While I strive to be fair and balanced in my coverage of companies–especially those that in any way compete with Endeca–somehow I seem to come down hard on Google.
But today I’m glad to have the opportunity to point readers to a post by Matt Cutts, the head of Google’s Webspam team (and a speaker at the SIGIR ’09 Industry Track!), defending Google against the oft-repeated charge (this time by Om Malik) that Google has run out of big ideas.
I do note that, other than the deep web research, which I covered in an earlier post, I don’t see much about how Google is innovating in search. Perhaps Google is done with search, and is focusing its innovation efforts elsewhere? While I’m personally interested in solving the open problems in the search space, I don’t doubt that investigating alternative energy is imporant too.
39 replies on “Matt Cutts: Google Still Has Big Ideas”
I think Google is a little disjointed in their research, no that’s unfair, I think they are trying many things that are not currently related to their core business so we don ‘t hear a lot about it.
While I think they could still use innovation in the search space (it’s still wide open & how I long for faceted results in general search) I do like companies that can afford to pursuing good pure research. MS for all it’s faults, perceived or real have been doing this for years but they publicize their endeavors a little better by doing their open research days.
Plus with everything converging around information management a lot of what looks like unrelated research actually 5 years out might be the new hotness.
It’s a benefit of all those revenues enabling them to play in many different areas without having to show results.
It’s not just that MS is friendly toward research. For all their faults as a company, that is one thing that they do right. It’s that Google, I’ve noticed, is actively hostile toward most, if not all, research that looks more than 6 months, maybe a year, down the road. But I’ve seen small evidences that their attitude is changing.
I think you’re correct, Christopher, about how what looks like unrelated research now turns out to be highly relevant five years down the road. I know this, because I could tell ya stories about how seemingly unrelated research that I was doing five years ago.. even eight years ago.. is highly relevant now.
That’s what research is.
>>how I long for faceted results in general search
But what facets could reliably and consistently be employed with web content?
In a product catalog with neatly structured data it is easy but what about general web content?
Dates are hard to determine, locations equally hard and I’m pretty sceptical about some global unified taxonomy populated by automated categorisation from text.
How do you reallistically see this working?
Mark, I think you’re setting the bar too high. But you have good company.
Thanks for the link, Daniel. I’m a relative newcomer to your site so excuse my ignorance of previous topics.
From the comment in your link:
“while I concede that faceted search for the web is hard, I think it’s possible ”
Can you elaborate on what facets you think might be possible that Google is missing?
Does this information come from better “on-page” analysis e.g. based on entity extraction or is it “off-page” sources e.g. del.ici.ous tags or “most-closely-related Wikipedia article”.
As you say the adversarial nature of the web makes life challenging.
Choosing “correct” facets in the open web search space is a tough nut too crack but we are seeing a lot of examples (especially in the enterprise space) that give me hope we are getting there.
Companies like Evri.com are making strides and all-though I would not call their search a totally open search space it is a hint at what can be done.
We are making great strides extracting specific contextually accurate items like dates & locations from documents. For extracted facets they do need to have context around them as simply pulling out all dates or all locations from a document is rarely going to offer good facets to use for narrowing the space – there will be too much garbage and noise.
Finally while I agree a globally agreed upon taxonomy may be a tad pie in the sky there are a number of high-level standardized facets that could be employed without a truly unified group approach to creating a taxonomy.
Why does there need to be a global taxonomy? Why does there even need to be a folksonomy? Why not just let the data speak for itself?
We should be seeing more systems like Vivisimo/Clusty, which just uses raw machine learning to find clusters in the data, and then extract ad hoc labels for each of those clusters. Is a cluster the exact same thing as a facet or an aspect? Maybe, maybe not. But I have personally found them to be extremely useful.
What sometimes really convinces me that Google is out of ideas is that they won’t (can’t?) even give us something like the Scatter/Gather algorithm, for interactive information exploration.
You don’t need a taxonomy: Every single word in a document can be treated as a taxonomic label for that document. And then, by implementing a Scatter/Gather interactive algorithm, you allow the user to explore this data in a much, much better way than simply walking linearly through a ranked list.
