Active Question Answering (ActiveQA)

2018: Release and disclosure of codes

In October, 2018 the Google company provided the new system of artificial intelligence which processes informal conversation and with a high accuracy answers the asked questions. Such technology will make interaction of people with voice assistants, including Google Assistant, more natural — as though the person communicates with similar.

A system received the name Active Question Answering (ActiveQA), Google laid out its source codes in a public access. The software engineer in division of Google AI Michelle Chen Huebscher calls development "the agent who over and over again interacts with the voproso-response systems, using a natural language for providing the best answers".

For example, if to ask date of birth of the physicist and inventor Nikola Tesla ActiveQA, having asked the question "When Was Born Tesla?", then a system reformulates a question in "In what year Tesla" and "When a birthday of Tesla is given birth?", will send information to the database and finally will issue the most exact answer — on July 10, 1856. The logic of this example is illustrated on the animated image below.

Google notes that each rephrasing of a question is estimated, making a start from  that, the answer received on  its  basis is how good. If the answer satisfies to a request, then  the algorithm of training  will use more likely this reformulation.  If the answer is incorrect, ActiveQA will give to this option a smaller priority.

In process of training a system begins to ask more specific questions thanks to what results on an output improve.^[1]

ActiveQA is available in the form of a packet to the platform of machine learning Google TensorFlow. The company also offers previously trained system of the sequences adapted using TensorFlow Neural Machine Translation Tutorial Code.

Notes