Our Analysts' Insights

The explosion of enthusiasm for MT that followed Google’s announcements about neural machine translation (NMT) last year shows no signs of abating. CSA Research has seen a few trends in our observation of new entrants and existing players:

• Security is increasingly important. Issues around security and where data is housed are becoming more important for enterprise users. Many companies face legal restrictions on data disclosure to third parties or have concerns about data leakage to online engines. CSA Research found that data security is the second most-important concern of implementers – after quality – and awareness is growing around the topic.
   
• NMT is here to stay. Far from being a short-term trend, MT development is consolidating around neural systems at an accelerated rate. Although it isn’t ready to replace older statistical systems such as Moses in all contexts, it is clear that industry momentum is headed to NMT, and industry pioneer Systran has bet its future on this approach.
   
• Rules are dead: Long live rules! For much of the last decade it appeared as if rule-based MT (RbMT) systems were likely to disappear. But as the discussion below will show, they still have their place. In addition, NMT developers are increasingly turning to systems that utilize rule-based components to improve their output. The way forward appears to be one that uses pragmatic combinations of neural, statistical, and rule-based approaches, combined with the input of skilled computational linguists helping some providers stand out from the masses.

As various companies have either jumped on the MT bandwagon or repositioned their offerings to meet the current demand, we found a few recent standouts from the crowd that highlight some of these trends.

Omniscien Relies on Complex Pipelines
Formerly Asia Online, Omniscien Technologies launched its own NMT services as part of Language Studio 4.0 in April, 2017. CTO Dion Wiggins briefed us on the difficulties of NMT and how his company has addressed them. NMT is notoriously resource-hungry, and early generations required farms of specialized servers based on graphical processing units (GPUs), which happen to be very efficient at processing neural networks as well as images. Omniscien worked to optimize the code behind the networks and reduce the hardware footprint. The company’s internal benchmarks show a throughput increase of almost 1400% percent per GPU since development began. The result is an economical system that can scale to meet a broad range of volumes of demand.

Omniscien had long stood out from the crowd because of its sophisticated pre- and post-processing pipelines. Users could customize these themselves or leave development to the supplier. These processes can correct common errors in the source, reorder content to facilitate MT, and make programmatic changes. In effect, they make Language Studio a hybrid system that combines the strengths of NMT, robust rules, and statistical components in order to address the limitations of each one.

According to Wiggins, the customer-specific capabilities of this hybrid are particularly useful in translating e-commerce text, which often consists of unstructured text and ungrammatical strings that serve primarily to make content discoverable for search engines. With the move to NMT, the company had to ensure that changing engines would not break these pipelines. The new system functions as a drop-in replacement for its older systems, which remain available to customers.

The net result is that Omniscien has successfully implemented a neural-statistical hybrid – with accompanying quality increases – in one of the most complex MT workflows on the market. Its offerings should appeal to enterprises that want to exercise maximum control and that can invest in developing processes designed to improve output.


Yaraku Delivers Adaptive NMT for the Asian Market
We first encountered this relative newcomer in June, but Japanese MT developer Yaraku stands out as the first company we know of that offers adaptive machine translation – a critical component in the augmented translation stack – for Japanese in its YarakuZen system. This technology learns in real time from observations about how users correct its output or translate text on their own in order to produce results more like their output. Yaraku is also one of the first to combine adaptive workflows with NMT, using a multi-step process that chops saved data into phrases to create rule-based corrections to the underlying neural layer. This approach solves some of the processing difficulties that have stymied efforts that focus on retraining neural systems in real time. This solution is currently available for Japanese and English, with Chinese and other Asian languages slated for release in the near future.

Source: Yaraku

The company focuses on an unusual niche: It is aiming for the mid-point between free online services that make too many mistakes and LSP-driven services that promise full human quality at a relatively high price point. Although LSPs and professional translators can use it – and Yaraku counts some of Japan’s largest translation providers among its clients – the primary audience consists of businesspeople who know some English but who need help with their communication. Members of the this group can use it to assist them to produce e-mails or documents in their non-native languages.

The product includes collaboration abilities and – if individual users decide that the results need more attention – the click-button ability to order professional post-editing of the results. In addition, the company offers APIs to get MT as a call-back and to order professional translation. With the 2020 Tokyo Olympics fast approaching, CSA Research expects Yaraku to get a real work-out as demand for language services increases.

SDL Moves to Neural in ETS
In June SDL officially joined the NMT club, with a new version of SDL Enterprise Translation Server (ETS) that added neural systems to its existing statistical MT (SMT) lineup. As was the case with Omniscien, SDL worked to ensure that the shift was a drop-in for the previous approach and would not disrupt its existing users. SDL reports significant improvements both in the BLEU scores favored in research and in the subjective assessment of users. When CSA Research reviewed sample English-German translations from the new system versus the SMT version, they were clearly more natural sounding and demonstrated the typical improvements seen from moving to NMT.
 

Source: SDL

A significant portion of SDL’s customer base relies on on-premise machine translation solutions, which raises the performance bar. Most NMT systems take a brute-force approach that throws lots of computing power at the problem, but SDL wanted a scalable solution that could run on relatively modest hardware. It spent many months optimizing code so its NMT software could run on a typical enterprise server setup. It claims that ETS can run on single servers and scale up to support high-volume requirements. For buyers that do not require in-house set-ups, SDL also offers secure cloud-based solutions.

Blast from the Past: ULG Adopts Rule-Based MT
Even as the other three companies discussed in this post are moving rapidly into the NMT frontier, United Language Group (ULG) has shown that established technologies still have a role to play, with its acquisition in March 2017 of Lucy Software. Lucy – an established, if somewhat niche, player – long offered highly-tuned rule-based MT (RbMT) and statistical solutions prior to the acquisition, but had never quite managed to break through to a larger audience. However, its strengths should make it a strong contender now that it is part of ULG.

RbMT offers advantages in cases where the source language is relatively controlled and proper terminology data is available because it can run on very modest hardware while still delivering high throughput and accurate results. It can also be easily trained to preserve document formatting – an area where statistical systems have struggled – to reduce the need for costly DTP. On July 27, ULG announced that it was using the acquired technology in a way that plays directly to the strengths of RbMT approaches, by offering a secure and efficient proxy-server-based translation option that preserves native document formatting. The company has also combined it with optical character recognition and language identification components to handle the translation of scanned documents. Such support emphasizes the role that Lucy’s components play in the delivery of secure, high-performance options in this area.

Conclusion: MT Solutions Are Blooming
Enthusiasm for machine translation tends to come in waves every few years. It is apparent that the current swell – which corresponds to a tremendous hype cycle for artificial intelligence in general – is a major one that corresponds to real change. Even though CSA Research maintains that much of the current media coverage promotes unrealistic visions about what MT can accomplish, the seemingly weekly emergence of new or revamped solutions shows just how rapid this period is in comparison to the incremental change that had characterized the field for the past decade.

As these and other solutions develop and find their place, expect to see an expansion of use of both raw MT and augmented services that combine the advantages of humans and machines to deliver higher quality at a lower cost. And given the increased suitability of machine-translated output and the proliferation of solutions at multiple price points, CSA Research envisages more and more solutions in the coming months that address needs the world doesn’t even know it has. Each of the four offerings discussed above helps address part of the overall puzzle and moves into areas where the alternative had previously been limited or zero translation.