As part of our continuous improvement efforts, we are happy to inform you of an enhancement upgrade to the NLU inference service. This upgrade provides great benefits from applying additional deep learning technologies. This is an out-of-band enhancement to the NLU inference service that improves the quality of predictions ahead of the upcoming Paris family release in September. The Paris release will also introduce additional features and enhancements in the NLU Model Builder, along with updates to the out-of-box models.
What is changing and improving with this update
These changes will result in NLU models that are more confident in their predictions. In some scenarios where there were no predictions, the NLU model will now understand and predict the intent.
These changes were effective on July 14, 2020, for all customers on all versions. The changes will apply to models that are currently being trained or retrained after this update became available. There is no impact on models that have already been trained and are currently being used in production.
Things you should know about this upgrade
The upgrade is in the NLU Inference Service, which comes pre-trained for language understanding and ready to use, regardless of which Intents, Utterances, and Vocabulary you choose to use in your NLU model. This upgrade requires you to retrain the model to take advantage of the improvements. Existing models will continue to work as they have been until they are retrained.
Several safeguards have been added to prevent issues in training the model and to improve accuracy. In one specific scenario, you might run into an error training the model even though the model previously completed the training successfully. Please check the following to resolve the issue. If you find this information not helpful to resolve your issue, please reach out for support.
Note. In prior versions of the NLU inference service, you may have seen some unknown word errors in some of your utterance examples. This is no longer the case, as the inference service now uses new technologies that derive the meaning of these words based on their surrounding words. It's still a good practice to provide vocabulary for any acronyms or words that are specific to your organization or domain.
| "Invalid synonym is found, synonym should not have reference to any other vocabulary." |
This error occurs when NLU Vocabulary words are used to provide meaning to 'words' that are n't naturally defined in common language. If a Base word is used as the Synonym, or the other way around, the error can occur. Ideally, Vocabulary Bases should contain words that don't exist in the language, and Vocabulary Synonyms should contain entries that provide meaning for the Base words, using words that do exist in the language. Above (click image to enlarge) you can see two Vocabulary definitions: First, there is the base 'worknote' with the two synonyms 'work note' and 'note'. This is a logical definition because 'worknote' isn't a defined word in the English language. The synonyms allow the system to associate 'worknote' with its actual meaning in the context of the system. Second, there is the base 'note' with a single synonym 'work note. As you can be seen here, the base 'worknote' has the synonym 'note', and 'note itself is also a Vocabulary Base with its own synonyms. This creates a circular relationship. This can become increasingly problematic with larger Vocabulary models; the NLU training now stops when it encounters this scenario, and it shows the error "Invalid synonym is found, synonym should not have reference to any other vocabulary." To work around this issue:
As a best practice, remove 'note'. 'note' exists in the English language, so it doesn't need to be re-defined with a vocabulary word. If 'note' has an alternative meaning specific to your model, it shouldn't be used as a synonym for 'worknote'. |
Who is impacted and is there an opt-in or opt-out?
With this upgrade, your model will be smarter and yield more confident predictions. All customers are impacted by this change. There is no opt-in or opt-out available.
If you experience any issues training your models after this update, please contact Customer Support.
