Archive for the ‘Artificial Intelligence’ Category

High-quality personal filtering

August 25, 2008

Imagine a personal filter that connects people to web-content and other people based on interests and memories they uniquely have in common – enabling discovery of new knowledge and personal connections. High quality matches offering freedom from the chore of reading every item so you don’t worry about missing something you care about in your email, on the web, or in a social network. 

Background

There are 1.3 billion people on the web and over 100 million active websites. The Internet’s universe of information and people, both published and addressed to the user is growing every day. Published content includes web pages, news sources, RSS feeds, social networking profiles, blog postings, job sites, classified ads, and other user generated content like reviews. Email (both legitimate and spam), text messages, newspapers, subscriptions, etc. are addressed directly to the user. 

The growth of Internet users and competition among publishers is leading to a backlog or heap of hundreds or thousands of unread email, rss, and web content in user inboxes/readers — forcing users to settle somewhere between the extremes of either reading the all the items or starting fresh (as in “email bankruptcy”) .

Search is not enough.  If you don’t have time to read it, a personal filter reads it for you.  If search is like a fishing line useful for finding what you want right now, a personal filter is a fishing net to help you capture Internet content tailored to your persistent interests — for which it is either painful or inefficient to repeatedly search and get high quality results on an hourly or daily basis.

Techniques

Personal filtering can help users recover the peace of mind that they won’t miss an item important either in the email inbox, on the web, or in a social network — provided matching algorithms deliver sufficient signal-to-noise ratio (SNR). For example, an alert for “Devabhaktuni Srikrishna” produces relevant results today because it is specific – i.e. SNR is high. Only 1:100 or so alerts for “natural language processing” or “nuclear terrorism” end up being relevant to me. Signal is low, noise is high. So SNR is low but can be improved through knowledge of personal “context” gained via emails, similar articles, etc. Successful/relevant matches are all about context – what uniquely describes a user, and can we find it in other news or users?

  1. The relevance of a match should be easily/instantly verifiable by the user.
  2. Statistical analysis of phrases is proven to be a useful method to extract meaningful signal from the noise and find precise matches.
  3. Phrases describing each person are in what users write, read, and specify.

Framework

Let’s define the user corpus to be all content describing users, which includes the user’s own email, blogs, bookmarks, and social profiles. Interests can also be derived from items or links the user decides to upload or submits specifically by email – “like this.” As new information arrives the most likely and statistically useful interests are automatically detected through natural language processing algorithms (aka text mining), similar to Amazon’s Statistically Improbable Phrases (SIPs). The interests are displayed on a website analogous to friend lists in social networking, the interest list is automatically ranked, and the user vets the list to keep only the most meaningful interests.

 

Once a user’s interest list is generated, it can be used to filter all information in the entire web corpus via open search engine APIs (such as Yahoo-BOSS and Live Search API 2.0) to create a personal search index. The filter may also be aimed at a subset of the web which we call the user-targeted corpus configured directly by the user – including incoming email, alerts, RSS feeds, a specific news source (like the Wall Street Journal), social networking profiles, or any information source chosen by the user. There are now open-standard APIs to email (like IMAP for Gmail and Yahoo Mail Apps), social networks (like Facebook API and Opensocial for Linkedin, Hi5, MySpace, and other social networking sites), and newspapers like the New York Times.

 

The result: A manageable number of high-quality filtered matches, ranked and displayed for the user, both viewable on the website and in the form of a periodic email (daily update).

 

The Personal Filtering Space

References statistical analysis of phrases

Personalized news/RSS filtering

“Recommended News” in Google News, Yahoo’s personalized news portal, My Hakia, Technorati, Newsfire, Digg, Socialmedian, Alerts.com, Dailyme, Loud3r, Feedhub, the personal portals from WSJ/NYTimes/WashPost, NewsGator/RivalMap, AideRSS,  DayLife, Reddit, Feeds 2.0, Twine, Feedrinse, and Particls, Filtrbox, Meehive, Pressflip, Factiva by Dow-Jones, HiveFire (see blog post and demo), Sprout,  my6sense  (see video and review), iDiscover (iPhone app), Skygrid for mining financial information in blogs (see article), Uptodate and JournalWatch for tracking medical research by specialty,  Anchora.

