Source: devclass.com
To give more people a chance to play with natural language understanding, Google AI Hub has introduced Semantic Reactor, an experimental Google Sheets Add-on, to use in service bots and the like.
The Semantic Reactor is promoted as a tool that “allows the user to sort lines of text in a sheet using a variety of machine-learning models”. It is said to be useful for experimenting with chatbot responses, and testing word associations for games among other things. Moreover, Google sees a variety of applications for the approach, ranging from semantic search tasks for forums to building digital personal assistants. machine-learning The Semantic Reactor is promoted as a tool that “allows the user to sort lines of text in a sheet using a variety of machine-learning models”. It is said to be useful for experimenting with chatbot responses, and testing word associations for games among other things. Moreover, Google sees a variety of applications for the approach, ranging from semantic search tasks for forums to building digital personal assistants.
To get going, users will have to add phrases fitting the use case at hand to a Google Sheet, and will then be able to select a model and ranking method for further processing. The candidate list can then be queried from the outside and provide the enquirer with a ranking of all phrases and weights or scores for them.
The ranking methods provided for now are Input/Response and Semantic Similarity. While the former rates the line most appropriate as a conversational response higher, Semantic Similarity puts text with a similar meaning to the input on the top of the list. When writing a conversational bot for example, Input/Response might be suitable since it is hard to predict what people might ask and the number of potential responses is quite high.
Meanwhile Semantic Similarity would be a good choice for bots for use cases with a very narrow focus where questions are easy to anticipate. An example could be a product-related customer support bot. When writing such a thing, Google states that users will have to compile a list of questions and answers, and once a question is asked, the Reactor will find the one in the list that is most similar and answer with the sentence paired to that.
Models used within the Semantic Reactor include a minified and a basic version of the universal sentence encoder, as well as one trained on pairs of questions and answers in a variety of languages, that can be tested against each other for experimentational purposes.
An example application making use of the Semantic Reactor can be found on GitHub. However, since the tool is still only an experiment, those interested in building something similar themselves will have to apply at the AI Hub to get approved and receive installation instructions.
