Customer Insights – MT in Action
If you read our last blog post, then you’ll know all about the trials and tribulations of using unedited machine output. In this article, we’ll show you how you can leverage MT to your advantage. We asked two of our major clients – a pioneering company specialized in talent management software and a world leader in contact center solutions – about how they make the most of MT to boost cost efficiency while still producing high-quality content.
What challenges do companies face in localizing their content today?
In the era of big data, many companies are struggling to handle massive content volumes. This problem is exacerbated when this material needs to be localized. When it comes to our contact center solution provider, international users want its online software documentation to be made available in their mother tongue. But this presents the company with a dilemma – in dealing with this higher demand for localization, it has to ensure that costs don’t skyrocket or that the quality of this content doesn’t deteriorate. This is precisely where MT enters the picture.
How do our clients rate the quality of MT? Does it meet their quality requirements?
MT has become the latest buzzword in the world of localization, with many companies being swept away by incredible cost savings of up to 50 per cent. But as we’ve said before, using raw machine output that hasn’t been revised by a human expert can backfire terribly – with disastrous results. This is something that our clients also pointed to. While conventional language combinations like English to Spanish initially appear to work quite well with MT, the output produced tends to be quite crude and unsuitable for immediate use. They outlined the following problems in particular:
- Fluency: Machine translations are notorious for being difficult to follow, often irritating users in the process. Individual sentences may not scan well, or pieces of information may not be linked properly.
- Readability: One of our clients noted that MT doesn’t render the correct word order necessary to make a text read well. For example, if three compound nouns are used in one sentence, the machine output turns out to be incredibly clunky, thus impacting user-friendliness.
- Formatting issues: Document types with special formatting elements like XML and PowerPoint don’t interact well with MT, often resulting in word-for-word translations that are almost identical to the original text, misplaced line breaks, or other formatting complications.
- Coherence & cohesion: Machines obviously don’t possess the same text comprehension abilities as humans, which means that grammatical constructions like gender are not handled particularly well. For example, while objects are simply referred to as “it” in English, this “it” could be “sie”, “er”, or “es” in German.
Our client from the contact center industry also pointed out that one way of improving MT quality could be to have texts written in line with clearly defined standards so that they are suitable for localization later on. After all, if an original text is poorly written, then we can’t expect a machine to interpret it correctly and produce high-quality output. This approach could help to further improve MT content in the future.
So where and when does it make sense to use MT?
Both companies employ machine translation for text types like product documentation, help content, and software UI, reflecting an overall trend towards using MT for large content volumes that otherwise would have been too expensive if purely human translation were used. What’s important to note here is that this content tends to be technical – meaning that comprehensibility takes precedence over stylistic perfection. For example, it’s ideal for texts like technical instructions, as the language used here doesn’t need to be particularly polished.
How does machine translation hold up against conventional translation?
Our client from the field of talent management handles huge swathes of online help documentation that needs to be localized. A traditional translation workflow was out of the question for cost reasons, while even conventional post-editing was too expensive for its enormous content volume. This meant its localization managers found themselves in balancing act – caught between the demands of upper management to keep costs down, and end users such as international customers who needed localized software instructions. Not only that, but the internal sales team also wanted high-quality material to help it close deals.
While raw MT proved unbeatable in terms of cost efficiency, it soon transpired that it would be too risky to publish this content without human revision.
After an in-depth consultation with Milengo and extensive testing, the client opted for a particularly innovative and pragmatic solution in the form of a neural translation system combined with a human quality assurance step, which centers around ensuring that the software UI is correct and pre-defined key terminology is used. This has allowed the client to significantly reduce translation costs for product documentation, while also upholding essential quality criteria. Managed MT – as Milengo dubbed the solution – has since proved itself to be a gamechanger when it comes to making the company’s product documentation available in other languages.
Finding the perfect balance with MT
Exponentially growing volumes of content paired with high cost pressure mean that we’ll be seeing a lot more of MT. By mitigating the pitfalls of MT and including a human quality assurance step, both of our clients are now able to enjoy the best of low costs and high quality. The shape that this revision step takes frequently depends on the text type, as printing instructions for a technical device are sure to warrant higher stylistic requirements than a troubleshooting article consigned to the depths of a knowledge base. No matter the approach, one thing’s for certain: machine-supported solutions that strike the perfect balance between quality and cost requirements are the future of the localization industry.