How AI Has Changed the Game for Machine Translation

Thumbnail Image
Yeffeth, Isabella Anne
Issue Date
Research Projects
Organizational Units
Journal Issue
Alternative Title
In this essay, we will be discussing all of the issues that any machine translator will have to overcome in order to create a “perfect translation”. First, we will be discussing the difficulties of translation itself, specifically with a focus on translation between English and Japanese. We will talk about how the target language, the language that we want to translate to, and the source language, the language that we are translating from, interact with each other. There are many things to consider in this category, such as the differences in vocabulary between the two languages, words with double meaning, words that do not have a direct translation in the target language, and how the two languages’ grammars differ from each other. Grammar is a difficult hurdle to cross, especially when the languages are very different, and the sentence structure becomes convoluted in more advanced works. However, this is far from the most difficult thing about translation. The next step deeper into the art of translation would be context and cultural significance. This can be a difficult thing to translate, even if the translator is fluent in both languages. In particular, as the works become more casual and more culturally influenced, this grows in difficulty. The more colloquial phrases, word play, allusions, and cultural phrasings (such as being indirect for the sake of being polite or other social norms that are represented through language but not explicitly stated) that are implemented, the more difficult the work can be to translate. The “perfect translation” requires a fluency in both languages, as well as a perfect understanding of the translation piece and its context. After addressing the complications of translations, we will look at what a computer can do to solve these problems, if anything yet. For this section of the piece, I will be focusing on machine translation, or any computer that is constructed with the intention to translate works, and what it can do without artificial intelligence (AI), which will be discussed afterwards. Without AI, machine translation is very limited, and so we will primarily be discussing those limitations, in order to more fully understand what AI does to take machine translation to the next level. Specifically, I will be focusing on dissecting the computer’s ability to understand the works it’s translating, what it means for a computer to understand, and why that is so important. For the next section, I will be taking a step back from translation to describe artificial intelligence and how it works. For a clearer explanation, I have chosen to use the example of image recognition in order to explain AI, because I think that that is the most straightforward and understandable way that AI can be used today. Key concepts such as neural networks and neurons and weight will be broken down and described. This section is for giving an overview of AI, so that it can be more easily understood how the computer could build upon itself, and start to learn the concepts we describe earlier, like the difficulties of language, and understanding a human language. Finally, I will pull all of these concepts together to describe what an AI-enhanced machine translator would look like. A lot of the ideas that I bring up in this piece have yet to be implemented fully, due to the complexity of the problems described, however, I do bring up the machine closest to understanding language at this point, Google’s BERT. BERT is still a work in progress, but I describe the method at which he learns, what he is capable of at this point, and why his methods are so effective. There are also studies from the way that babies learn languages, and how that method of learning reflects on BERT’s successes.
39 p.
U.S. copyright laws protect this material. Commercial use or distribution of this material is not permitted without prior written permission of the copyright holder. All rights reserved.
PubMed ID