Interesting New Search Engine

http://www.stremor.com/samuru.html

by jcr • 13 years ago

21 12 13 years ago

jcr2• 13 years ago

There isn't much detail on how their "Liquid Helium" tech actually works but so far I've managed to find the following pdf and video channel.

http://www.stremor.com/lhe.pdf

http://www.youtube.com/stremortv

drakaal6• 13 years ago

Liquid Helium works through heuristiics. LHe is the symbol for Liquid Helium in science, and is the intitials for Language Heuristics Engine. (we are so clever we know)

http://www.youtube.com/watch?v=9pb_F1mwLCo explains it oversimplified.

Basically we create a very efficient set of classifiers using sieves.

Imagine you want to find a "How To".

What makes a great How To? It should be instructional. Have steps. It shouldn't have a lot of opinion. It should have good descriptions. It should be written as commands in the "you" form.

These are all language constructs we can identify. Then we score a document to see that it meets a minimum requirement to be a "how to".

When you do a search we filter and rank results based on what kind of search we think you were performing. We determine the type of search based on what other people have searched, what kinds of documents come back and we build an "ideal result profile" which we attempt to meet with the documents in our result set.

If you are looking for a review we try to find a balance of positive and negative.

If you are looking for a song we try to find lyrics and videos of that song.

We also summarize each result.

Some of these things happen only after someone has done a search for the first time. So if you are the first person to try a search you may have to wait 30 seconds and try again.

hnriot• 13 years ago

So, like google except not anywhere near as fast?

I tried it, the results were ok, I didn't see any result that was better than Google (or bing).

I'm suspicious of anything that starts out by saying search is broken. It isn't.

All these things you have built like sentiment analysis and summarization etc do nothing new than using page rank. The reason is, those balanced review, well written pages are the same ones that humans have curated and used in their own citations. People do a much better job of language understanding than some simple sentiment analysis and the other other things you do. Because humans do it better, when google comes along and leverages all that human curation, the results are staggeringly good. Sure, there are easy to find edge cases, just like there were with your engine when i tried it, but I don't see anything new here. Maybe you could provide some example queries that demonstrate it producing better results than google or bing?

If, you could do comparison queries like Nikon vs canon and it produced a summarized report, that would be interesting, but it didn't do that.

Song lyrics are easy! That doesn't demo anything, nothing could be easier for a search engine than song lyrics...

TYPE_FASTER• 13 years ago

From the PDF, it sounds like they are exploring advanced information retrieval techniques like local context analysis, sentiment analysis, and person/company/location recognizers.

These ideas have been around for a while, and some of them are starting to become more prevalent (phone number recognizers in phones, location recognizers to link to a map, etc.).

Here's an example paper that talks about local context analysis:

http://www.cs.uml.edu/~haim/teaching/iws/tirsaa/sources/ACM_...

jcr2• 13 years ago

Thanks. I'll give it a read.

In addition to the samuru.com search engine, they've also created another display of their "Liquid Helium" tech at:

http://www.unpartial.com/

There are some youtube videos about it as well:

http://www.youtube.com/watch?v=1DCLvBK_sEM&list=PLZ_j4Zk...

drakaal6• 13 years ago

Unpartial runs an older version of LHe. we have been meaning to upgrade it, but it was optimized for Politics with a focus on the election. To make it fast it doesn't have the full set of tools in it, so it's summaries are not quite as good.

drakaal6• 13 years ago

This kind of stuff is part of what we are doing. Extraction of semi-structured data, like phone numbers and adresses is something we do, along with review scores and dates and social media pages. But the bigger thing is the unstructured parts.

Imagine you want to know everything that was ever said about your product, your competitors' products, and comparisons between them.

If you are Apple that is a LOT of stuff. So what if you wanted a machine to do it?

You'd need a copy of the entire Internet. (like a search engine does)

You'd need something that could parse sentences to determine all the features talked about and if they were talked about in a good way or a bad way.

You'd need something to summarize the articles so that humans wouldn't have to flip through every 3000 word review but could still get the highlights. (like our TLDRStuff.com does)

You'd need something to check out the authors to see what their audience reach is, audience demographic, how often they talk about the topic, and how authoritative they are on the subject.

We have all those parts.

We think to build the next generation of search and research engines you have to have all those parts. Link Authority is too easily gamed but putting these parts together we don't need link authority.

Our challenge is making this fast enough to use all the factors, and figuring out how to not penalize brands for searches of their own names since so many brand home pages don't have good information, but shouldn't be number 36 in search results either.

drakaal6• 13 years ago

I (Brandon Wirtz CTO of Stremor) am glad we keep hitting the front page of HN. The response has been a bit overwhelming, so I will do my best to keep up with the comments.

Some answers are already on the other two threads.

https://news.ycombinator.com/item?id=5579804

https://news.ycombinator.com/item?id=5579336

polynomial• 13 years ago

After trying Samuru out, it looks like a great way to broaden a search, however the results seem to skew toward the random.

That is, searching for specific concepts and things, it did not return the most relevant results, but instead results that were heuristically interesting- but not what I was looking for.

I'm also curious about the emphasis on the "Liquid Helium" name(from Language Heuristics engine) which although clever sounds a little silly.

Potentially useful (we do need more search tools) but a bit wide of the mark for what they seem to be claiming.

arindone• 13 years ago

I am really impressed -- With the results in their current state, I'm excited to see how they improve with scale (i.e. more and more people searching and adding data to the machine learning algorithms)

I think a Bing vs. Google vs. Samuru challenge should be done a la what Bing did in their most recent marketing spree; I'm curious what the rankings would be

drakaal6• 13 years ago

Building one of those was harder than we expected, Google won't let you frame their results page.

Interesting New Search Engine

Related Stories