What is a ChatGPT and how it can change the translation industry?
Developed by OpenAI - ChatGPT (GPT- Generative Pre‑trained Transformer) became one of the most popular AI tools for natural language processing. Continuous learning makes it even better daily, allowing use for various functions - text generation, information research, code creation, translation, and many more. The constant development - OpenAI released a more powerful GPT 4 in March 2023, unlike GPT 3.5 - expands abilities even further. For example, the chatbot can describe what is shown in the image, prepare a menu(recipe) based on a photo of food, create a structured agenda based on the input information, and many more.
The prediction of ChatGPT usage in the future is very optimistic - AI will be able to work almost independently on content-generating tasks (better context understanding) and provide support for users as a live assistant.
In this article, we want to overview one of the potential areas where ChatGPT can be continuously used - text translation. Let's look at it compared to other popular systems and ChatGPT's strong and weak sides.
How does ChatGPT work? A short overview
Neural networks (NN) are generating a lot of interest in society. Every time a new product is released, headlines like "Is AI dangerous?", "Soon ChatGPT will replace employees," "When will robots take over the world?" and so on, start to appear everywhere. Before moving on to the article's main topic, we will explain in a few words how all the NN "magic" actually works.
ChatGPT is a neural network based on the Transformer - deep neural network architecture introduced in 2017 by Google Brain. At one time, this technology "shook" the entire industry and contributed to its development. Transformer consists of two main components: an encoder and a decoder, so it takes one set of sequences as input and outputs the same set of sequences but converted according to some algorithm.
Unlike earlier solutions that processed data sequentially, the Transformer can process data simultaneously and "focus" on different parts of the input sequence depending on the problem to be solved. This way, models can work efficiently with long sequences and consider the context in processing each element.
The first successful experience with using Transformer for generating texts by the neural network was in 2018 - the release of GPT-1. OpenAI proceeded with the project's scalability - they added more parameters (from 177M to 1.5B) and upgraded the dataset (40 gigabytes of text from Reddit and its hyperlinks). The differences between the first and second models were astounding. While GPT-1 could generate small texts and had trouble with the correct context, GPT-2 had been creating an essay on the proposed topic.
GPT 3 and GPT 3.5
GPT 3.5 is something we know well, thanks to its growing popularity. But let's take one step back. GPT 3 - is an even more fed model than the second version. It has 175B parameters and 400 gigabytes of text (dataset.) The main difference between the 3rd and 3.5 models is that the second was trained with human feedback. People "communicated" with the neural network and rated its responses.
On March 14, 2023, OpenAI launched ChatGPT 4 - the outperformed version with more parameters to produce more accurate responses (40%) than GPT-3.5.
ChatGPT only needed 5 days to get the 1 million users audience. For comparison, for Netflix, it was 3.5 years; for Facebook, 10 months; for Spotify, 5 months.
Using ChatGPT for text translating: benefits
At this stage, we already understand how GPT works and can predict the main advantages of its use.
- Support for different languages. ChatGPT knows over 50 languages, including German, English, Chinese, Spanish, and others. With such a significant background, its usability and benefits for text adaptation are undeniable.
- Accuracy. Thanks to the neural networks, ChatGPT provides high accuracy of text translation, especially when working with complex and ambiguous phrases.
- Constant learning. The system always develops by collecting new knowledge and data from the users. Also, every new release from OpenAI gives us a more powerful neural network with a better understanding of context.
- Successful work with complex text constructions. Unlike other machine translation methods, ChatGPT does not rely on pre-built dictionaries and can learn independently from large amounts of text data. Such an architecture allows it to find more exact matches between words. You can ask the chatbot to provide different translation styles or use other constructions. For example, you can ask about adding synonyms, being more confident, using specific terms, etc.
- Speed. ChatGPT can process many texts in the shortest possible time due to its scalability and the ability to process data in parallel.
Disadvantages of using ChatGPT for text translating
One of the critical abilities of ChatGPT is natural language processing. Unlike Google translate, DeepL, and other systems, the chatbot uses a contextual understanding of languages to ensure more accurate translation.
Since the system was trained on the content available to it (Reddit, Wikipedia, etc.), not all languages will have the same translation level. This research describes how the quality of translation depends on the language's rareness in the web environments.
Based on the analysis, the ChatGPT performs flawlessly with the widespread languages and can compete in language pairs like German/English and English/German. On the other hand, English/Romanian and English/Chinese are not so good.
Thus, the disadvantages of using chatGPT in translation include the following:
Limited vocabulary (more or less) for some languages. The system knows only learned words and can have problems with unique terminology. This issue can be larger if we talk about non-common web languages.
As a result - limited understanding of context. ChatGPT cannot parse the context in which a word or phrase is used and, therefore, may not capture the nuances of the translation.
If ChatGPT is not perfect, why has it become so popular?
The neural network systems, an example, GPT 3.5, were available for usage even at the start of 2022. Why is there so much hype for well-known ChatGPT? The main reason is a very simple and familiar web interface for users. If earlier you needed to use an API (the model in the API existed in free access ~10 months before ChatGPT release) and be a confident user at least how to work with it, now every person can sign up and access the chatbot.
When the technology became available to the masses, constant feedback and posts on Twitter/Reddit increased interest and user engagement.
Google Translate, DeepL, ChatGPT, what to choose?
The best translating machine is still a human) Each technology can translate content quite well and be the basis for pre-translation. We observe with interest the development and emergence of new engines and algorithms in the translation field to improve Lingohub clients' user experience.
Google Translate provides more languages and DeepL deals better with a specific vocabulary. At the same time, ChatGPT can understand context better in popular language pairs, provide more translation variations, use synonyms if you ask, or add clarifying information.
In the context of localization, if you need a stable translation tool for a lot of data from different languages, we advise you to turn to a couple - DeepL and Google Translate.
In our Machine Translation feature, we use both (DeepL and Google Translate) to provide accurate pre-translated text for segments and speed up the localization.
Cherry on the top
ChatGPT is a promising and developing tool. Due to the massive wave of popularity that it has generated, even more, companies have begun and will begin to engage in language models, which will contribute to the qualitative growth of technology. And who knows, maybe one day we will use ChatGPT and similar tools for complete translation from any language.