We Saw in my previous blog that the first idea for Statistical Machine Translation was introduced by Warren Weaver as far back as 1947. As with many innovations, the idea was way ahead of the technological ability to make it happen. But with the growth of super powerful computers, the introduction of the Cloud as a repository for these super-computers, and the provision of these to LSPs at an affordable price permitted the theory of SMT as a solution from translation of languages to become commonplace.

But as with all technological breakthroughs, once the tipping point is reached things grow exponentially. While SMT could provide a translation model – albeit in a limited capacity – scientist began to look at the next iteration of machine translation – Neural Machine Translation.

A generic neural network is a real or virtual computer system that is designed to mirror the brain in its ability to “learn” to analyse bodies of complex data. Neural networks, particularly those with multiple hidden layers (hence the term ‘deep’ – the more layers, the deeper) are remarkably good at learning input->output mappings.

Such systems are composed of input nodes, hidden nodes, and output nodes that can be thought of as loosely analogous to biological neurons, albeit greatly simplified, and the connections between nodes can be thought of as in some way reflecting synapse connections between neurons in the human brain. As stated, the ‘deep’ element of the concept refers to a multi-layered processing network of neuron filters.

In the human process, instructions and information flows from the brain through neurons connected by synapses. In the machine translation equivalent artificial neurons are used to fine-tune and refine input data as it is passed through the translation ‘engine’ with a view to achieving a predicted output. In addition, the process called Deep Learning learns – as does the human brain – from experience and can adjust and remember processes accordingly. Consequently, the more an engine is used the more refined and powerful it becomes. The process becomes a life-long R&D development.

So, in translation, the language engine uses deep neural networks to essentially predict the next word in a sequence. The Neural Network is built using a bi-lingual corpus and is trained on full sentences. (These are referred to a sequences.) This is important, as NMT fully understands what a sentence is, whereas SMT only understands what a phrase is. This shift to training Neural Networks on sequences means that we eradicate almost all the syntactical and grammatical errors frequently found in SMT outputs.

For example, NMT systems make significantly fewer errors when compared to SMT systems:

(NOTE: every graph below 0 is the percentage decrease in error types for NMT systems compared to SMT systems.)

Therefore, NMT provides a demonstrable improvement in machine translation outputs, resolving many of the problems associated with the statistical machine translation (SMT) model. Effectively, in the last two years it has solved many of the translation shortcomings of the SMT system; deficiencies that SMT users have been trying to resolve for the last two decades. NMT-based R&D has been moving with a speed far out-pacing the slow pace of SMT development.

Amazingly, Google managed to condense in to four years the development of its NMT solution. That compares with 15 years it spent on developing and refining statistical machine translation (SMT). This quantum leap in output quality for NMT systems is the result of being able to train the translation models on ‘sequences’ (i.e. full sentences) rather than words, and the use of deep learning to train and refine the neural networks on an ongoing basis.

So, a Neural Network is a mathematical model seeking to imitate how the human brain functions. Obviously, scientists can only do this in a modest way. The human brain consists of approximately 100 billion neurons, and these are highly connected to other neurons creating super-complex relationships. In a Neural Network a neuron is a ‘word’ within a sequence (sentence), and each word is highly connected to words within the same sequence creating a complex web of grammatical and syntactical relationships.

By creating these neutral models with lots of sequences, we can build up an elaborate network of word relationships. We can use this network to predict translations within the context of the training data used to build the translation model.

To determine if a model is good at predicting translations (relative to the training data), we use a mathematical number called Perplexity. We want to build networks with a Perplexity as low as possible. If we achieve this, we can predict good quality translations that contain fewer syntactical and grammatical errors when compared to SMT systems.

These highly efficient models can then provide fast and fluent translations, and at economically affordable prices. Today, over 90% of the daily traffic on the KantanMT platform is processed by our NMT services. Slator.com has reported that the number of LSPs now using NMT has quadrupled since 2018. This gives an indication of the high regard customers have for the efficacy of the NMT platform. The powers of deep learning, the availability of huge amounts of data and the ever-increasing processing power of computers makes it inevitable that this branch of artificial intelligence will be with us for the long haul.

One of the concomitant effects of the introduction and widespread use of viable machine translation solutions within LSPs is the inevitable re-organising of the translation process, and the realignment of existing jobs to fit with the newest technology trends. I will look at this in my next blog and try to predict how job roles might evolve over the coming years.

