What is custom training/custom trained model?

The process of providing language assets (bilingual data) as teaching data to a machine translation (MT) learning service for machine learning is called "Custom training" (sometimes called "Adaptation" or "Relearning"). As the result of custom training, a custom trained model is created. Once the custom training has been completed, the custom trained model can be used in Quick MT/PE and LAC (Custom Glossary Model Registration) of this system.

Information in Custom Trained Model Registration screen

In this screen, you can register a new custom trained model. The general steps are as follows:

1. Select a MT learning service for which you want to perform the custom training. With this system, you can use the following three MT learning services.
   
   • Google AutoML
   • Kawamura NMT
2. Select the language assets already registered in LAC as the teaching data for custom training. Press [Language Asset Selection] button to display a selection window. You can specify criteria to search for language assets that have been imported. Select the check box for the language assets you want to use and press [OK] button. You can select multiple language assets.
3. The main screen displays a list of the selected language assets. Please set necessary items according to the MT learning service you selected (described later).  *If there is any language assets you don't need to include into the custom learning, you can delete them by checking the left-most checkbox and clicking Trash can icon.
4. If you want to remove the inline tags from the language assets during the custom training, please check [Remove inline tags from Language Asset data] check box. It is checked as default. If you uncheck it, data (source texts/target texts) including the inline tags will be processed as the corpus, which may reduce the accuracy of the custom training.
5. When all settings are completed, click [Custom Training Request] button. The custom training is scheduled to run and will appear in Custom Trained Model list. Registration of the custom trained models may take hours to days to complete, depending on the size of the provided corpus data, load on MT learning services, and so on. You will be notified via email as soon as the custom training model has been registered.

Custom trained model registration settings

Depending on the selected MT learning service, you will need to set/select the items marked "〇" in the table below. Please refer to the explanation below the table for each item.

• Type: Type of Kawamura NMT (described later)
• Training Method: You can choose from the following three methods;
  
  1. Adaptation + EBMT: Creates a hybrid translation model by combining the methods 2 and 3 below. It is recommended when using language assets of 10,000 lines or more without noise (mistranslation or miscorrespondence).
  2. Adaptation: A standard method. Executes the custom training to the base model using the provided language assets. It is expected to improve the accuracy of machine translation for sentences that are similar to the training data, but the accuracy of machine translation for sentences that are not similar to the training data may be lower than the base model. It is recommended when using with noisy language assets of 10,000 or more lines.
  3. EBMT (Example-based Machine Translation): If there is a sentence in the training data that is similar to the input sentence (source text for the machine translation), translates referring to it. If there is no similar sentence to the input sentence, translates referring to the base model. It is recommended when using noise-free language assets of less than 10,000 lines.
• Setting: There are the following four settings. Please note that the following requirements depend on the MT learning service you have selected.
  
  • Automatic (Google AutoML only): The specified language assets are automatically split and used for the following three purposes accordingly.
  • Training: Used for actual custom training.
  • Tuning: Used to prevent overlearning. (This setting can be called 'Validation', 'Adjustment' or some other names depending on TM learning service.)
  • Testing : Used to evaluate generalization performance.

• Reverse Language: If you want to use the language asset by reversing the language direction, check the check box. If you check the box, for example, language assets with the language direction of English-Japanese (en->ja) will be used as corpus data for Japanese-English (ja->en). Note that the check box will be checked by default only for language assets whose import category is "TM Generation (AI Utilization)". This is based on a study finding that it is more effective to give machine-translated text as the source text instead of the target text. In other words, if you machine translate Japanese into English and create a Japanese-English TM, it will be more effective to use that TM for custom training for English-Japanese. If you do not wish to reverse the source and target text, please uncheck the check box.

TIPS: How to execute the training more effectively with less corpus data (smaller language assets)?
We have found that selecting "EBMT" as the training method using "Kawamura NMT" can execute custom training to language assets with fewer rows effectively. Even when you create a model by specifying the same language assets for "Training", "Tuning" and "Testing", you can also customize your training model.

TIPS: How do I execute custom training when using noisy bilingual data?
You may be able to reduce the impact of noise by providing a small amount (about 1,000 lines) of noise-free language assets and specifying them for "Tuning".

TIPS: How do I compare the performance of multiple custom trained models?
Use the language assets specified for "Testing" in the custom training to evaluate the custom trained model. To evaluate the differences in multiple language assets for "Training", specify the same language assets for "Tuning" in addition to "Testing".

About Kawamura NMT engine type

There are eight types of models (combinations of types and language pairs) for Kawamura NMT. The characteristics of each type are listed below. Please select the one that best suits your purpose.

• General NT: Engine can be used for the translation in any field/purpose of the source documents.
• Patents NT: Engine built specifically for patent translation, and supports specific terms and phrases.
• Finance Services: Engine built specifically for financial services translation, and supports specific terms and phrases.
• Finance NT (IR/Disclosure): Engine built specifically for the translation of IR/disclosure information related to financial service, and supports specific terms and phrases.
• Legal documents NT: Engine built specifically for the translation of legal documents, and supports specific terms and phrases.
• Science: Engine built specifically for translation in fields suitable for academic papers mainly in biological and medical fields, and supports specific terms and phrases.
• Patent decision etc.: Engine built specifically for the translation of notices of reasons for refusal of patent applications and decision information, etc. among patent-related documents, and supports specific terms and phrases.
• News: Engine built specifically for the translation of news, including new words, new topics and current topics, and supports specific terms and phrases.

About Customization Credit (CC)

• Customization Credit (CC) is required to create custom trained models in LAC's Custom MT Management menu. If you have applied for a LAC trial contract, you will receive a specific CC according to your plan. After the termination of the trial period, CC will not be automatically granted once transiting to the main contract. You need to purchase CC every time when necessary. If you wish to purchase it, please go to [Organization Management] tile in Service menu ➡ Organization Management menu, and click [Change plan/limits] button on Billing Information screen and access XMAT Support Contact Form or please contact our sales representative.
• To check your CC history (addition/consumption), please go to [Organization Management] tile of Service menu ➡ Organization Management menu, and click [Customization Credit History] button on Billing Information screen.
  
  • Current Customization Credit Balance: "Unused (Available)" is the number of CCs that are currently available, and "In use (Pending finalization)" is the number of CCs that are being used in the ongoing custom training process.
  • Processed Date/Time: The date and time when the CC was actually added/consumed or expired.
  • Category: The reason of the addition/consumption
  • Changed CCs: Number of CCs increased/decreased
  • Application: MT learning service name for which CCs were used
  • Updated By: If the category is "Custom Training", displays the account name of the user who executed the custom training. In case of "Additional Purchase", "Added by Admin", "Added for Trial Use", or "Expired", the character string "SYSTEM" is displayed as fixed.

About upper limit of models

If your organization already own a maximum number of custom trained models, you cannot request custom training process. Use one of the following measures to ensure that there is enough room to reach the upper limit.

• Delete some of custom trained models you have: Please go to Custom Trained Model List from the navigation menu and delete the models your organization own.
• Increase the Custom Trained Model Limit: If you do not want to delete any model, you can increase the limit. Please access XMAT Support Contact Form or please contact our sales representative.