Machine translation (MT) has been a popular branch of NLP and, by extension of AI in general. It has come a long way in recent years thanks to the advent of neural machine translation and since 2023, GenAI or LLM translation. Nevertheless, it still has some way to go, particularly when we think about expressions, turns and idioms, One of the biggest challenges facing MT is the lack of such high-quality parallel corpora. Today, we are going to deal with the scientific need for human-like corpora in machine translation and why systems will require constant updates.
First things first: Parallel Corpora – A Definition
Parallel corpora are sets of text that are aligned in two languages, so that each sentence in one language has a corresponding sentence in the other language. These corpora are essential for training MT systems, as they provide the data that the systems need to learn how to translate between languages.
There are two main types of parallel corpora: human-like corpora and synthetic corpora.
a) Human-like corpora are created by human translators, who translate text from one language to another. This type of corpora is the result of translators using CAT-tools and exporting their work as a TMX file.
b) Synthetic corpora are created by computers, using algorithms to generate artificial translations.
Synthetic corpora have some advantages over human-like corpora. They are typically much larger than human-like corpora, and they can be created at scale, more quickly and easily. However, synthetic corpora also have some disadvantages. They often contain literal translation errors, and they may not be as representative of real-world language use as human-like corpora.
Another technique is back translation. This is a child technique from synthetic corpora and that can be used to improve the quality of MT systems that are trained on synthetic corpora. Back translation involves translating a text from one language to another, and then translating it back to the original language. This process can help to identify and correct errors in the synthetic corpora.
However, back translation is not a perfect solution. It can introduce even new errors into the corpora, and it can be time-consuming and expensive, as well a means of introducing simplified translations and word-by-word versions, not reflecting the nuances of the original language.
For these reasons, it is important to use human-like corpora whenever possible to train MT systems. Human-like corpora are more accurate and representative of real-world language use, and they can help to produce better translations.
Benefits of using human-like corpora to train MT systems:
- Accuracy: Human-like corpora are more accurate than synthetic corpora, because they are created by human translators who are experts in their field. This means that the translations in human-like corpora are more likely to be correct.
- Representativeness: Human-like corpora are more representative of real-world language use than synthetic corpora. This is because they are created from real-world text, such as news articles, blog posts, and social media posts. This means that the translations in human-like corpora are more likely to be natural and fluent.
- Relevance: Human-like corpora are more relevant to the needs of MT users than synthetic corpora. This is because they are created from text that is relevant to the topics that MT users are interested in translating. This means that the translations in human-like corpora are more likely to be useful to MT users.
Challenges of using human-like corpora to train MT systems:
- Cost: Human-like corpora are more expensive to create than synthetic corpora. This is because they require human translators to translate the text.
- Time: Human-like corpora take more time to create than synthetic corpora. This is because they require human translators to translate the text.
- Availability: Human-like corpora are not always available for all languages. This is because there may not be enough human translators who are fluent in both languages.
Despite the challenges, the benefits of using human-like corpora to train MT systems outweigh the costs. Human-like corpora are more accurate, representative, and relevant than synthetic corpora, and they are more likely to produce better translations.
Conclusion
Human-like corpora are essential for training MT systems that produce high-quality translations. While synthetic corpora and back translation can be helpful, they cannot replace the need for human-like corpora. As MT technology continues to develop, the demand for human-like corpora will only increase.
