Quality estimation
Assess the quality of your AI translations even before post-editing.
MTQE (Machine Translation Quality Estimation) allows the quality of machine-generated translations to be assessed even before post-editing. Traffic-light systems and quality scores indicate which segments require closer scrutiny or further action. Quality estimation therefore delivers rapid results and enables more efficient post-editing processes.
As QE does not require a human translation as a reference, it is also suitable for new translations and all language directions.
oneword has extensive experience with various QE systems and integrates them seamlessly into your translation processes. We also tap into the QE results for optimisation via custom prompts and for automated preparatory work in the form of automatic post-editing. This enables us to get the most out of your AI translations for you.
Our certifications:
Why quality estimation is worthwhile for your business
More efficient post-editing with professional quality estimation
Quality estimation can specifically reduce the amount of post-editing required for machine translations. We regard QE as one component of the MTPE process, which supports – but is not the sole basis for – the evaluation of translation output. Because even though a machine can, in principle, evaluate a machine-generated result, we rely on people to derive meaningful next steps from these results.
In the process, we focus on:
- The best QE system: we regularly test QE systems to ensure we are always working with the best possible results.
- Targeted prompting: any QE system is only as good as the prompts used to instruct it. We incorporate your requirements into the evaluation process so that, in addition to general language quality, company-specific factors are also taken into account.
- Human control: we use QE results as the basis for efficient translation processes, but never as a substitute for human expertise. Every AI project undergoes post-editing, which can be carried out more effectively and quickly using QE.
This enables us to integrate quality estimation specifically into your MTPE process and reduce the amount of post-editing required.
When and how we use quality estimation
QE can be used at various stages of the AI translation process:
1. We estimate the post-editing workload
Our assessment shows in advance, at segment level, which segments are likely to require editing and how many there are.
2. We identify sources of error
QE can quickly identify problematic segments. Some of the mistakes detected can then be corrected using automatic post-editing (APE).
3. We check against company-specific requirements
The prompts you have saved determine which aspects the QE system evaluates. Through targeted custom prompts, we take individual requirements into account and assess whether these have been adhered to.
Segment by segment rather than an overall score
With QE, we adopt a segment-by-segment approach to quality assurance as the quality of individual segments within a document can vary significantly and even the smallest mistake can make a key difference.
A summary assessment of an entire document is not meaningful and can even be risky: even a good overall document score may still contain segments of very poor quality and critical mistakes.
Would you like to learn more about quality estimation? You can find out more by speaking to one of our experts in person.
FAQs
Does quality estimation work for every language?
Quality estimation is based on neural networks and can therefore be used for any language combinations. You do not need reference translations or separate training for each language.
How reliable is machine translation quality estimation (MTQE)?
The quality of the evaluation depends on the system used and the prompts deployed. General QE primarily takes language quality into account, but does not consider company-specific characteristics. With targeted QE prompts, you can also check individual specifications and tailor the evaluation to your requirements.
Do segments rated as good no longer need to be checked by a human?
Whether a human review is required for segments rated as good depends on factors such as the level of risk, domain and usage scenario. After all, one machine is evaluating the output of another machine and may itself make mistakes in the process. However, with QE you can easily identify segments containing mistakes and then edit them.
How much does quality estimation cost?
The amount of work and cost involved depends, amongst other things, on the LLM used, the creation of company-specific QE prompts and the feedback loops we use to check the reliability of the evaluation. QE results can support your post-editing work and help you exclude segments containing no mistakes from projects, thereby cutting costs.
Our AI translation services.
Let’s discuss your next project!
Please use our contact form or give us a call:
Make an appointment now:
