTs, they are generally more expensive. At first glance, directly comparing the pricing tables might suggest otherwise. However, the size of the initial context, the prompt length of the task, and the output length all factor into the pricing. If contexts and inputs are not planned efficiently—that is, if the goal of producing more efficient outputs with less context is neglected—it can result in high costs. NMTs, on the other hand, have a simpler pricing model based on the number of characters in the texts to be translated.
10. Data Requirement
The training data set required by NMTs is much larger than that needed by LLMs. If you do not use a generic NMT and want to train a customized MT yourself, you need to provide bilingual language data across many languages and domains. However, if generic models meet your needs, you do not need to plan for such a dataset. Many NMT service providers already collect these data sets, allowing them to offer translation services of sufficient quality using generic models.
On the LLM side, the context you provide for training purposes is much easier to use. The LLM’s ability to interpret complex languages directly affects its ability to interpret training sets, enabling you to achieve the desired translation quality with relatively fewer data sets. LLMs support zero-shot translation models, meaning the AI can translate between language pairs even if it has never been specifically trained to do so.
Final Decision: LLM or Machine Translation?
- Low-Risk Translations: MT or LLM + Automated LQA
- Medium-Risk Translations: MT or LLM + Post-Edit with LLM + Automated LQA
- High-Risk Translations: MT, LLM, or Human Translation + Human Post-Edit + Human LQA
If you want to make translation and localization across all risk levels within a single platform, Lugath can be your solution. With the intelligent MT hub and generative AI translation, you can establish a high-quality, creative, and efficient translation workflow. You can also get contributions from human translators and create an approval workflow with the collaborative translation product.