Content-to-content matching

Sphere, Angstro (mining the web for mentions of your friends using their social profiles), Pique by Aggregate Knowledge, Loomia, MashLogic (see TechCrunch writeup) and Zemanta search for and display “related” content on blog posts and news articles currently browsed — analogous to how Google AdSense displays contextually relevant ads, except these services serve up blog/news/other content. Inside Gmail, Google displays related content links below their ads under “More about…”

Person-to-person matching

EHarmony, Twine, Redux, Searchles, PeopleJar, and several Facebook apps have tried to do things like this FriendAnalyzer.

Persistent search

Factiva Alerts by Dow-Jones, Google Alerts, Filtrbox (see Mashable interview), Yotify (see Mashable writeup), CleanOffer for real estate listings, Everyblock for news stories about your own block through RSS or email, Trovix searches jobs on Monster.com, and Oodle searches multiple classified sites. Operating systems and email clients have had a user-configured “Virtual folder” feature for some time – configured by the user (manual), or auto-generated whenever a user performs a search (over-active). , Pressflip for news sources like Reuters, UPI, etc. There is also a way to do Persistent search in Gmail

Some examples of academic/research projects

Creating hierarchical user profiles using Wikipedia by HP-Labs, Interest-Based Personalized Search (Utah), The CMU Webmate, UIUC’s “AN INTELLIGENT ADAPTIVE NEWS FILTERING SYSTEM”, “Hermes: Intelligent Multilingual News Filtering based on Language Engineering for Advanced User Profiling (2002)”. Fido from Norway –“A Personalized RSS News Filtering Agent”. HP researchers compared three ways of detecting tags in “Adaptive User Profiles for Enterprise Information Access“ using del.ico.us bookmark collections. “Writing a Personal Link Recommendation Engine”  in Python Magazine demonstrates a web document ranking system based on content in del.ico.us bookmarks.

Natural Language Processing & Search

Cognition, Hakia, and Powerset (acquired by Microsoft). Google personalization incorporates Personalized PageRank and uses Bayesian Word Clustering to find related keywords — see “Introduction To GoogleRanking” and “Technologies Behind Google Ranking“.

Collaborative Filtering

Digg, Socialmedian, StumbleUpon, FastForward by BuzzBox, Amazon.com (for books), “Recommended News” in Google News, Twine, Filtrbox.

Personal email analysis and indexing

Xobni, Xoopit, ClearContext, Email Prioritizer by Microsoft Research (see writeups here and here)

Personal Music Filtering

Pandora

Recommender Systems

See the list of recommender systems in Wikipedia.

Related Posts

Bill Burnham writes,

…no one has yet to put together an end-to-end Persistent Search offering that enables consumer-friendly, comprehensive, real-time, automatic updates across multiple distribution channels at a viable cost.

Personalized Clustering: It’s too hard, say developers

Because let’s face it, Personalization + Clustering is the next big step in RSS. If 2005 was about Aggregation, then 2006 is all about Filtering.
Nik wrote up his thoughts today, in a post entitled Memetracking Attempts at Old Issues. While he mentions lack of link data as being an issue, it seems to me the crux of the problem is this:

“generating a personal view of the web for each and every person is computationally expensive and thus does not scale, at all.”

He goes on to say that “this is why you don’t have personalized Google results – we just don’t have the CPU cycles to care about you.”

So it’s mainly a computational and scaling problem. Damn hardware.

I think enabling technology will be new and better algorithms. Also see the links in Filtering Services and “Why Filtering is the Next Step for Social Media”.

Web 3.0 Will Be About Reducing the Noise

…if you think it is hard enough to keep up with e-mails and instant messages, keeping up with the Web (even your little slice of it) is much worse…. I need less data, not more data.

Bringing all of this Web messaging and activity together in one place doesn’t really help. It reminds me of a comment ThisNext CEO Gordon Gould made to me earlier this week when he predicted that Web 3.0 will be about reducing the noise. (Some say it will be about the semantic Web, but those two ideas are not mutually exclusive). I hope Gould is right, because what we really need are better filters.

I need to know what is important, and I don’t have time to sift through thousands of Tweets and Friendfeed messages and blog posts and emails and IMs a day to find the five things that I really need to know. People like Mike and Robert can do that, but they are weird, and even they have their limits. So where is the startup that is going to be my information filter? I am aware of a few companies working on this problem, but I have yet to see one that has solved it in a compelling way. Can someone please do this for me? Please? I need help. We all do.”

