Google Semantic Experiences AI Lets You Talk To Books And Play Word Games
Google is working hard to improve natural language understanding and one of the key areas that is being used to improve that understanding is the development of word vectors. A word vector allows machine learning algorithms to understand the relationship between words based on actual language usage. This has led to Semantic Experiences such as Talk to Books that Google unveiled this week. The tech is similar to what Google uses in Smart Reply for Gmail.
Talk to Books is a new way to explore books that starts at a sentence level. This allows you to enter a sentence or ask a question and the tool will find sentences in books that respond to your question. There is no dependence on keyword matching according to Google. The search giant says that the tool essentially allows you to talk to the books and the responses will help you determine if you are interested in reading the books.
Google says that once you ask a question using the tool it searches all the sentences in over 100,000 books to find sentences that respond to your input using the semantic meaning at the sentence level. Google says there are no predefined rules that bind the relationship between what you type in and results you gain.
There is room for improvement in the tool according to Google because a good matching sentence can be taken out of context. The tool also doesn’t add weight to books that are more popular; it only looks at how well the sentences match up. That might be annoying if you are hoping to find popular books that answer your question, but Google figures it might also help you to find what you are looking for via unexpected titles.
The same tech powers a word association game called Semantris. It's fun to play, but the system has limitations. In my testing, Motorcycle and Engine came up, but the AI didn’t see the correlation to "Honda" as a clue. It did, however, correlate my clue of "Swimming Teeth" with shark.