Here is a good example:
What Christopher said. Or take a look at what Endeca’s done for the ACM and for a leading sports programming site. It’s not that hard to develop a vocabulary that can then be used for tagging–and it’s not even that far fetched to organize it into facets. In fact, I’m surprised that Google, Yahoo, and Microsoft haven’t made more credible efforts to do so. I’d have a field day with access to their data and resources–and especially their query logs.
You could argue that the lack of bells and whistles is not due to a lack of ideas but some other drivers:
KISS – the desire to keep the interface as simple as possible
Revenue – anything that involves complicating the criteria (eg dynamically formed clusters) also complicates the commodity they sell – ads connected to simple keyword searches. .
Cost – turn a 10 ms query into a 20 ms query and you’ve just doubled the size of the servers needed to run this beast.
Indeed, I’ve heard Google execs make those arguments. But, to quote Einstein, “Make everything as simple as possible, but not simpler.”
As for cost, a better interface could reduce the number of queries, even if some (or even all) queries require more computation. It’s not clear how that balances out.
Finally, you’re right that they may be using the only interface they know how to monetize. This is an area where they may eventually need to learn from Endeca. 🙂
Endeca teaching Google how to monetize? That certainly did deserve a smiley 🙂
(1) The very fact that you characterize such functional information retrieval interaction mechanisms as “bells and whistles” means that you’ve already biased yourself against any other, potentially better, solution,
(2) As for the KISS principle, let me introduce to you the idea of “necessary complexity”. Here is a great writeup of it, from someone reviewing the failure of Google’s “Lively”:
If you dumb something down far enough, very few people will actually want to use it. We’re not ragging on Lively here. Instead, we’re aiming to learn from its principles and performance. Let’s introduce a new principle called necessary complexity. The idea here is that any interactive system has a certain amount of complexity, usually involving the number and type of tasks which can be performed. Obviously, it is detrimental if the interaction interface is more complicated than it needs to be. That just makes things harder. What’s a little less obvious is that reducing the complexity of the interaction interface too far makes things harder as well. Either it makes it hard to perform the tasks, or it reduces the number of tasks which can be performed.
(3) For what you say in terms of revenue, Google claims that its search efforts do not interfere in any way with its advertising efforts. Either they need to start telling a different truth, or they need to stand by their word. To me as a user, it seems that if an interactive information seeking mechanism can give me better results, with less effort, than another, then that is the one that offers higher user satisfaction. And Google says the needs of the user comes first.
(4) Cost: Yes, but Google is smart. They can make it work.
Don’t get me wrong – I make a living creating and selling search “bells and whistles” and am equally passionate about advancing the search UI.
While I have seen great solutions built on smaller, cleaner sets of data using these advanced interfaces, I’m pessimistic about how that can be applied to the particular case of Google (but more than happy to be proved wrong!) For example, I see these “advanced” interfaces such as clustering requiring significantly more compute resource to perform their analysis (say 20 milliseconds queries easily creeping up to 100 milliseconds, a 5 fold increase) and I don’t reallistically see that balanced by a five-fold decrease in the overall number of queries. Think of the servers required. Google may be smart but they can’t defy physics.
Incidentally, funny you should mention the scatter/gather interface. Doug (Lucene) Cutting was behind that and I remember reading a comment from him recently where he said that he thought the single-search-box, 10-links style of interface was like the bicycle of UIs – hardly a ferrari but adequate for most people’s needs and unlikely to disappear any time soon for large-scale search.
You’ve got two good points, Mark. First, yes, it’ll take more resources. Second, I postulate an 80/20 rule on the bicycle/ferrari.
I never said anything about10-link going away completely. I only said that, if one wants to interact with the search engine more, there should at least be the ability for the search engine to let the user do so.
Think about it this way: A search engine should have the 10-links as the default. That’ll satisfy 80% of the users, on 80% of their queries. Most people are just looking for someone’s home page, or the answer to a factual question, anyway. And for that they don’t even need 10-links. 5-links would suffice. Or maybe even 3- or 2- Heck, if you think about it, the “I’m feeling lucky” button is the 1-link interface. It’s already getting rid of the 10-link interface, and taking you straight to the 1-link interface.
But when users want to interact more with the engine, when they have a more complex need, or have a simple need that is not so simple to express in 2-3 query terms, then there should be a prominent button, somewhere on the search page, that says “Do the more advanced clustering” or faceting or whatever.
So I am not saying that clustering should be automatically done, for every single query that every single user issues. Rather, it should be something that the user can “turn on” when they need it.
If it’s done that way, then it will not be anywhere near the load on Google servers that you worry about. Because it won’t be happening for every single query — only for the few, chosen, most useful and needful queries.
That’s what I mean about Google being smart. Google would not be smart if they turned on clustering, by default, for every single query. Not every single query needs it.
This is not an all-or-nothing game. It doesn’t have to be all 10-links, or all-clustering. But your response.. Doug Cutting’s response.. and Google’s response.. all seem to be coming from the mindset that it has to be all-or-nothing. That we either have to have 10-links and nothing else, or else we have to cluster, and get rid of 10-links completely.
That just isn’t true. And Google, and others, need to step up and realize this, and stop playing the all-or-nothing game.
Furthermore, I submit to you that if the “necessary complexity” button is marketed/presented in the correct way, then the search engine should be able to get away with load times longer than 100 milliseconds. For example, if the button said something like “I’m Feeling Deep”, or “Heavy Processing Search” or something like that, then you could easily sell the user on the idea that the computer was doing lots of complex work for you, and that it would be worth even a 3-4 second (3000-4000 millisecond) wait to get your fancy clustered results back. Why do I believe this? Because I see the fact that users are patient in other areas. When users buy a song from iTunes, there is not a 100 millisecond response time. It usually takes a few seconds, if not more, before the user can start interacting with the information that they’ve requested.
But the user understands this.. they understand that they’re asking iTunes for a lot of data, and that this data takes more time to deliver.
So if the search engine made clear to the user that they could get deeper processing of the data, I think the user would be totally fine with a few seconds wait. Because an intelligent searcher knows that waiting 3-4 seconds to get a really good, comprehensive answer takes much less time than trying ten different searches, one after the other, to try and construct all those facets yourself, manually. Right?
I love the idea of an “I’m Feeling Deep” button. I’ll see if I can sneak that in at the Jeff Jarvis talk tonight.
So for various reasons we’ve converged on the notion that 10-link search is sufficient for 80% of queries and more exotic searches/exploration should be on-demand features away from the main service.
Is this not roughly where Google are now?
I can’t speak for the hive, but I don’t think 10-link search is the best approach for 80% of queries. That’s why I switched from Google to Duck Duck Go as my day-to-day search engine.
Well, I guess I’m only speaking for myself, not Daniel 🙂
But I look at your links, and I’ll admit: I’m confused. Google Insights for Search only tells you information about other people doing the same search that you are doing. It doesn’t actually give you any better way of sorting through and understanding the 352,000,000 pages that are returned by an actual search on the query [obama].
So no, Google Insights for Search has nothing to do with my actually information seeking behavior. It’s probably more of a tool geared toward marketers/advertisers, to help them understand how to place keywords. Not a tool for searchers, to get better, exploratory, results.
And Google timeline view is nice. But let’s assume that there are indeed 100 of every 500 Google queries that need an exploratory interface (20%). Daniel says the number is higher. But let’s just go with 100 of 500 for the moment. I would say that, of those 100 exploratory queries, only 1 or 2 come from information needs that require a timeline view. How are the other 98 or 99 exploratory queries being handled? Not at all.
So no, that’s not where Google is now.
Furthermore, even if timeline view were the be-all and end-all of exploratory search: As a user, how do I know that the timeline view is even available? Where does Google show me the button or link to switch to timeline? You’re telling me I have to know some hidden “view:timeline” operator? How was I supposed to learn that this exists? I click on the “advanced search” button on the Google homepage and it’s not there. I do a normal search, and it’s not there.
What good is a tool, if it’s hidden/buried? That’s a terrible user interface!
Again, this is not roughly where Google is now. Google is not supporting exploratory search.
DuckDuckGo is nice but looks to just be using Wikipedia content for disambiguation and categories.
You could argue the manually-maintained Wikipedia has a role to play as the “global taxonomy” we alluded to earlier. It appears there’s certainly no algorithmic smarts behind DuckDuckGo that is clustering on the fly and intelligently creating category labels.
Would you suggest Google simlarly adopt Wikipedia to form its non-10-link view of the world? That would certainly seem computationally more achievable than fancy clustering algorithms and likely to generally produce better results.
I agree that Duck Duck Go is squeezing the juice out of low-hanging fruit, if you’ll forgive the mixed metaphor. But opportunism is hardly a crime, and Google does its own crowdsourcing by using the link graph, anchor text, etc.
Moreover, I’ve heard reports that 30% of Google queries return Wikipedia results. That seems high, but Google isn’t transparent enough to reveal the true number–and I’ve asked someone privy to the information. Regardless, it’s clear that Google certainly leans heavily on Wikipedia. Why not make better use of it?
Jeremy, maybe the fact that “insights” targets search terms rather than pages is because it is significantly easier to identify the “where” and “when” of search requests rather than web pages? That would seem like a reasonable explanation to me.
Locations/timelines aside – what type of “exploratory searches” are the other 98 requests you mention? You claim on-the-fly clustering is useful e.g. Clusty , but I personally have found that typically less than useful e.g. taking our “obama” example: http://clusty.com/search?input-form=clusty-simple&v%3Asources=webplus&query=obama
Mark: What about that query are you finding less than useful? I type Obama, and I see a whole lot of categories or facets that I was unaware of, previously. I see that there are a lot of local blogs about Obama. Love the blogosphere. Oh, I see that there is a cluster about his voting record. That’s interesting.. let’s take a look at that. Ooh.. there is another cluster of “fights and smears”. How tabloid! I would have never realized that such a set of pages existed, if I had to go through the plain 10-links interface, one link at a time.
So the exploratory interface brings me information that I would have never thought to look for on my own, and that isn’t readily available through the 10-links interface.
But that’s just for the query [obama]. Let’s go back to Doug Cutting’s example, for the query [Criminal Actions Against Officers of Failed Financial Institutions].
So the premise here is that I am not looking for any one web page. There is no “CriminalActionsAgainstOfficersOfFailedFinancialInstitutions.org” page that I can “I’m feeling lucky” arrive at. Rather, I am looking for a whole set of information. And I am exploring; I don’t necessarily know the best way to get to it all.
First, let’s try Google:
Oddly, Google brings up the Doug Cutting page first. Not what I was looking for at all — in fact the complete opposite. But how well does Google do after that? I see a couple of links to lawyers. Maybe a few links to one or two specific examples / court documents. But I have no idea how to go forward from there. Do I keep clicking “next page”, 10 links at a time, and reading everything?
Let’s look at Clusty, now:
Ok, Clusty also gets the Doug Cutting page at the top. That’s funny. But even before we get to the top of the ranked list, which is in the middle of the page, we have a whole series of options on the left hand side. I see that information before I see the Doug Cutting page. And those options/cluster tell me that there are a lot of different ways I could go about exploring through the information I’m trying to learn more about.
(1) I could look more directly at the Directors and Officers involved. Who were they? What did they do?
(2) I could look more directly into the Banks involved (the “failed financial institutions” — notice how I didn’t even use the word “bank” in my query, and Clusty still gives me an aspect/facet with that name on it.)
(3) I could look into the litigation that has been done — actual court cases
(4) I could look into the laws themselves, that regulate these scenarios
(5) Oooh.. “class actions”, huh? What does that have to do with anything? I was thinking “criminal actions” in my original query. But this is telling me that maybe some sort of shareholder legal action often accompanies a criminal action. That could be a useful avenue!
(6) Same for civil actions. Interesting.
Finally, let’s see how well “Google Insights for Search” does:
Worldwide, 2004 – present
Not enough search volume to show graphs.
My conclusion: Fail. 🙂
I’m not saying that Clusty is the ultimate answer for exploratory search. I am just saying that it at least tries, whereas Google does not. I can’t get from Google, either from Search or from Insights, what I can get from Clusty.
Mark: I have a longer post that is probably held up in Daniel’s “flagged as potential spam” wordpress queue, probably because it contains more than 4 links. So check back; I will have an answer for you.
Jeremy, sorry about the delay in your comment getting through. I’ll have to have a word with my spam filter about that.
No worries on the delay. I only guessed/suspected, because I recently set up a wordpress installation from scratch, and noticed that the defaults say “hold for my attention of post contains 4 or more links”. I thought to myself.. you’ve probably got the same default 🙂
Look forward to hearing about the event.
I guess its about target audiences. For me Clusty doesn’t pass the “Mum test” – as in if my Mum saw the obama links I was seeing e.g. “elect” and “vote” she would think WTF? (possibly more politely though). Consistently generating useful distinct clusters and labelling them well is very hard and I’m sure Jakob Nielsen would have something to say about the end-user experience with some of those suggested links. It just doesn’t feel ready for prime-time use on Google where the average user would be less forgiving than you or I when results are less than ideal.
I guess hiding it away from the main flow of search traffic in an “advanced” area would remove my Mum from the equation.
BTW, thanks for an interesting discussion!
My personal take-aways, at least in respect of Google/large scale search, are:
1) We didn’t come up with any sensible facets that could be reliably mined from web-based data e.g. location, price, date
2) In the absence of 1), clustering text provides an alternative means of “exploratory search” but this perhaps is not for everyone due to a combination of usability (poor labelling/organisation) and compute power required.
3) Wikipedia has an interesting potential role as a human-maintained means of grouping or providing perspective on text content/searches. This avoids the problems of entirely machine-based grouping/labelling when exploring but is constrained by what subjects are covered in Wikipedia.
I agree re: Clusty. I’m curious if Vivisimo manages to deliver a better experience to its enterprise customers, since Clusty for me actually strikes me as an anti-demo for their clustering functionality.
As for your take-aways:
1): I agree that no one has done it at web scale. I don’t concede that no one can. Microsoft researchers (Sue Dumais et al.) seem ready to concede, and Google isn’t very loquacious. But I still have the audacity to hope.
2) Clustering text is a *lot* easier when you have search logs. We’ve exploited this at Endeca for tagging, and it works like a charm. Just imagine what you could to with web-scale logs!
3) Agreed, agreed! And that’s why I’m so fond of Duck Duck Go: they worked with the army they had and achieved, if not a “mission accomplished”, at least an impressive start towards winning the war on legacy search.
Daniel: Agreeing re:Clusty.. was that agreeing with me, or with Mark?
Mark: Either way, the larger point, where we started this discussion, is on the topic of whether Google still has big ideas, especially in web search. And I have not yet seen a compelling case that they do. I see them as having given up. And that is where, personally, a lot of my frustration lies.
Whether or not Clusty does it “right” is beside the point. You’re being too limited if you think that it is only about “com[ing] up with any sensible facets that could be reliably mined”. That is only one way of thinking about clustering, and doesn’t mean that there aren’t other ways. In fact, coming up with other ways is exactly the sort of “big idea” that I have been expecting Google to come up with for years now. Just because the three of us can’t do it, in 24 hours, doesn’t mean it can’t be done. I share Daniel’s idealism.
FWIW, I recently had a great experience with Clusty, that I didn’t have with Google. One of my current active research areas is on explicitly collaborative information retrieval. I did both a Google search and a Clusty search on this phrase [collaborative information seeking] yesterday. The Google search gave me links to all the stuff I already knew about.. the literature from the past year or two.
On Clusty, however, I immediately saw one very interesting cluster that popped out: “CSCW’98”. CSCW stands for “computer supported collaborative work”. Turns out there was a workshop 11 years ago on this topic. And so Clusty’s exploratory mechanism led me to very rapidly discover this, when I had not known about it before.
So I went back to Google, and went through all the results, until I found a link that mentioned CSCW. I didn’t find one until the 5th page.. the 57th result. Most people give up after page 1, if not page 2. But Clusty’s exploratory mechanisms help me discover something that I otherwise would not have, something that the page-rank based wisdom of crowds completely missed.
Talk about Jakob Nielsen / user unfriendliness.. would you rather have a poorly labeled cluster, or have to read through 5+ pages, 10 links at a time, until you got to that first new and interesting link? Seems like the Google approach is much less friendly.
So whether or not Clusty does everything right doesn’t matter to me, so much as it demonstrates to me that there is not only validity, but also real utility, in an exploratory search system.
Anyway, thank you also for a very interesting discussion. This is why I like the blogosphere 🙂
Speaking one more time of Jakob Nielsen and labeled clusters.. what do you think of the following notion: What if Clusty didn’t do ANY labeling of clusters at all? What if they instead just did the clustering, and gave you 5 or 10 buttons/links on the left, for you to switch between the various clusters, at will?
That way, if you were satisfied with the flat list, you wouldn’t have to do anything. But if you thought to yourself: I need to get a different perspective on this information…but short of going through hundreds of links, one at a time, how do I do that? Then breaking the results set into some sort of latent clustered factorization, and letting you browse through each subset, would give you an easy, and I think extremely usable, way of understanding everything in your initial ranked list.
I am saying that there is no need to label the clusters, because the 1-2 sentence snippets of each of the top 5 links from each of your unlabeled clusters would essentially be a a summarization of that whole cluster. In fact, you could present a paragraph-long aggregation of those sentences as your summary for the cluster as a whole.
Would it take a little bit of extra time, to read this paragraph, for every cluster? Yes. Would it take less time to read each paragraph, than you step through the original Google ranked list, one 10-link page at a time? Also yes.
>>would you rather have a poorly labeled cluster, or have to read through 5+ pages
I actually quite like “significant keywords/phrases” in a UI e.g. when http://www.gigablast.com suggests “H5N1” for a search on “bird flu”. I think that approach can provide “beyond page 1” insight without being based on the principle of results needing to be organised into strongly distinct groups. In some apps sometimes it is the single phone number or email address that may occur only twice that is hugely significant..
>>What if Clusty didn’t do ANY labeling of clusters at all?
Yes, I think listing the page titles would typically make it faster to get a sense of a cluster’s content. Title info is typically very information-rich.
My objection is to Clusty is not the interface, but the content. Though my suspicion is that they chose their interface too ambitiously: it sets expectations they simply can’t meet. They probably would have been better off to show significant phrases rather than clusters.
Or they could compute those phrases for the entire result set and then cluster the phrases (which is what we do at Endeca), rather than first clustering the top few hundred documents based on inter-document similarity and then computing phrases for each cluster (which is what I believe they do).
Hmm.. but once you’ve extracted phrases, by what mechanism do you (Endeca) then cluster the phrases? Certainly raw phrases have no natural clusterability, beyond simple subphrasing, no?
The issue of labeling aside (I like Mark’s suggestion to just use the titles of the few top-ranked pages in each cluster as the “summary”, rather than assigning any explicit label, by the way), what’s wrong with doing it the 2nd way.. the full-text inter-document similarity method? Do you know of work showing how/why that approach doesn’t work? (And for what it doesn’t work on?)
But ultimately, wouldn’t it be even more transparent to let the user roll their own cluster? Maybe start the user off with the raw, full ranked list.. and then give them tools to try rolling the cluster either by-phrase or by-content.. or both?
With dialogue-based, interactive, exploratory search, might it not be useful to let the user chop and slice however they want? The more tools the better? Isn’t giving the user the ability to poke around, themselves, rather than deciding for them an important part of an exploratory search?
Maybe by rolling it one way, and then the other, the user can get a better sense of the results, overall. That’s because the difference between the two clustering approaches might reveal both similarities and differences that gives the user more insight into the information being explored. Maybe.
I ask these questions half rhetorically, but half seriously, too. I don’t really know the answers.
Oh, the clustering is based on co-occurrence or some other relative information measure. My point is that the sequence should be documents->phrases->clusters, not documents->clusters->phrases. I can’t prove that the latter doesn’t work, but my experience with and knowledge of document similarity measures is that they are opaque and somewhat unreliable. It’s better to work at the level of phrases.
In any case, showing exemplar titles for clusters is a good idea, independent of the clustering method.
As for letting users roll their own, I agree that they should be able to. But I understand that users may not want to do all that work. I say, use automation to make a best first cut, but give users the ability to refine it.
Thanks for the phrases explanation.
By roll their own, I mean give ’em a simple button they can click, to turn it on/off. That doesn’t seem like too much work. Certainly not as much work as scrolling to the bottom of each page, clicking next page, and repeating that five times in a row.
Agreed, that’s a no-brainier. I’d take it further, offering users clusters that they can modify. What Koenneman and Belkin call penetrable relevance feedback, only in a more general sense.
Yahoo had an experiment a few years ago, where they gave you a slider that dynamically (real time) let you control the mixture of results to a query. The two endpoints of the slider were, if I remember, “commercial” and “academic”. So if you wanted all commercial results, you could pull one way. Vice versa for all non-commercial results.
I’m all for more of this kind of thing.