Let's read the text
Required reading: 5 minutes
1. How would you explain machine translation to an eight year-old?
Imagine there is a new game you would like to play, but the instructions are only available in English - which you don't understand. It would be great to have them in German, wouldn't it? But nobody of your friends or family can translate it for you. Somebody from your family certainly does, your English teacher could do it for sure. Unfortunately, translating all these instructions would need much, much time, which they don't have. So, it would be great if a computer could do it for you, wouldn't it? And guess what? It is possible! It is called "machine translation", because it is not a person who translates from one language to another, but a machine (computer). The computer may not be as perfect as your English teacher, but it would be enough to understand the game instructions. And if some parts are not completely clear, you can ask somebody to help - this would not take too much time.
2. Machine translation or human translation? - What are the strength and weaknesses of a machine in comparison to a human being?
The strengths of a machine lies in the speed and availability of online tools. Anyone can access a machine translation on the web within seconds. The quality of these translations has increased significantly in recent years. There are now good options where mistakes can be tolerated. If a user wants to find a news article Understand or something on the web in another language, the grammar or style is not that relevant. Even if the meaning of the original text is not completely transferred, it usually has no serious consequences. Machine translation is therefore a good tool if a text does not need to be perfect.
In addition, machine translation can be the basis for professional, human translation. However, I would like to add that many professional translators would disagree. They are often forced to use machine translation to minimize the workload and therefore the costs. Their attitude towards MT is more positive when they can choose what and how much they want to process.
"It is not recommended to forego human participation."
However, there are also clear weaknesses. Although machine translation has improved in recent years, even the best systems are still susceptible to error and can not be fully utilized. It is therefore not recommended to use MT for "serious" texts - such as analysis or legal documents - and to refrain from human involvement.
All in all, the degree of human intervention required for a particular translation task depends on the purpose of the translation and the value of the content.
3. Is there any MT tool that you would recommend?
For daily, simple usageI would recommend
- Google Translate (supports a large number of language pairs),
- DEEPL (English, German, French, Spanish, Italian, Dutch, Polish)
For the translation of large amounts of text Amazon translate is one possible option. (Arabic, Chinese, French, German, Portuguese, Spanish, Japanese, Russian, traditional Chinese, Italian, Turkish, Czech)
And for training own systems for desired language pairs and domains, based on own data: SockeyeHttps://github.com/awslabs/sockeye)
"It is astonishing that neural networks can only learn many different things by combining mathematical functions."
4. Which upcoming ML trends are most fascinating at the moment?
Neural networks, For various reasons. It's amazing that you can only learn a lot of different things by combining mathematical functions. You can also enter different types of input features. As part of machine translation, this allows not only learning through words, but also word units, linguistic characteristics or even terms from other languages.
There are many different architectures for artificial neural networks and this is fascinating. I must mention that It is often unclear what really happens in a neural network. Therefore, it is difficult to analyze and understand translation errors. This is a clear disadvantage, but understanding such complex processes is a challenge that interests me.
5. How did you discover machine translation/NLP/NLG as a topic for yourself?
The first step was my Master Thesis which deals with speech processing. Afterwards, I've worked in a research institute on speech recognition, which combines speech processing and machine learning. In the meantime, I've learned about other language processing tasks and natural language processing in general, and at some point I've got an opportunity to start PhD in machine translation. Since then, I am involved mainly in machine translation, but also in some other NLP NLP tasks.
"Excessive media coverage is harmful to users and MT developers."
6. What advice would you give a non-expert (layman) that just wants to start learning about machine translation?
To start with the book of Philipp Köhn "Statistical machine translation" to get an idea about what's going on. It covers only phrase-based MT though, but additional chapter about neural MT is available online.
Also, if possible, to attend the MT Marathon, where MT experts provide lectures and internships for beginners.
7. What do you tell people who are afraid of machine learning or AI?
Don't be afraid 🙂
In recent years, improvements introduced by neural networks have been accompanied by a good deal of media hyperbole, some of which suggesting that several professions, including translation, may be under threat. Such statements, suggesting that machine translation reached the level of human translation, are harmful. Both - for users as well as for MT developers. It is important to know that MT is not a solved problem,and will not be any time soon,
Photo by Paola Assenmacher, EyeEm
Maya Popović Studied at the Faculty of Electrical Engineering at the University of Belgrade and continued her studies at RWTH Aachen, where she earned her doctorate with the work "machine translation: Statistical approach with additional language knowledge". She then researched at the DFKI Institute and at the Humboldt University of Berlin and now at the Adapt Centre at Dublin City University. Her research focuses on various approaches to the evaluation of machine translations as well as other – partly related – NLP tasks, such as text simplification. In addition, she worked on machine translation for underrepresented European languages and combined linguistic knowledge and statistical/data-driven methods for NLP.