Ask HN: Access to the Corpus - What Would You Do?

So .. Google <i>have</i> these data. Their hacking chops are not unremarkable. What have they done with it? Well, basically, you can ask for a word and they'll find other pages that feature that word ...<p>And that's it.<p>They have access to "the world's combined knowledge" and a zillion PhDs, and that's <i>all</i> they can do?! It's shocking. But it's not, really, because the data are basically useless without annotations.<p>So let's go to Disneyland and pretend we have a genius NLP engine or an annotated web. Then,<p>1) 360 on a company/product. In particular: who are <i>all</i> the stakeholders, and how do they feel? I'd sell this to e.g. analysts. Same thing on people's online identities.<p>2) Memetracing. I'd sell this to advertisers. (So, we follow historical product releases and see exactly what memes spread about them and how, and through who. Related to (3) on ismarc's post)<p>3) Rumors (this would require your feed to be real-time). I'd also probably sell this to stock traders (the idea here is to monitor e.g. forums frequented by GE employees to guess scoop)<p>OK, so those are basic. More interesting:<p>4) Organization-tracing. If you can label social graph edges with influence levels / information intakes, you can start playing with predicting organizational decision-making (behavioral economics / game theory ..There's a TED talk about this).<p>5) Games: procedural content generation that looks really real, i.e. worlds full of people whose identities are plausible, whose interactions with others are plausible, etc.<p>Those all require some analytical / NLP firepower on a scale which I don't think is really doable at the moment. The problem is that bags of words are meaningless without a social context - the data are pretty worthless unless your computer can figure out who it's important for and why.

SEO? With something like that you could programattically find semantic locations with lower coverage and then sell the knowledge that you could possibly attack that keyword. For instance, I'd love to see a chart which plotted frequency of use in an English corpus against some hypothetical Google-coverage variable. Anything that's a strong outlier from something like an exponential curve could be a good target.

You have to consider that all that data isn't just a list of web pages. That includes information provided by the web page and the contents of the page itself (as well as associated metadata).<p>1) Create a map of the web (what links where and how), enabling an enhanced "browsing" experience (no more perusing a site for links to other interesting places)<p>2) The contents of those pages contain a large volume of technical documentation as well as a large number of opinions of the technologies that rose from the documentation. With both, and a large enough history of the creation of pages, the release of the documentation and the opinions presented, a model can be built to predict the success rate of any particular technology.<p>3) Given sufficient time, there is a large enough set of text that can be rendered to audio through speech synthesis (higher quality the better) in order to train a speech recognition system to a previously unseen level of accuracy.<p>4) Given a proper algorithm, sites with security vulnerabilities can be discovered just from what is crawlable and most likely should not be accessed. From that list, and given the total number of unique websites contained in the database, you can calculate the ratio of harmful to potentially non-harmful websites and provide a risk threshold of any given link on any given page.<p>5) Provide a search engine that uses regular expressions on different sets of the data (metadata, tags, text, text in a specific tag, etc.) to present a search engine with a previously unseen level of accuracy (accuracy of the results is dependent on the individual doing the searching, not the systems ability to guess that by std::list, I didn't mean "Sexually Transmitted Diseases STD List" (seriously, <a href="http://www.google.com/search?sourceid=chrome&#38;ie=UTF-8&#38;q=std::list" rel="nofollow">http://www.google.com/search?sourceid=chrome&#38;ie=UTF-8&#3...</a> has that as the second result).<p>Edit: I failed at formatting

I'd like to see which search results were clicked on for a given search, and then associate those sites. And then do an amazon style: people who landed at this site also went to _____. As well as other relations based on user browsing, not based on site links.

I would use it to fight crime/corruption. It isn't too hard to identify people from a few google searches. Examples:<p>Food places near X will give you a good indication where they live and their income Find what they are searching during normal work hours. Often this will give you where they work. Searches of themselves.<p>Then just use this information to hunt them down or expose them.

Create new tools similar to how Google, Yahoo and MS already do when leveraging this asset. I know this is somewhat a BS answer, but really what other answer is there? This is like asking the question "What new features should Google, Yahoo or MS build on top of their search?"

Download all of the porn on the internet. I'm only half joking.

I would keep the meta data and reverse engineer the algo!

Sell it to someone, who'd know what to do with it?

I wouldn't create a business. I'd write a program that could learn (text-based) language. Then I'd write it up and submit it to Computational Linguistics. Then I'd die in obscurity as someone took my idea and figured out how to monetize it.

Start the world's most underhanded (and successful) SEO firm.

Better machine translation.

This scenario is reminiscent of what happened a few years back when AOL released their whole corpus of search queries with associated user-ID hashes. Most of what was done with that data set amounted to amateur CSI work and general meanness. I don't recall any profound revelations or can't-miss business opportunities coming out of it. The value of raw data is overrated.