Language translator for training

EBA NLTraining now uses natural language translator, enabling EBA to learn from non-English langauge inputs. EBA natively supports translation for a variety of languages. If EBA detects that a non-English language is used, it will execute in parallel both the original input and an English translation of the input. EBA will then resolve the user’s input by selecting the variant with the highest score. Before this release, it was not possible to train EBA on a system-generated English translation in the event that it was selected to resolve a question, as the system could not actually understand the original input. With this new translator for training support, it is now possible to provide a clarification to a translated input such that EBA will understand translations in subsequent interactions.

An illustration

Below is an example where EBA intially cannot resolve the Russian input ‘расскажи мне весёлую историю’. EBA neither understands the original input as is, nor the English translation ‘tell me a funny story’. As a clarification, we can tell EBA that ‘расскажи мне весёлую историю’ is conceptually equivalent to ‘tell me a joke’. EBA will then generate the NL components on the translated text during its traning process. In the example below, this means that EBA will introduce a pattern [funny story](:Joke). Please note that such training is not limited to the non-English language. If you ask in English ‘tell me a funny story’, EBA will apply the generated pattern to resolve your input. As a high level interaction, this feature gives users the ability to teach EBA Russian by using English clarifications, similar to a human-to-human interaction with a new foriegn language learner.

edit this article