Aidan Collins, Marketing Manager at KantanMT.

The first ideas of Statistical Machine Translation were introduced by Warren Weaver as far back as 1947. He explained that language had an inherent logic that could be treated in the same way as any logical mathematical challenge. He contended that logical deduction could be used to identify “conclusions” in the target (untranslated) language based on what already existed in the source (translated) language. With the advent of the cloud, and affordability of powerful computers, the theory of Statistical Machine Translation (SMT) became a practical option.

Therefore, SMT is a machine translation paradigm where translations are generated based on statistical models, whose parameters are derived from the analysis of bilingual text corpora (text bodies) – a source text of translated material and a target text of untranslated material.

Statistical machine translation starts with a very large data set of approved previous translations. This is known as a corpus (corpora is plural) of texts that is then used to automatically deduce a statistical model of translation. This model is then applied to untranslated target texts to make a probability-driven match to suggest a reasonable translation.

Where are these large data sets sourced? Well, over many years most large global organisations, for example the EU, UN, World Bank, World Health Organisation etc develop enormous domain-specific corpora, in multiple source/target language combinations. Many of these have originated as human-translated texts. These are then made accessible to machine translation users/developers to further refine and use for MT purposes. The process is an evolutionary model with each existing corpus being refined, added to and updated on an ongoing basis.

This SMT paradigm is based on what’s known as ‘probabilistic mathematical theory’. Such theory suggests the chances (probability) of something occurring depending of different variables likely to influence the event. For example, tossing a dice gives the probability of one dice being a certain number is 1/6. So, a gambler has a one in six chance of landing on their chosen number. The probability of two dice being the same number is 1/6 x 1/6 = 1/36. In poker, the professional gambler keeps mental track of the cards being played and using probability theory he/she decides whether a gamble has a good chance of winning based on the odds they have calculated in their head.

In SMT, the MT engineer builds a Translation Model using the frequency of phrases appearing the training corpus into a table. This table stores the phrase and the number of times this repeats over the entirety of the training corpus. The more frequently a phrase is repeated in a training corpus, the more probable the target translation is correct. Each phrase (stored in the Phrase Table) can range from one to five words in length. This phrase table is referred to as the Translation Model.

Consequently, the MT engineer is using a probability model to hit on the right source/target translation combination. The process is evolutionary as the corpus is refined and adjusted after each translation run to eliminate/adjust any anomalies. The more frequently the corpus is used, the more perfected it becomes. The development of the corpus quality is a continuous organic R&D process of a highly valuable translation asset.

Additionally, the MT engineer builds a secondary model using the target translation data. This model helps determine the order in which the engineer needs to assemble phrases (from the Phrase Table) in order to optimise translation Fluency i.e. to give the translated text its natural language flow. Fluency ensures that literal translations (i.e. the words are all there, but the sense of the sentence is not) are replaced by a more natural sounding translation.

In order to translate a source sentence, the MT engineer goes through the following decode process (as outlined in the diagram below):

1. He/she breaks the source language sentence into phrases. (You can see the phrase as individual grey blocks in line two of the diagram.)
2. He/she then looks up each of these phrases in the Phrase Table/Translation Model and generates the target language translations. (You can see this in line three of the diagram.)
3. The engineer then uses the Phrase Table/Translation Model to re-order these phrases to optimise translation Fluency. (You can see this reordering in line four of the diagram.)

In my next blog, I will give a short introduction to the increasingly popular MT model Neural Machine Translation.

Aidan Collins, Marketing Manager at KantanMT.

In a blog I wrote in May of 2018 I commented that when it comes to IT nothing develops faster than new technology (“In Technology, Things Never Slow Down – Quite the Opposite”). It seems that once a new technology is born, its exponential growth becomes unstoppable. And that – I would argue – is exactly where we are with machine translation. We are watching an amazing exponential growth, with Neural Machine Translation (NMT) now the new, hot kid on the block.

