The Future of Machine Translation

It’s an exciting time for translators indeed, with December 2017 seeing the launch of two AI systems able to teach themselves any language. According to Global Market Insights, the translation industry will have a $1.5 billion net worth by 2024, with “the ability to translate different languages according to customer preferences and the lack of existing translator for several specialized fields and language combinations providing tremendous growth opportunities for the industry.” This is good news for the future of machine translation, and both for individual translators and translator tech developers. Both will be called upon to deliver increasingly specialized translation solutions to fulfil an ever-growing global demand.

Technology does not Aim to Replace Translators

Before delving into the machine translation systems that will be dominating the industry, it is important to understand that these systems are not meant to replace human translators. For example, Korea has one of the highest robot populations in the world and hardly 0 unemployment. Higher automation creates more "human" and better cognitive jobs. Translation continues to be an attractive profession that is directly in line with the millennial demand for online-based income sources. The annual mean wage for this profession stands at around $45,910, though salaries can be considerably higher. The spirit behind new technology is one of co-existence and of easing the burden of translators working on large volumes.

The Global Machine Translation Market - Forecasts from 2018 to 2023 Report

This report delves into market dynamics, global machine translation, and competitive intelligence. It identifies four major technologies that will be influencing the global machine translation market: statistical machine translation, rule-based machine translation, hybrid machine translation, and neural machine translation. Its authors note that “The major factors driving the market are the growing volume of big data, the need for cost effective translation, and increasing online content. Moreover, globalization is raising the demand for location based content across various industry verticals such as electronics, travel, e-commerce and hospitality.”

Neural Machine Translation (NMT)

This is the latest approach in industry and the closest related to artificial intelligence. It reduces the post-editing effort thanks to its focus on fluency. In many ways, it is an "ultra-statistical" approach, with several layers processing the information and nodes passing on information and ensuring accuracy. It has been called neural as the way layers and nodes work resembles very much the way neurons work in humans (and mammals in general). There are three interesting features that Neural Machine Translation can offer. Some of its features help it understand the similarities between words, analyze entire sentences, and evaluates the fluency of a sentence in the target language by analyzing a few words at a time. NMT is just one of many exciting ways to elicit more natural, fluent sounding translations in the target language. In fact, academia is still struggling to understand what happens inside the "black box" once the neural network starts to train. The "conclusions" are pretty amazing and sometimes surprisingly accurate when it comes to human language processing because of the number of calculations, weights, gradients applied and attention models, even between non-related languages and languages with a rich morphology.

Statistical Machine Translation

This system employs a learning algorithm to a large body of previously translated text, using this algorithm to translate texts that have not been encountered before. The system is suitable to documents that focus on one particular subject. Its advantages include the plethora of existent current platforms and algorithms (which make the system cheap and quick) and the fact that it requires little space (i.e. does not need a server of its own). Training is done in CPU servers and it is easily deployed. Decoding is also fast and it serves pretty much as large "translation memory" that places n-grams (groups of words) next to each other. Most systems we are familiar with were statistical until the end of 2016, when Google published its paper on neural machine translation. Its weak point is that it works in a specific context and is not adept at colloquial language and idioms and also it works better with close languages but it does not perform well syntactic re-ordering (as it would be required when translating from English into German or Japanese).

Rule-Based Machine Translation

This is the first system developed in the world of machine translation and it dates from the 1950's; it relies on the rules of grammar, lexicon, and software programs to process rules. Translation occurs by pattern matching rules. The key attribute of this system is that it permits the avoidance of matching unfruitful rules. The strong point of rule-based translation is its ability to deeply analyze language at the semantic and syntactic levels. Its weak point is the large numbers of rules that govern each language which may end up contradicting each other and the "mechanic" sound of it some times.

Hybrid Machine Translation

This method uses multiple machine translation approaches within one system. There are many types of hybrid machine translation. One popular approach combines a statistical engine with a rules-based approach (both at the pre- and post-processing stage). The system offers great flexibility, control, and precision, and differs from a purely statistical-based approach. Historically, it aimed at adding the best of rule-based MT (syntactic information, that is "understanding" of collocation and where words should go, their proper declension in the case of morphologically-rich languages, re-ordering, etc.) combined with the power of data-driven statistical MT.