{"id":6066,"date":"2018-04-24T15:21:04","date_gmt":"2018-04-24T19:21:04","guid":{"rendered":"https:\/\/news.elearninginside.com\/?p=6066"},"modified":"2018-04-24T15:22:13","modified_gmt":"2018-04-24T19:22:13","slug":"google-talk-to-books","status":"publish","type":"post","link":"https:\/\/news.elearninginside.com\/google-talk-to-books\/","title":{"rendered":"Will Google’s Semantic Search in ‘Talk to Books’ Supplant Keywords?"},"content":{"rendered":"

On Friday the 13th<\/sup> this month, Ray Kurzweil took the stage at TED 2018 in Vancouver to talk about his latest work at Google. Their new service, Talk to Books, allows you to, well, talk to books. And not just a few books. More like the 100,000+ volumes available in Google Books.<\/p>\n

Here\u2019s how it works. An interested or curious reader can type in a question, a statement, a sentence, a topic\u2014anything really\u2014much as they would speak it in a conversation. Talk to Books then scans all the titles in Google\u2019s library (in roughly half a second) and produces several short segments of books that likely relate to your search.<\/p>\n

Semantic Search in Talk to Books<\/h1>\n

That\u2019s definitely cool, but I\u2019m also kind of burying the lead. Talk to Books does not use a keyword-based search, as does its search engine and nearly every other online library and database. Instead, it uses what Kurzweil calls \u2018semantic search.\u2019<\/p>\n

\u201cSemantic search,\u201d writes Kurzweil in a recent blog post<\/a>, \u201cis based on searching\u00a0meaning<\/em>, rather than on keywords or phrases. Developed with machine learning, it uses \u201cnatural language understanding\u201d of words and phrases.\u201d<\/p>\n

\"Kurzweil\"
Kurzweil at TED2018. Steve Jurvetson, Flickr.<\/figcaption><\/figure>\n

Talk to Books is part of a larger project at Google<\/a> implementing AI-powered semantic recognition. It also features word association games to help test (or train?) their algorithms along with a growing library of end-to-end models to be used by developers.<\/p>\n

Quartz reports<\/a> that Kurzweil does not intend semantic search to replace keyword search. They do not, however, provide a direct quote and we have a tendency to respond by asking, \u2018are you sure?\u2019<\/p>\n

Keyword searches are generally unintuitive applications. Let\u2019s take, for example, JSTOR, one of the most popular academic online databases. If I wanted to search for articles authored by Lawrence Lessig, I would need to navigate to the advanced search menu, toggle to the category of \u2018author,\u2019 and then type in \u2018Lessig, Lawrence.\u2019 If I accidentally typed \u2018Lessig, Laurence<\/em>,\u2019 I would miss out on all relevant results. In some databases, you need to know particular search syntax. (On JSTOR, the above search would look like (au:\u201dLessig, Lawrence\u201d).)<\/p>\n

In the online worlds of research, marketing, journalism, legal matters, personal directories, social media, virtual communities, and countless others, users navigate primarily via search. That search function will be like JSTOR\u2019s at best, and much more basic at worst. The fact of the matter is that in our current environment of hypertext and abundance of information, it can be very difficult to find what we\u2019re looking for.<\/p>\n

Talk to Books, and Kurzweil\u2019s team\u2019s research on semantic search, seems poised to revolutionize how we navigate the internet in general.<\/p>\n

Machine Learning Algorithms Have Issues of Their Own<\/h1>\n

Still, deep learning algorithms do not come without their shortcomings. To paint in broad strokes, the semantic search algorithm seeks to understand and recreate (sort of) human language by learning it from other humans. And other humans have biases. As Kurzweil reiterates in a separate post<\/a>, \u201cIn Talk to Books, while we can’t manually vet each sentence of 100,000 volumes, we use a popularity measure which increases the proportion of volumes that are published by professional publishing houses. There are additional measures that could be taken. For example, a toxicity classifier or sensitive topics classifier could determine when the input or the output is something that may be objectionable or party to an unwanted association. We recommend taking bias-impact mitigation steps when crafting end-user applications built with these models.\u201d<\/p>\n

\n

The systems are mainly trained using white faces and male voices. This “…can lead to misperceptions of black faces or female voices, which can lead to the AI making negative judgments about [other folks].” https:\/\/t.co\/GBho5WBKpv<\/a><\/p>\n

\u2014 chris g (@hypervisible) April 22, 2018<\/a><\/p><\/blockquote>\n