Undoubtedly, NMT is now mainstream and LSPs are rapidly adopting it as a viable translation technology option. Machine Translation has followed the exponential growth path of other technological innovations. Amazingly, Google managed to condense in to four years the development of its NMT solution. That compares with 15 years it spent on developing and refining statistical machine translation (SMT). In November 2016, Google give up on developing SMT and unveiled its newly developed NMT system (GNMT). Google announced that their new NMT technology was already able to translate eight language pairs. Within 12 months GNMT was supporting another 90 language pairs. The exponential growth was astonishing: the new kid on the MT block had arrived and was soon centre stage.

A Transformer Neural Network

 Image Courtesy of: “The Illustrated Transformer” by Jay Alammar

In tandem to this rapid NMT development, the emergence of relatively cheap, super-powerful computers – coupled with practically unlimited storage capacity due to the emergence of the “Cloud” and the Cloud’s more recent iteration “the Edge” – the way was opened for most language companies to adopt complex NMT technology as a standard customer service offer. The way has been created for the language industry to use automated translation technology to process even more content, in to more languages, and faster than ever before.

In 2018, a www.Slator.com report revealed that in just 12 months the number of LSPs offering NMT as a service had quadrupled from a low base of five to 20. Perhaps more pertinent than the numbers are the status of the LSPs adopting the technology as their go-to language solution for bulk translation projects. Significantly too, NMT is now the standard solution of major global entities in both the private (e.g. Amazon, Google, Microsoft etc) and public sectors (e.g. the EU).

So, as LSPs move speedily to integrate Machine Translation into their service offerings and workflows it’s perhaps a good time to look at the ongoing debate about “quality” in NMT production. Quality has always been a hot topic and fiercely debated within the industry. Whole conferences, often in exotic locations, have been given over to passionate debates trying to define the “gold standard” for quality.  But the debate has one inherent flaw, there’s no consensus on what constitutes “quality”.

“Central to these issues is the acceptance that there is no longer a single ‘gold standard’ measure of quality, such that the situation in which MT is deployed needs to be borne in mind, especially with respect to the expected ‘shelf-life’ of the translation itself.”

Source: Andy Way “Quality expectations of machine translation

The conundrum is what exactly constitutes “quality”, how is it measured, and who decides what is universal qualitative measure is acceptable?  As with all empirical matters, it is not surprising that this vigorous debate is still ongoing, fuelled by the growing use of NMT. Slator.com’s 2018 NMT shone a light on the debate and gave the views of different industry LSP players on this complicated subject.

The current discussion primarily focuses on the efficacy of technical quality testing using such standards as the BLEU score. BLEU is not seen by many as a suitable measuring standard. Its use simply followed because its existence pre-dated the emergence of NMT. Again, to quote Andy Way:

“When it comes to NMT, however, the improvements over predecessor MT— not to mention the differences in design (i.e. NMT usually runs on character-level encoder-decoder systems) — makes BLEU even less suited to quantifying output quality.”

Indeed, there are those in the industry who would declare BLEU has now become obsolete and should be replaced by measures such as BEER, chrF and characTER (see Slator.com).  These technologies all measure quality at a more granular level, unlike BLEU which measures at a sentence level. Then again, there are those who argue that the ultimate assessment of quality can only be done by a human: a linguistic, subject-matter expert. Underlying this belief is the assertion that there needs to be within the industry an accepted gold standard. An “absolute” standard that can only be achieved by a human. And is this “absolutist” stance that is the Achilles heel of their argument.

“The advent of machine translation and post-editing has focused attention on the very nature of quality: Is it proximity to a “gold standard” of perfection or is it characteristic of a product that simply serves its purpose well enough to satisfy the needs of the consumer? In other words, is quality something that should be measured and judged in absolute terms or in relative terms?”

Source: The Thorny Problem of Translation and Interpreting Quality: Geoffrey S. Koby, Kent State University, Kent, Ohio, USA; Isabel Lacruz, Kent State University, Kent, Ohio, USA

The same academicians state that the “absolute” stance asserts that some requirements are always understood and absolute and constant and can therefore remain unstated.  Whereas the “relative” stance asserts that in general, the best practice is to explicitly state all requirements as specifications, because they can vary from project to project. For both ends of the specification’s axis, some degree of accuracy and fluency is required. The difference is whether the minimum levels depend on audience or purpose or both.

Image Courtesy of: Pickering Electronics

It is in the last sentence above that the defines the dichotomy in the debate. Absolutist fly the gold standard flag and are unforgiving of less. The relative stance says, no – it’s the intended audience that defines the required quality standard. Their argument can be best explained in the following two examples:

1. You are a production manager in a law firm. You have been told you need a million words of Discovery Files to be translated from German in to English with two weeks. Only then can the firm begin to understand the evidence placed before them. As a production manager you could contact an LSP and ask them to provide an army of expert translators in law who can deal with the language pair. Good luck with that – and the cost. Or you could ask the question what are we trying to achieve with this translation?

The answer comes back that the lawyers need to have a “gist” of what the files contain, at that point they will decide which files are relevant and have them fully translated. The audience is not demanding a gold standard, they want a basic quality that will allow them to decide on the next translation move. This is undoubtedly a job for NMT in that this technology will satisfy the quality required, will do it within the narrow time frame and will do it without bankrupting the firm. Case closed: NMT wins hands down;

2. You are the production manager in a company that supplies mission critical technology. Perhaps a medical firm that supplies an IVD product that must be implanted successfully to ensure the survival of patients. The product’s instruction for use cannot have any ambiguity, or error. There needs to be zero errors (although, arguably not possible?). Certainly, this is not a case where you would use solely NMT to produce a raw (unedited) translation. However, you could use both NMT (because of volumes and time pressures) and a linguist/s with the subject-matter expertise. In seriousness, no-one would currently argue NMT alone should be used alone used. There would be a need for a strong human input. It’s a no brainer!

So those who argue that there is alone a gold standard are wrong in their argument. The standard is decided by the audience of the translation. If a company wants real-time translations it should be willing to accept a lesser standard of quality. More and more global companies are looking for this service, so much so LSPs are now embedding NMT in to their workflows with some PMs also finding their job spec changing as they become initiators and monitors of a constant flow of words being translated in real-time and displayed online. Alongside them you also have PMs working with the traditional human translators and/or human and machine translations.

The translation paradigm has changed dramatically in just a short number of years. At the end of the day, companies with an eye to their bottom lines will specified their quality standard. For them, this will be dictated by the audience they are aiming the translation towards. If it is an internal document and is large volume and time sensitive the customer will likely lower their quality needs and accept raw NMT. If the audience is more important, they might opt for both NMT and human editing.

Other companies, as outlined above, will declare they have a zero tolerance of errors and demand the human only track. One thing is clear, it is no longer a binary choice of human versus machine: the new paradigm sees the need for both. The production model is evolving as too are the job functions at all levels within the LSPs. From finance through to production, all will need to fully understand how the paradigm has shifted and restructure accordingly – to snooze is to lose.

Aidan Collins, Marketing Manager at KantanMT

If you had a very valuable and talented member of staff, would you send him/her on an errand to collect something from a supplier; a task that might tie them up for hours? Or, if you had a second option whereby you can request the supplier to call you when the product is ready for collection allowing your staff member to make a timely collection, which option might you take? Most people, I would venture, would choose the second option. However, many would favour having both options. Flexibility is always a good thing to have as part of any workflow.

With KantanMT’s announcement that its Software Development Kit (SDK) has been enhanced by the addition of a new Asynchronous Interface the community of KantanMT users has been given this very flexibility. This development in the SDK provides a high speed, high volume asynchronous programme interface in to both Statistical and Neural MT engines. This option complements the already existing synchronous interface option.

For the uninitiated, the synchronous and asynchronous nature of an Application Programming Interface (API) is a function of the time it takes from the initial request for a service to the return of processed data. In the case of synchronous APIs, the expectation is that there will be an immediate return of processed data. The API requests data and then pauses and waits for the processed data to be returned. This has the impact of tying up that hardware whilst it awaits a result to its request. Hence my analogy above of tying up a valuable resource while he/she waits for the supplier to supply the requested product.

In the case of asynchronous APIs, it proceeds to send data on the basis that a resource may not be immediately available, and therefore, the requested service may not be immediately responded to. The API will receive a call back when the required service is available. The asynchronous method avoids the server (and an engineer?) pausing and lying dormant while it awaits a response, thus allowing that server to be employed in other tasks.

Asynchronous processing is used in the case where the user doesn’t want to halt processing while an external request is being dealt with. A synchronous API will block the caller until it returns thus tying up resources until the action has been completed. On the other hand, an asynchronous API will not block the caller and typically will require a call back that will be executed once the work is completed. Asynchronous requests are useful in maintaining functionality in an application rather than tying up resources while waiting on a request.

The buying of hardware is a big capital investment for many companies, particularly start-ups. And staying on top of the biggest and the best is also an endless drain on company resources. Therefore, server time must be optimised to the nth degree. A tied-up server equates to a wasted resource – hence my analogy of the valuable staff member twiddling his/her thumbs while they queue for a collection.

The provision by KantanMT of the asynchronous API function allows users to reduce spending by optimising server time, so making less need for investment in multiple servers. KantanMT is confident that this addition to its SDK will be welcomed by many of its clients; users who while requiring high volume translations, do not need it done in real-time. The Asynchronous interface will provide a solution to any possible bottleneck in server capacity by queuing non-urgent translation requests to be processed when capacity is available. This, KantanMT is confident, will ensure users can avoid spending on costly hardware, thus reducing outlays and so keeping TCO (Total Cost of Ownership) to a minimum for the KantanMT Community. Having a bypass to a bottleneck is always a good thing to have, especially when it equates to less spending.

Aidan Collins is a language industry veteran. He is Marketing Manager at KantanMT.

As the localization industry strives at a fast pace to integrate Machine Translation into mainstream workflows to increase productivity, reduce cost and gain a competitive advantage, it’s worthwhile taking time to consider which type of Neural MT provides the best results in terms of translation quality and cost.

This is a question that has been occupying our minds here at KantanMT and eBay over the past several months. The fact is, Neural MT comes in many variants – with the different models available yielding remarkably different quality results.

Overview of Neural Network Types

 The main models of Neural MT are:

• Recurrent Neural Networks (RNNs) – these have been designed to recognize sequential characteristics of data and use the detected patterns to predict the next most likely sequence. Training happens in a both forward and backward direction; hence, the use of the descriptor recurrent. RNNs have been the predominant neural network of choice by most MT providers.

Fig 1: Image Courtesy of Jeremy Jordon

• Convolutional Neural Networks (CNNs) – these are the main type of networks used in computer image processing (e.g., facial recognition and image searching) but can also be used for machine translation purposes. The model exploits the 2D-structure of input data. Training process is simplified and CNNs require less computational overhead to compute models.

Fig 1: Image Courtesy of Jeremy Jordon

• Transformer Neural Networks (TNNs) – The predominant approach to MT is based on the recurrent/convolutional neural networks model of connecting the encoder and decoder through an attention mechanism. However, the Transformer Neural Networks model uses only the attention mechanisms aspect (e.g., contextual characteristics of input data). This completely avoids using the recurrence and convolution structures of the other models. This has the effect of simplifying the training process and reducing the computational requirements for TNN modelling.

Fig 3: Image Courtesy of “The Illustrated Transformer” by Jay Alammar

The eBay NMT Experiment

To determine which model yields the best translation outcomes, eBay and KantanMT collaborated and set up a controlled experiment using the KantanMT platform, which supports all three types of Neural Models.

The language arc English => Italian was chosen, and the domain defined as eBay’s Customer Support content. Each Kantan model variant was trained on identical training data sets which consisted of:

• eBay’s in-domain translation memory
• eBay’s glossaries and lists of brand names
• Supplementary KantanLibrary training corpora

The Test Reference Set was created by the eBay MT Linguistic Team by sampling the eBay Translation Memory to mirror its segment length distribution (e.g., 10% short segments, 30% medium and 60% long).

To provide a comprehensive comparison and ranking of the performance of different models, the translation outputs from the following systems were included in our joint experiment:

• Kantan TNN (Transformer Neural Network, customized)
• Kantan CNN (Convolutional Neural Network, customized)
• Kantan RNN (Recurrent Neural Network, customized)
• Bing Translate (Transformer Neural Network, generic)
• Google Translate (Transformer Neural Network, generic)

Human Translation (HT) was also included in this comparison and ranking to determine how neural machine translation outputs compare to translations provided by Professional Translators.

The evaluator was an eBay Italian MT language specialist with domain expertise and experience in ranking and assessing the quality of machine translation outputs.

