OPUS-CAT is a collection of software which make it possible to OPUS-MT neural machine translation models in professional translation. OPUS-CAT includes a local offline MT engine and a collection of CAT tool plugins.
OPUS-CAT MT Engine is a Windows-based machine translation system built on the Marian NMT framework. OPUS-CAT MT Engine makes it possible to use a large selection of advanced neural machine translation models natively on Windows computers. The primary purpose of OPUS-CAT Engine is to provide professional translators local, secure, and confidential neural machine translation in computer-assisted translation tools (CAT tools), which are usually Windows-based. To that end, there are plugins available for two of the most popular CAT tools, SDL Trados Studio and memoQ (OPUS-CAT can also be used in the Wordfast CAT tool as a custom provider). OPUS-CAT MT Engine provides pretrained MT models for a very wide selection of language pairs, courtesy of the OPUS MT project (listing of OPUS MT models).
NOTE: Fiskmo MT Engine application needs to be running on the computer, when the plugins are used.
Fiskmö Trados 2019 plugin includes a functionality for fine-tuning models with bilingual material that is specific to a domain or style of translation.
The base models that can be downloaded from the OPUS model repository generally produce useful machine translations. However, since the base models have been trained on a mix of different bilingual texts from various textual domains and styles, the translations may not be suitable for many contexts. In a professional translation setting in particular, there is an expectation of terminological and phraseological consistency within documents and adherence to style guides and glossaries.
By fine-tuning a model with suitable fine-tuning material (bilingual texts from the relevant domain), a base model can be modified to produce translations that adhere to the conventions used in the fine-tuning material.
The fine-tuning functionality in the Trados 2019 plugin is implemented as a custom batch task. The batch task can be invoked via the Batch Tasks button on the Home ribbon (or via context menus):
The fine-tune batch task can also be used to pregenerate the translations for the project. Note that this means pregenerating the translations in the cache of the Fiskmö MT engine so that they will be instantly available as matches when translating in Studio, the translations will not be copied in the sdlxliff files by the custom task. You may choose to fine-tune and translate, fine-tune only, or translate only (the model will have to be selected in that case). The different modes are shown in the following picture:
Fine-tuned models are identified by model tags (see the Model tag box in the picture above), which usually describe the model. Fine-tuning settings can defined more specifically in the Fine-tuning settings tab:
Generally the most useful fine-tuning material is the translated text that is already present in the project (i.e. exact match segments). These segments will always be extracted for fine-tuning. However, you can also extract fuzzy matches from the TM to use as fine-tuning material, which is especially important in cases where there's not enough translated text present in the project. You can set the minimum fuzzy percentage and the maximum amount of fuzzies to extract for each segment in the Fine-tuning settings tab.
When you are ready to start fine-tuning, click Finish in the Trados Batch Processing screen. Extracting the fine-tuning material may take a while, depending on project and TM size. Once the fine-tuning material has been extracted, the batch task will send it over to the Fiskmo MT engine running on the same computer in order to start the fine-tuning.
The fine-tuned model will be shown in the local model list with a status of Customizing, and the text Fine-tuning in progress is displayed in the bottom of the MT engine screen. The fine-tuning will generally take several hours, and once it's finished, the status of model will change to OK and the text at the bottom will disappear.
You can use a fine-tuned model for translation by selecting the tag of the model in the settings of the Fiskmo translation provider in Trados:
If the fine-tuning is still in progress for the model with chosen tag, the connection status message will state so.
If the project contains a lot tags, it's possible to include textual representations of tags into the fine-tuning material, in which case the fine-tuned model learns to transfer tags from the source text into the target text. The tag learning feature is experimental, so it might not work optimally in all cases (the fine-tuning material needs to contain enough tags to learn from, and the tag structure should be relatively simple). You can choose to learn placeholder tag positions (Include placeholder tags as text) and tag pairs (Include tag pairs as text).
Models are normally installed in the MT engine with the Install model from Web functionality, but they can also be installed from files. This is useful for distributing customized models and for installing models in systems where no Internet connections are possible. To install a zipped model, click Install model from zip, browse to the location of the model zip in the file system, select the model zip, and click Open.
Right-click the file and choose Properties.
Check the Unblock box in the bottom right and click OK.
The path of the installation folder varies in different memoQ versions, for memoQ 9 the default installation folder is C:\Program Files\memoQ\memoQ-9.
Here's the easiest way of locating the memoQ installation folder:
Once the FiskmoMTPlugin.dll file has been copied to the Addins folder, the plugin should be loaded when memoQ is started. As the Fiskmo plugin for memoQ is an unsigned plugin, you need to confirm that it can be loaded with memoQ. Select Yes in the dialog box:
MT is used in memoQ by first defining a set of MT settings and then attaching those settings to a project. You can create a set of MT settings by performing the following steps:
To use MT during translation, you need to add the MT settings that you created in the previous section to the translation project. You can do this performing the following steps:
Often it's best to pretranslate documents with the Fiskmo MT plugin before starting translation, since this means that the translations are available immediately (when translating segment by segment in the editor, generating the MT may take some time).
You can pretranslate documents with the Fiskmo MT plugin in memoQ by performing the following steps:
It's possible to obtain translations from the Fiskmo MT engine via an HTTP API. Note that the HTTP API is not enabled by default, as doing so requires system administrator permissions. To enable the HTTP API, open the command prompt as an administrator and use the netsh utility:
netsh http add urlacl url=http://+:8500/ user=DOMAIN\user
(Substitute your domain and username for DOMAIN\user. 8500 is the port number used by the HTTP API.)
Translations can be fetched from a local Fiskmo MT engine instance with requests using the following format:
A list of available language pairs can be fetched with requests using the following format:
http://localhost:8500/MTRestService/ListSupportedLanguagePairs
This work is part of the fiskmö project. It implements a self-contained MT plugin for SDL Trados Studio that runs a translation engine based on MarianNMT locally within the plugin. Please acknowledge the project if you use our tools and resources.
Source:
git clone https://github.com/Helsinki-NLP/fiskmo-trados.git
Copyright 2019 Tommi Nieminen
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.