Slate approaches every translation segment similarly to your work with difficult translation segments. Let’s look at how you work on a difficult translation. Then, we’ll see how Slate uses the same principles with its every suggestion. Finally, we’ll explore how this helps you improve Slate’s performance.
Your Most Challenging Translations
Let’s face it. When you do the work, some segments are more difficult to translate than others. Some are difficult because you can’t find the right words or grammar or syntax. Your research skills kick in. You have a breakthrough and move on to the next segment, but what do you do when draft two, three or more candidates that express equally appropriate translations? How do you you decide which one send to your customer?
Knowledge Goes Only So Far
Like other difficult segments, you start with your knowledge. Maybe you expand your knowledge. You resort to dictionaries, glossaries and grammar books. You research every resource you know and then some. You ask your colleagues and Google.
Despite your exhaustive research, you still have three perfectly valid and equally appropriate candidates.
How do you choose the final translation when you have multiple target language candidates that express the same meaning, tone and subtext as the original source text?
Going The Distance
It’s difficult to see at first, but the solution is easier than the process that got you to this point. You choose the one that feels right. The mechanics might vary from translator to translator. Maybe you close your eyes and imagine the target sentence in a real-world context. Maybe you chant the choices and listen to how they sound. The point is, at this late stage, you ignore the source language segment and choose the target language because it feels right.
Slate’s Everyday Processes
Slate doesn’t have feelings. It can’t use its knowledge for some segments and “feelings” for others when it translates. Instead, it has an SMT model with two sub-models, the translation model (a.k.a. the front-end) and the language model (a.k.a. the back-end). They work together mimicking your actions to generate the suggestions you see.
Mimicking Your Knowledge
Slate’s front-end uses a large knowledge database that it created from your translation memories during the build process The database, called the translation model, is a bilingual concordances with source-target fragment pairs of 1 word, 2 words, 3 words… through 9 words. Although other concordance databases have only words groupings that humans understand, the translation model has every possible combination of fragments.
The front-end breaks the source segment into all possible fragments of 1 word, 2 words… etc. It searches the database for the corresponding target fragments. From that knowledge, the front-end assembles the target fragments into upwards of 10,000 or 20,000 candidate translations. Clearly, the vast majority of these have the wrong vocabulary and word order. One or two or three, however, can be exactly what you’re looking for, and Slate needs to find it. The problem? Computers doesn’t have feelings.
Mimicking Your Style
Slate’s back-end uses a second database that it created from your translation memories during the build. This database, called the language model, is a monolingual concordance with only target language fragments of 2 words, 3 words… up to 7 words and the frequencies they occur in your translation memory.
The back-end runs the same process for each of the thousands of target candidates it received from the front-end. It breaks the candidate into all possible fragments of 2 words, 3 words… etc. It searches the language model database for the corresponding frequencies. From those numbers, it gives a score to each target candidate based on the the most frequently occurring fragments and the most frequently occurring order. Slate finally presents the candidate with the highest score to you as its final suggestion.
Slate’s back-end process doesn’t directly mimic your feelings, but let’s take a take a closer look. Your feelings are rooted in a sense of familiarity. That is, we feel comfortable with the familiar. Our familiarity comes from frequency of your experiences over a long time. Frequency drives our feelings. In the end, Slate and you pick your final translations based on the frequency of experiencing the target words and their order.
Improving Slate’s Performance
With this understanding, it’s easy for you to control Slate‘s performance because Slate converts your translation memories to the translation model and language model databases, and Slate uses them to create suggestions.
With Slate, you can create and use a wide variety of translation engines from legacy machine translation to customized.
A personalized translation engine is a customized engine that drafts personalized translations. Slate learns what you expect by studying your translation memories and the engine mimics the terminology and style you expect.
This is possible because Slate runs on your personal computer to serve only you. With a personalized translation engine, you finish more work in less time without customers or competitors spying on you through the cloud.
You can experience how much faster you can work in 3 easy steps during your risk-free 30-day trial money-back guarantee.