The following Key Performance Indicators (KPIs) were chosen to determine the comparative fluency and adequacy of each system:

• Fluency = Fluency determines the translation follows common grammatical rules and contains expected word collocation. This KPI measures whether the machine translation segment is formed in the same way a human translation would
• Adequacy = Adequacy measures how much meaning is expressed in the machine translation segment. It measures whether the machine translation segment contains as much of the meaning as if it were translated by a human

Each KPI was rated on a 5-star scale, with 1 star being the lowest rating (i.e., No Fluency) and 5 stars being the highest rating (i.e., Human-Level Fluency).

KantanLQR was used to manage the assessment, randomise and anonymise the Test Reference Set, score the translation outputs, and collate the feedback from the eBay MT linguist.

The Results

Our Conclusions

The Custom Kantan Transformer Neural Network (Kantan TNN) performed the best in terms of Fluency and Adequacy. It outperformed RNNs in terms of Fluency by 9 percentage points (which is statistically significant), and 11 percentage points in terms of Adequacy. While there is still some way to go to achieve near-human-level quality (as depicted by the HT graphs), Transformer Neural Networks provide significant improvements in MT quality in terms of Fluency and Adequacy, and they offer the best-bang-for-your-buck in terms of training time and process simplification.

Since this blog was first published, comparative analysis has also been carried out for English=>German, English=>Spanish and English=>French language combinations and in all cases Kantan TNNs out-performed CNNs, RNNs, Google and Bing Translate.

What is Neural Machine Translation (NMT) all about?

Neural Machine Translation is an approach to machine translation that uses large neural networks to produce translations that are more natural sounding and achieve greater levels of fluency. These networks are trained on sequences (or sentences), which means they solve many of the syntactical and grammatical errors previously associated with Phrase-Based Statistical Machine Translation.

With the emergence of relatively cheap, super-powerful computers, coupled with practically unlimited storage capacity due to the emergence of “the Cloud”, we can now compute these complex NMT models in several hours.

These highly efficient models can then provide fast and fluent translations, and at an economically advantageous price. Today, over 90% of the daily traffic on the KantanMT platform is processed by our NMT services. This gives you an indication of the high regard our customers have for the efficacy of our NMT platform.

Why is there so much hype around NMT?

Simply because it provides a demonstrable improvement in machine translation outputs, resolving many of the problems associated with the statistical machine translation (SMT) model. Effectively, we have in the last two years solved many of the translation shortcomings of the SMT system; deficiencies that we’ve been trying to resolve for the last two decades! So, you can imagine how excited we are to be able to move with such speed compared to the pace of development we were lumbered with when working with SMT.

An interesting factoid about NMT is that we actually don’t train them using “whole words”, but we in fact train them using “word pieces”. And even if we don’t have parallel training data for a language combination, we can build a zero-shot network that will be capable of producing translations in these languages! It’s completely amazing what we can do once we set up the deep learning approach and throw super-computers at the problem. In Deep Learning (DL) we use highly sophisticated, multi-layered, pattern of ‘neurons’ to process huge chunks of data looking to refine the information contained within that data. The DL process can take an abstract jungle of information (word pieces), as is contained with data, and using the power of super-computation refine the data in to clearly understood language.

Can you imagine how good Neural MT will be in a further two years? Will Moore’s Law of exponential technological growth apply to Neural Machine Translation too? I believe it will. It will be amazing to see then how powerful NMT will be. It is certainly something that excites us here at KantanMT.com.

Which languages have made the greatest progress for NMT?

Any language that has a deep and complex grammatical structure can now be efficiently modelled using Deep Leaning and Neural Networks. For example, take the grammatical characteristic of the humble German verb – under normal circumstances it needs to be positioned at the end of a sentence. That would seem a straight forward enough challenge? However, SMT struggled to position the German verb accurately. To overcome this, we at KantanMT.com used advanced part-of-speech reordering approaches to improve this accuracy. This was a very complex, time-consuming and computationally intensive approach. However, NMT (because we train the engines on full sentences) almost always correctly positions that illusive German verb. This methodology also allows us to meet the challenges of languages such as Hungarian and Finnish. These are now well within our capabilities, allowing us to produce very good translation outputs using NMT.

Where do you see the translation industry in the next 5 years?

What an exciting time to be in the Localization Industry! We are on the cusp of a massive explosion in Artificial Intelligence (AI), which will impact all facets of the localization industry’s workflow and processes.

The industry will use automated translation technology to process even more content, in to more languages, and faster than ever before. Translators should not fear, as they will be the main beneficiary of this transformation. As this technological evolution grows, translators will be able to produce more words per day and consequently, significantly improve their income levels. I envisage a scenario whereby translation from scratch will be viewed as old-school and passé. The translation model will change in the same way as Computer Aided Translation (CAT) transformed the industry for the better. In the new NMT paradigm, the post-editing of a constantly improving machine translation output will be seen and accepted as the modern, progressive way of working. And the industry will be the better for it.

AI will also enable better job matching and candidate selection – so translators will be selected based on their relevant skill sets, domain knowledge and previous job performances. This is not to be feared, as essentially this is they way we choose our dentists and doctors today. AI wil

l become a driver for greater competition and increased professionalism in our industry.

I also see AI being becoming part of the project management workflow system, and the project management role.  PM systems will be expected to handle real-time translation workflows. A system that will combine automated translation and “human touch post-editing” to provide almost instantaneous results.

On the quality side, translation errors and problems will be identified by AI checkers and automatically routed for automatic recovery and fixing. The time between job arrival and completion will reduced in some cases to seconds. These “micro-jobs” will be driven by the requirement for new content to be translated in effectively “real time”. This fast system will be required for content such as blogs, wikis, live user forums, reviews, internal corporate content, help chat lines etc.

What should we expect from KantanMT in the next few months?

We’re working on a new type of Neural Network that will provide even better translation outputs than before, with a significantly reduced training time. These new networks are already in testing with one of the largest eCommerce companies; so, stay tuned for further news of this major step forward in the evolution of NMT.

Additionally, we have figured out a way of measuring the quality of an automatically generated translation. This Quality Estimation Score system was developed by KantanMT.com for the European Commission. The good news is, we shall be open-sourcing this technology in early 2019.

You’re also going to also see a new, improved version of KantanLQR that will support multi-lingual quality projects. It will give you the means to measure how individual language arcs are performing across your enterprise.

This article first appeared on http://www.argosmulitlingual.com in October 2018: http://www.argosmultilingual.com/blog/shaping-the-path-to-neural-machine-translation-with-tony-odowd

Today’s blog is aimed at helping the novice understand the technology that is Deep Learning (DL). To do this, I will need to discuss in-depth Linear Algebra, Statistics, Probability Theory and Multivariate Calculus. Only Joking! Nothing would turn the novice readers off than trying to hack our way through the above complex disciplines. We’ll leave that for the nerds. Today’s blog – like my last on Machine Learning – will try and use an analogy to help explain what is without doubt a very multifaceted, intricate subject to fully master.

For myself, the more I read about Deep Learning, and the more I spoke to the engineering masterminds at KantanMT the more I realised that the discipline of using a Deep Learning model bore a similarity to sculpting. Let me expand: I don’t know to whom this quote is attributed, but for me it certainly describes the methods of Deep Learning:

“The sculpture produces the beautiful statute by chipping such parts of the marble block as are not needed – it is a process of elimination.”

Indeed, I think it was no less than Michelangelo, who when asked about sculpting, said that the angel lay within the marble block; it was simply his job to release it. Michelangelo’s minimalist explanation, and the above quotation, encapsulate in its simplest form what the Deep Learning progression involves. The engineer is the sculpture. The marble block represents the huge block of dense data to be processed.  The act of processing the data is the chipping away of unwanted information by neural networks. The act of fine tuning the deep learning neural engine represents the technique of the sculptor carefully finessing the shape of the emerging form in to a recognisable figure.

In both the role of sculptor and engineer there is a vision of what the ‘fine-tuning’ activity should produce. I am confident that if you as a novice accept this simple analogy you are going someway to grasping the very fundamentals of the Deep Learning process.

As a concept, Deep Learning is less than two decades old. The origin of the expression is attributed to Igor Aizenberg, Professor and Chair of the Department of Computer Science at Manhattan College, New York. Aizenberg studies, amongst other things, complex-valued neural networks.  He came up with the concept of an Artificial Neural Network system based on that of the human neural network – the network of the human brain.

The ‘Deep’ element of the concept refers to a multi-layered processing network of neuron filters. The equivalent process in the human brain is that of information flowing through neurons connected by synapses. In the machine equivalent, artificial neurons are used to fine-tune and refine data as it is passed through the ‘engine’. The process of Deep Learning also learns from experience and can adjust its processes accordingly. In sculpting, it is the equivalent of the experienced sculptor chipping and refining the marble to release Michelangelo’s hidden angel.

Jeff Dean, a Senior Fellow at Google’s ‘System and Information Group’ – the group behind many of Google’s highly sophisticated machine learning technologies – said:

“When you hear the term ‘Deep Learning’ just think of a large neural net. Deep refers to the number of layers typically, and so this is kind of the popular term that’s been adopted by the press.”

For many novices there is a confusion around the terms Machine Learning (ML), Artificial Intelligence (AI) and Deep Learning (DL).  There need not be this confusion as the division is quite simple: Artificial Intelligence is the catch-all term to cover Machine Learning and Deep Learning. Machine Learning is an over-arching term for the training of computers, using algorithms, to parse data, learn from it and make informed decisions based on the accrued learning. Examples of machine learning in action is Netflix showing you what you might want to watch next. Or Amazon suggesting books you might want to buy. These suggestions are the outcome of these companies using ML technology to monitor and build preferences profiles based on your buying patterns.

Deep Learning is a subset of ML. It uses a highly sophisticated, multi-layered, pattern of ‘neurons’ to process huge chunks of data looking to refine the information contained within that data. It takes an abstract jungle of information, as is contained with data, and refines these in to clearly understood concepts. The data used can be clean, or not clean. Clean data is the processing of refining the pre-processed information to remove any clearly irrelevant information. Clean data can be processed quicker than data that has not been cleaned. Think of it as the human brain blocking out extraneous information as it processes what is relevant, and discards what is irrelevant. Something the human brain does every minute of every day.

But why has Deep Learning suddenly taken off so spectacularly? It is because of the ability to train Artificial Neural Networks (ANN) to a level of accuracy when trained with huge amount of data. ANN can synthesise complex non-linear processes with a high degree of accuracy. DL is also becoming predominant because of the following boosters:

• The emergence of Big Data
• The increase in computational power
• The emergence of The Cloud
• The affordable availability of GPU and TPU
• The development of DL models using open source code

Today it is estimated that Big Data provides 2.5 quintillion bytes of information per day. Now, if you are like me, you’ll will never have heard of the measure quintillion. Well, apparently, it is 1 million billion. Not that helps give it finer focus!

According to IBM:

“90% of the data in the world today has been created in the last two years. This data comes from everywhere: sensors used to gather shopper information, posts to social media sites, digital pictures and videos, purchase transaction, and cell phone GPS signals to name a few. This data is big data.”

It is safe to say that the amount of data available will only increase over the coming years. Institutions such as the European Union, the United Nations, the World Bank, the World Health Organisation, Social Media companies etc make huge volumes of data available daily, and in multilingual form. The importance of this resource of massive data is underlined by Andrew Ng, Chief Scientist at Baidu, China’s major search engine, who said:

“The analogy to deep learning is that the rocket engine is the deep learning models and the fuel is the huge amounts of data we can feed to these algorithms.”

The advent of Cloud Computing has allowed even small companies to have virtually unlimited storage space, and access to fantastically powerful computational power. Processors of the power of tensor processing unit (TPU) are available via Cloud computing. Some examples of Cloud computing sources would be Amazon’s Web Service, IBM’s SmartCloud or Google’s Cloud.

TPUs were developed by Google to specifically deal with the demands of ANN. Previously, graphics processing unit (GPUs) reduced from weeks to hours the machine learning process. TPUs have speeded up that process exponentially. Without this level of computing power, it is unlikely Deep Learning would be a viable technology.

Finally, Intel is reportedly developing a device called a Neural Stick which they claim will allow companies to bypass the Cloud to do their processing at a local level (i.e. non-Cloud level). This will be a boost to those companies who baulk at the security implications of processing data in a remote location. It will also increase the speed of processing as all the crunching will be done at the local level. Intel say it is their intent to make DL work “everywhere and on every device”. If they succeed, Deep Learning will expand to a huge degree. Interesting times lie ahead for Artificial Intelligence.

Aidan Collins is a language industry veteran. He is Marketing Manager at KantanMT. This article first appeared in Multilingual in the December 2017 edition: https://multilingual.com/all-articles/?art_id=2592