Google today announced a pair of new artificial intelligence experiments from its research division that let web users dabble in semantics and natural language processing. For Google, a company that’s primary product is a search engine that traffics mostly in text, these advances in AI are integral to its business and to its goals of making software that can understand and parse elements of human language.
The website will now house any interactive AI language tools, and Google is calling the collection Semantic Experiences. The primary sub-field of AI it’s showcasing is known as word vectors, a type of natural language understanding that maps “semantically similar phrases to nearby points based on equivalence, similarity or relatedness of ideas and language.” It’s a way to “enable algorithms to learn about the relationships between words, based on examples of actual language usage,” says Ray Kurzweil, notable futurist and director of engineering at Google Research, and product manager Rachel Bernstein in a blog post. Google has published its work on the topic in a paper here, and it’s also made a pre-trained module available on its TensorFlow platform for other researchers to experiment with.
The first of the two publicly available experiments released today is called Talk to Books, and it quite literally lets you converse with a machine learning-trained algorithm that surfaces answers to questions with relevant passages from human-written text. As described by Kurzweil and Bernstein, Talk to Books lets you “make a statement or ask a question, and the tool finds sentences in books that respond, with no dependence on keyword matching.” The duo add that, “In a sense you are talking to the books, getting responses which can help you determine if you’re interested in reading them or not.”
It is a legitimately neat and super polished product, from my experience using the web interface. Ask it a question like “why is the sky blue?” and you’ll get a number of different answers displayed in clear text, sourced from books on the subject, like, “The Rayleigh scattering of light by molecules in the atmosphere gets stronger as the wavelength decreases.” But, as opposed to using standard Google Search and having to click a link and parse an article or webpage, the Talk To Books algorithm does that work for you.
“The models driving this experience were trained on a billion conversation-like pairs of sentences, learning to identify what a good response might look like,” Kurzweil and Berstein explain. “Once you ask your question (or make a statement), the tools searches all the sentences in over 100,000 books to find the ones that respond to your input based on semantic meaning at the sentence level; there are no predefined rules bounding the relationship between what you put in and the results you get.”
Of course, as you might suspect, there are some limitations here. The tool is better for answering raw factual questions and doesn’t perform quite as well handling complex geopolitical questions or topics of modern cultural and historical importance. But as a simple web tool, and one Google says helps improve products like Gmail Smart Reply, Talk to Books is a fun way to explore the web in a semantically natural way. It also gives us a glimpse of what future interfaces might look like when AI is actually sophisticated enough to handle almost any query we throw at it.
The second of the two experiments released today is far more interactive. It’s a game called Semantris, and it basically tests your word association abilities as the same software that powers Talk to Books ranks and scores the words on-screen based on how well they correspond to the answers you input. For instance, if you’re given the word “bed” at the top of a collection of 10 words, you might think to type “sleep” as a response. Semantris will then rank the 10 words and give you points based on how well it thinks the semantic relationship between bed and sleep is in comparison to the relationship between “bed” and every other word in the list.
It should be noted that a lot of these Google experiments are also ways for the company to gather user data, which can help inform its technology by giving it ample human-grade information on word relationships and so on. That appears to be the case with Semantris, but regardless, the game is a fun way to test your own abilities and to see how well the software judges the associations between words. You can also play a Tetris-like version of the game that lets you input words to clear blocks from the screen, based on your own assumptions of what associations the software may draw between words written on the colored blocks and the answer you type into the bottom.
Like many of Google’s past AI experiments, like the recent Teachable Machine tool for letting users train their own basic algorithm and past ones focused on doodling and music-making, these web games and tools are valuable ways to interact with and learn more about artificial intelligence in the ways it’s more readily applied in the real world. AI, as well as terms and phrases like machine learning and neural networks, is often an abstract concept we hear thrown around a lot without much context or in a way meant to obfuscate or gloss over what’s really going on under the hood of the world’s most powerful software applications and platforms. But with experiments like these, Google is able to demystify the technology in a way that’s beneficial for everyone.