The Music Genome Project powers Pandora

The Music Genome Project, created in January 2000, is an effort founded by Will Glaser, Jon Kraft, and Tim Westergren to “capture the essence of music at the fundamental level” using over 400 attributes to describe songs and a complex mathematical algorithm to organize them. The company Savage Beast Technologies was formed to run the project.

A given song is represented by a vector containing approximately 150 genes. Each gene corresponds to a characteristic of the music, for example, gender of lead vocalist, level of distortion on the electric guitar, type of background vocals, etc. Rock and pop songs have 150 genes, rap songs have 350, and jazz songs have approximately 400. Other genres of music, such as world and classical, have 300-500 genes. The system depends on a sufficient number of genes to render useful results. Each gene is assigned a number between 1 and 5, and fractional values are allowed but are limited to half integers.[1] (The term genome is borrowed from genetics.)

Given the vector of one or more songs, a list of other similar songs is constructed using a distance function.

To create a song’s genome, it is analyzed by a musician in a process that takes 20 to 30 minutes per song. Ten percent of songs are analyzed by more than one technician to ensure conformity with the standards, i.e., reliability.

The technology is currently used by Pandora to play music for Internet users based on their preferences. (Because of licensing restrictions, Pandora is available only to users whose location is reported to be in the USA by Pandora’s geolocation software).[2]

How to Extract a Webpage’s Main Article Content

I had an idea to make a personalized news feed reader. Basically, I’d register a bunch of feeds with the application, and rate a few stories as either “good” or “bad”. The application would then use my ratings and the article text to generate a statistical model, apply that model to future articles, and only recommend those it predicted I would rate as “good”. It sounded like a plausible idea. I decided to start a pet project.

I soon learned that this idea wasn’t original, and in fact had been attempted by quite a few companies. The first to seriously implement this idea was Findory, later followed by Thoof, Tiinker, Persai, and probably others I’m not aware of. As of this writing, only Persai is still in business. Apparently, personalized news feeds aren’t terribly profitable. Why they’re not a commercial hit is a whole article in itself, so I won’t go into it now. However, before I admitted to myself that this project was doomed to failure, I decided to implement a few components to get a better feel for how the system would work. This is a review of a few interesting things I learned along the way.

TechCrunch writes about PeopleJar

The site’s most powerful feature is its robust search function, which allows users to search for others using many criteria. After creating a search, users can choose to have the site persistently monitor for any matches in the future.

Robert Pasarella writes about Tools for the Equity Research Toolbox

One of the paramount abilities of a good analyst is to spot trends early and realize their potential impact on a company or industry. What analysts are usually searching for is any hint of weakness or strength in competitive advantage. Sometimes the smallest trends start in the local newspapers. Google News makes locating those topics and stories much easier.

If you pair Google News with the enhanced filtering ability of Yahoo! Pipes, and your favorite feed reader; you can create some worthwhile tools that help your trend seeking abilities.

Here is an example I’ve been working on as part of a wider range of investment ideas on Oil.

My first approach was to set up a search in Google News that highlighted anytime OIL was in the title of a story. You can do that with the ‘allintitle’ operator and since I wanted US based sources I added the ‘location’ operator with USA as the source- it looks like this in the Google search window.
http://news.google.com/news?hl=en&ned=us&q=allintitle:oil+location:USA&ie=UTF-8&scoring=n

To see more useful operators check out the Google Cheat Sheet

There are choices on the page to make this search into an RSS feed. Clicking a link on the page will create a feed url in either RSS 2.0 or Atom. You can then take that feed and do further refining in Yahoo! Pipes. I like to create a broad search from Google News and then apply a layer of filters in Pipes for key terms that I think are important. Once I have configured Pipes to my liking it becomes a feed for my RSS reader. I also created a pipe that looks at the opinion and editorial feeds from certain newspapers. Those in the analyst community will recognize this technique as a kin to using Google alerts. Using RSS is the better mousetrap and it doesn’t clog you mailbox.

Cull Web Content With Alerts, by By KATHERINE BOEHRET, Wall Street Journal

For years, I’ve used Google Alerts as a way of keeping track of myself online. If my name is mentioned in a blog or if this column appears on the Web, such as on the site of a newspaper that syndicates it, a Google Alert sends me an email about it. Google Alerts can work for you to find a variety of things, such as telling you if a video of a favorite band popped up online or that a blogger posted something about last night’s episode of “Mad Men.”

In about a month, Google will begin delivering these alerts to users via feeds, as well as emails. Google certainly isn’t alone in the alerts arena, as Yahoo, Microsoft and AOL are also players. This week I tried two small companies that recently joined the mission to help users find the Web content using alerts.

I tried Alerts.com and Yotify.com, and found worthwhile features in both. While Google Alerts does a good job of finding search terms in news, blogs and videos, Alerts.com and Yotify use forms that are a cinch to fill out and let you pinpoint your searches.

Web 3.0: Not Yet by Navdeep Manaktala from New Delhi,

Major aspects of Web 2.0–vertical search, commerce, community, professional and user-generated content–have been panning out nicely, albeit slowly. But personalization, my cornerstone concept for Web 3.0, languishes.

I do believe, however, that we are going to move toward a more personalized and satisfying user experience within the next decade. After all, we went from Web 1.0 in 1995 to Web 2.0 in 2005. Only three years into Web 2.0, perhaps it is natural that we stay here and fine-tune for another five to seven years, before the real breakthrough innovations can come about and usher in Web 3.0.

Struggling to Evade the E-Mail Tsunami by RANDALL STROSS of the New York Times,

E-MAIL has become the bane of some people’s professional lives. Michael Arrington, the founder of TechCrunch, a blog covering new Internet companies, last month stared balefully at his inbox, with 2,433 unread e-mail messages, not counting 721 messages awaiting his attention in Facebook. Mr. Arrington might be tempted to purge his inbox and start afresh – the phrase “e-mail bankruptcy” has been with us since at least 2002. But he declares e-mail bankruptcy regularly, to no avail. New messages swiftly replace those that are deleted unread.
When Mr. Arrington wrote a post about the persistent problem of e-mail overload and the opportunity for an entrepreneur to devise a solution, almost 200 comments were posted within two days. Some start-up companies were mentioned favorably, like ClearContext (sorts Outlohok inbox messages by imputed importance), Xobni (offers a full communications history within Outlook for every sender, as well as very fast searching), Boxbe (restricts incoming e-mail if the sender is not known), and RapidReader (displays e-mail messages, a single word at a time, for accelerated reading speeds that can reach up to 950 words a minute). But none of these services really eliminates the problem of e-mail overload because none helps us prepare replies. And a recurring theme in many comments was that Mr. Arrington was blind to the simplest solution: a secretary.

When Mr. Arrington wrote a post about the persistent problem of e-mail overload and the opportunity for an entrepreneur to devise a solution, almost 200 comments were posted within two days. Some start-up companies were mentioned favorably, like ClearContext (sorts Outlook inbox messages by imputed importance), Xobni (offers a full communications history within Outlook for every sender, as well as very fast searching), Boxbe (restricts incoming e-mail if the sender is not known), and RapidReader (displays e-mail messages, a single word at a time, for accelerated reading speeds that can reach up to 950 words a minute).

Sarah Perez of ReadWrite Web proposes five solutions to email overload,

  • get it done
  • 4-hour work week
  • email as sms
  • folders/rules
  • email bankruptcy

In Email Hell, Ross Mayfield writes in Forbes

E-mail overload is the leading cause of preventable productivity loss in organizations today. Basex Research recently estimated that businesses lose $650 billion annually in productivity due to unnecessary e-mail interruptions. And the average number of corporate e-mails sent and received per person per day are expected to reach over 228 by 2010.

Clint Boulton writes in “Study: Collaboration Overload Costs U.S. $588B a Year “

“Information Overload: We Have Met the Enemy and He is Us,” authored by Basex analysts Jonathan B. Spira and David M. Goldes and released Dec. 19, claims that interruptions from phone calls, e-mails and instant messages eat up 28 percent of a knowledge worker’s work day, resulting in 28 billion hours of lost productivity a year. The $588 billion figure assumes a salary of $21 per hour for knowledge workers.
The addition of new collaboration layers force the technologies into untenable competitive positions, with phone calls, e-mails, instant messaging and blog-reading all vying for workers’ time.
For example, a user who has started relying on instant messaging to communicate may not comb through his or her e-mail with the same diligence. Or, a workgroup may add a wiki to communicate with coworkers, adding another layer of collaboration and therefore another interruption source that takes users away from their primary tasks.
Beyond the interruptions and competitive pressure, the different modes of collaboration have created more locations through which people can store data. This makes it harder for users to find information, prompting users to “reinvent the wheel because information cannot be found,” Basex said.
Basex’ conclusion is that the more information we have, the more we generate, making it harder to manage.

In “The threat from within,” COL. PETER R. MARKSTEINER writes,

If a technological or biological weapon were devised that could render tens of thousands of Defense Department knowledge workers incapable of focusing their attention on cognitive tasks for more than 10 minutes at a time, joint military doctrine would clearly define the weapon as a threat to national security.
Indeed, according to the principles of network attack under Joint Publication 3-13, “Information Operations (IO),” anything that degrades or denies information or the way information is processed and acted upon constitutes an IO threat. That same publication cautions military leaders to be ever-vigilant in protecting against evolving technologically based threats. Yet throughout the Defense Department and the federal government, the inefficient and undisciplined use of technology by the very people technology was supposed to benefit is degrading the quality of decision-making and hobbling the cognitive dimension of the information environment.
We all receive too much e-mail. According to the Radacati Research Group, roughly 541 million knowledge workers worldwide rely on e-mail to conduct business, with corporate users sending and receiving an average of 133 messages per day – and rising. While no open-source studies address how the Defense Department’s e-mail volume compares to corporate users’, my own anecdotal experience and that of legions of colleagues suggests a striking similarity. Without fail, they report struggling every day to keep up with an e-mail inbox bloated with either poorly organized slivers of useful data points that must be sifted like needles from stacks of nonvalue-adding informational hay or messages that are completely unrelated to any mission-furthering purpose.
E-mail is a poor tool for communicating complex ideas. Text-only communication, or “lean media,” as it is referred to by researchers who study the comparatively new field of computer mediated communication, lacks the nonverbal cues, such as facial expression, body language, vocal tone and tempo, that inform richer means of communication. Moreover, aside from its qualitative shortcomings and viral-like reproductive capacity, a growing body of research suggests e-mail’s interruptive nature is perhaps the most pressing threat to decision-making in the cognitive dimension.

In the Future of Search Won’t Be Incremental, Adam DuVander writes,

Personalization isn’t only coming, it’s here. Sign in to your Google account and you can activate it. Prepare to be underwhelmed. But even if it were as Carrasco describes, privacy concerns would stop personalized search from being adopted until the benefits were undeniable. It would take a radical shift.

When Google came along, it provided something that had never been seen before: good search results. Unlike all the other search engines, Google’s top few slots had what we were looking for. And it provided them fast.

It was a much easier time to make big changes. Someone has to make us realize that Google’s results are as antiquated as Yahoo and Excite were in the late 90s. A change in interface might be the most likely innovation.

Sphere which was acquired by AOL News displays articles “related” to the content of the page currently viewed by the user, and now powers over 100,000 sites including many major news outlets like Wall Street Journal, Time, Reuters, etc. This is also the back-end service (see content widget) used to generate “possibly related posts” on WordPress.

Sphere’s founder explains why they created it,

“We founded Sphere with a mission to make contextually relevant connections between all forms of content (mainstream media articles, archived articles, videos, blogs, photos, ads) that enable the reader to go deep on topics of interest,” wrote Conrad.

At the time of its acquisition, Sphere reaches a large number of webpages

Sphere’s third-party network includes more than 50,000 content publishers and blogs and is live on an average of more than 2 billion article pages across the web every month.*

Om Malik writes about Spheres original concept for Blog Search,

The way Sphere works is a combination of many tracks. Lets use an example say of what else, Broadband. The look for blogs that write about broadband, (including those with broadband in the title of the blog) to create a short list. If I am linking to someone who is also a broadband blogger, and vice-versa, Sphere puts a lot of value on that relationship. The fact is most of us broadband bloggers tend to debate with each others. Think Blog Rank, Instead of Google’s Page Rank. The company has also taken a few steps to out-smart the spammers, and tend to push what seems like spam-blog way down the page. Not censuring but bringing up relevant content first. They have pronoun checker. Too many I’s could mean a personal blog, with less focused information. That has an impact on how the results show up on the page.

John Batelle on how Sphere works,

It pays attention to the ecology of relationships between blogs, for example, and it gives a higher weighted value to links that have more authority. This will insure, for example, that when a Searchblog author goes off topic and rants about, say, Jet Blue, that that author’s rant will probably not rank as high for “Jet Blue” as would a reputable blogger who regularly writes about travel, even if that Searchblog author has a lot of high-PageRank links into his site. Sphere also looks at metadata about a blog to inform its ranking – how often does the author post, how long are the posts, how many links on average does a post get? Sphere surfaces this information in its UI, I have to say, it was something to see that each Searchblog post gets an average of 21 links to it. Cool!

Melodee Patterson writes,

Last week I was on vacation without Internet access. Now that REALLY slowed down my infomaniac impulses! I had to settle for flipping through the stack of magazines that I had brought with me, and reading the ebooks that I had downloaded previously and stored on my hard drive. (Thank God I had thought ahead!)

While on vacation, I did have time to think about how much time I did waste on aimless browsing and unfocused research. So I created a list of changes that I’m going to implement this week. That includes:

  • removing any Google alerts that have not provided themselves useful over the past 6 months
  • deleting any RSS feeds that have not added to my knowledge or imagination
  • hitting “unsubscribe” to ezines that I don’t really read
  • using a timer to keep my Internet rovings to 15 minutes (unfortunately, it has a 7 minute snooze button)

Posted in Artificial Intelligence, Internet, Statistical Natural Language Processing, Technology, Trust | 4 Comments »

Phrases, natural language processing, and artificial intelligence

July 3, 2008

Michael Arrington’s long interview with Barney Pell and the Microsoft product manager for search after announcing the Powerset acquisition gives hints as to the Powerset’s capabilities that MSFT search needs most vis-a-vis what Google already has incorporated into their search. A lot of it seems to fall into the “phrase” statistics category instead of “linguistic” understanding category — see below but also see this discussion about Powerset and other players like Hakia in semantic search. Given the performance of today’s mainstream search engines from Google/MSFT/Yahoo, I’m beginning to think that until the day we have artificial intelligence, phrases (bigrams, trigrams, etc) and related statistical techniques will remain the silver bullet for general purpose web search.

The Powerset/MSFT folks have a BHAG (big-hairy-audacious-goal) and it’s hard to fault them for that. The technical challenge that Powerset is running up against is the same sort of wall (i.e. steep learning curve) that shows up in modern computer vision and image recognition applied to things like image-search engines or face-recognition in photo collections. The point here is that the technology works in well-defined problem domains or with sufficient training/hints, and it is also massively parallelizable on today’s low-cost server farms for specific tasks — it either fails miserably or has degraded/unreliable performance when the software/code is not programmed to deal with increased scene or corpus complexity, i.e. combinatorial explosion of inputs that arises with diverse features. This happens when the scene or corpus is noisy and when scene/corpus features are ambiguous or new — exactly the situations where human beings still excel in making sense of the world through judgments, and why humans still write software. For example, see these papers from Stanford computer science class projects (1 and 2) that show how unreliable the face-recognition step can be for auto-tagging Facebook photo collections and in-turn how training can be used to achieve “100% Accuracy in Automatic Face Recognition.”

First of Powerset’s capabilities that MSFT wants is word-sense disambiguation — both inside user queries and in the corpus text — to produce more precise search results. Thhe example he gives is “framed.” Two attached papers listed below from the 1990s talk about two different approaches using statistical collocations (Yarowsky) and latent semantic indexing (Schuetze)… this is something MSFT may not be doing yet today, but Google is doing using perhaps Bayesian word clusters and other techniques.

RN: I think everything that Barney said was right on. I think you see search engines including Live Search and also Google and Yahoo are starting to do more work on this matching not exactly what the user entered but it is usually limited to very simple things. So now all of us do some expansion of abbreviations or expansion of acronyms. If you type “NYC” in a search engine these days, in the last couple years, it understands that it means the same thing as New York. These are very very simple rules based things, and no one understands that bark has one meaning if it about a tree and a different meaning if it is about a dog. Or an example that someone gave the other day was the question of “was so and so framed.” And framed could mean a framed picture or it could mean set-up for criminal activity that did not occur, and so on. And you have to actually understand something of it is a person’s name then it applies to one sense of the word framed if it is not then it doesn’t. So one of the things that Powerset brings that is unique is the ability to apply their search technology to the query to the user’s search in ways that are beyond just the simple pluralization or adding an “-ing” is that Powerset also looks at the document, it looks at the words that are on a web page and this is actually very important. If you look at just the users query, what you have available to you to figure what they are talking about are three words four words five words, maybe even less. That can give you certain hints. If you look at a web page that has hundreds or thousands of words on it you have a lot more information you can use if you understand it linguistically to tell what its about, what kind of quieries it should match and what kind of quieries it shouldn’t match. And Powerset is fairly unique in applying this technology in the index on a fairly large scale already and with Microsoft’s investment and long term commitment we can scale this out even further, an apply it even more of the web, not just the wikipedia content they have thus far.

Second is the potential to support complex but realistic queries in the future — they suggest that this can replace vertical search engines which seems a bit far out.

RN: I think what Peter Norwig is saying has some degree of accuracy and that he is also ignoring some things. So, just for normal queries, queries that are not phrased as questions, there is a lot of linguistic structure. If someone types in a query that is “2 bedroom apartments, under 1000 dollars, within a mile of Portero Hill.” That query is loaded with linguistic content. And that’s a realistic query. That is the type of thing that customers actually want to find on the web. Today there is a sort of helplessness, where customers know that certain queries are too complicated, and they wont even issure them to a search engine. They will go to some deep vertical search engine where they can enter different data into different boxes.

The “2 bedroom apartments, under 1000 dollars, within a mile of Portero Hill” example is interesting, since it requires logic and can’t be solved just using related keywords. Even with access to infinite compute cycles I don’t expect any general purpose search engine applied to the web corpus can do this query reliably today, including Powerset. For verticals like real-estate or travel, we will still need things like Craig’s list or Orbitz/Mobissimo that list classified ads and or build on top of user submissions, metadata, and targeted crawling.
Compared to phrases, linguistic analysis helps provide a better answer to the question of whether a given web page matches the user’s intended query or not. It certainly has the potential to improve recall through more accurate query interpretation and retrieval of web pages from the search engine’s index. By itself linguistic analysis does not improve the precision of search results (by re-ranking) in the way that hyper-link analysis did in Google’s PageRank and other web search engines.
The third use case is query expansion to related keywords using an automatically-generated thesaurus — which I thought Google is already doing.

nce; that is not really an area that is that interesting. But some of these more complex queries really are. For example, shrub vs. tree. If I do a search for decorative shrubs for my yard, and the ideal web page has small decorative trees for my garden, it really should have matched that page and brought it up as a good result. But today Google won’t do it, Yahoo won’t do it, and Live won’t do it. So even in these normal queries there is a lot of value in the linguistics.

I didn’t understand this “decorative shrubs” example. When I try it out on Google I find a result that seems relevant to me. Furthermore using the “~” option in front of any word, Google allows you to search using “related” keywords. If you try “decorative ~shrubs” you get results containing “decorative plants” as well.

Fourth, it seems 5% of queries are linguistic queries today (versus phrase queries), and they expect that fraction to expand as MSFT’s search engine capabilities mature to incorporate linguistic understanding of the query — mostly related to facts or solving problems.

BP: Let me answer the question of does anybody actually search this way. The answer is yes, people do this. It isn’t the most common mode, but we do see that probably 5% of queries are natural language queries. These are not all queries that are phrased in complete sentences, but they are queries where the customer has issued something that has some sort of linguistic structure. Almost any query with a preposition: X and Y, A near B, attribute A of Y, etc. Those things are loaded with linguistic structure…. RN: I have a list of some natural language queries in front of me. Can we just show you some queries that our customers have actually sent to us and are random examples. The first person to see the dark side of the moon. How to get a credit card in Malaysia. Enabling system restore in group policy on domain controller. Timeline of Nvidia. How to measure for draperies. What is the difference between Mrs. and women’s sizes? Does my baby have acid reflux? I could just go on and on and I. These fit in the category that we’ve labeled that match about five percent of queries and they’re really just cases where the customer can’t think of a simpler way to express it.

Finally, the people use case seems to be one of Powerset’s most visible strengths…

BP: We return answers. We actually synthesize, so if you were to say, “What did Tom Cruise star in,” you actually get not just the movies, but the cover art for the different movies. It synthesizes multiple pieces of information to give you a whole different kind of presentation. Or, if you were just to say, “Bill Gates” you’d be given an automatically generated profile of Bill Gates, pulled across many, many articles. It’s no longer just about 10 links, although we can certainly do more relevant job (and will) of the blue links, and a better job of presenting those links.

A friend of mine tried getting answers via Google, and got quick (< 5 minutes for all the searches) and relatively accurate answers without much hassle using phrases:

Those are examples that illustrate why phrases and statistical approaches are not only “good enough” for searching the web corpus, and but they have a very high “signal-to-noise” compared to all other machine search techniques that became available to us in the last 20 years.

The root issue seems to me to be that human symbolic and image comprehension is still well beyond today’s programmable machine capabilities — which has nothing to do with compute power or specific interface/modality (language or vision) — it has much more to do with human intelligence itself that resides inside the cerebral cortex and other parts of the central nervous system. I’d be willing to bet that to achieve their BHAG and make the real quantum leaps from today’s search technology that they advertise, Powerset + MSFT will need the equivalent of artificial intelligence, not just natural language processing (or better image processing in the case of scenes).

Powerset is not the only semantic search engine. Besides Hakia, there is Cognition Technologies — a company founded by a UCLA professor — that also has their own wikipedia search http://WIKIPEDIA.cognition.com. This white paper compares their results to Powerset, query by query, showing how Cognition’s precision is higher and most of the time recall is much lower (more relevant vs unrelated hits) — of course this is a test designed by Cognition, and therefore a display of their own strengths. In one example the show Cognition returning exactly one result and Powerset returning 259. Also see Top-Down and Bottom-Up Semantics by Cognition’s CTO. Cognition’s other white paper compares them query-by-query to Google, Yahoo, Verity, and Autonomy — see section VI. In these tests, only Google has an observable “statistical” boost but doesn’t quite work as well as Cognition — more precision and higher relevant recall. In summary, Cognition achieves these results using the semantic processing and natural language parsing technology described here,

Cognition’s Semantic NLP Understands:

  • Word stems – the roots of words;
  • Words/Phrases – with individual meanings of ambiguous words and phrases listed out;
  • The morphological properties of each word/phrase, e.g., what type of plural does it take, what type of past tense, how does it combine with affixes like “re” and “ation”;
  • How to disambiguate word senses – This allows Cognition’s technology to pick the correct word meaning of ambiguous words in context;
  • The synonym relations between word meanings;
  • The ontological relations between word meanings; one can think of this as a hierarchical grouping of meanings or a gigantic “family tree of English” with mothers, daughters, and cousins;
  • The syntactic and semantic properties of words. This is particularly useful with verbs, for example. Cognition encodes the types of objects different verb meanings can occur with.

Also see Q-Go, a European NLP technology company for customer service applications that partners with MSFT,

The intelligent, linguistic technology-based software that Q-go develops makes sure that customers receive a direct, immediate answer to all of their questions from a company’s website. Online, in their own language, and with at least the same comprehensiveness and quality as the answers provided by call centres. Not only is this easier for organisations, it’s also faster and cheaper. Q-go’s headquarters are located in Amsterdam, and the company has four local offices in Barcelona, Madrid, Frankfurt and Zurich.

A post by Don Dodge (MSFT) reveals that the Powerset’s “semantic rules” can be applied to MSFT’s existing index of the web and therefore likely to be as much about word word co-occurence statistics + clusters as it is using linguistic logic to gain an understanding from the corpus. The example is also illustrative of how Powerset breaks down a natural language query, and the post goes onto explain how Powerset may also be useful in vertical search applications…

Powerset is using linguistics and (NLP) to better understand the meaning and context of search queries. But the real power of Powerset is applied to the search index, not the query. The index of billions of web pages is indexed in the traditional way. The big difference is in the post processing of the index. They analyze the indexed pages for “semantics”, context, meaning, similar words, and categories. They add all of this contextual meta data to the search index so that search queries can find better results.

Who is the best ballplayer of all time? Powerset breaks this query down very carefully using linguistic ontologies and all sorts of proprietary rules. For example, they know that “ballplayer” can mean Sports. Sports can be separated into categories that involve a “Ball”. Things like baseball, basketball, soccer, and football. Note that soccer does not include the word ball, yet Powerset knows this is a sport that includes a ball.

Powerset knows that “ballplayer” can mean an individual player of a sport that includes a ball. They know that “best of all time” means history, not time in the clock sense.

Finally, an alternate theory that says Powerset was acquired for contextual ad placement, not search.

For the foreseeable future, phrases and statistical approaches will probably continue to deliver the greatest signal-to-noise-ratio for machine indexing of web content in the absence of a breakthrough in artificial intelligence. The evidence is not 100% conclusive, but during the last several weeks I’ve accumulated research papers to support this hypothesis…

Also some papers on techniques/tools for extracting keyphrases and performance evaluations that I came across…

  • Google made n-gram data available for all 5-grams (or smaller) appearing more than 40 times on the web, “All Our N-gram are Belong to You”
  • Open-Calais by Reuters can analyze for named entities, facts, and events and their API
  • The chapter “Natural Language Tools” in pages 149-171 in “Advanced Perl Programming” by Simon Cozens (O’Reilly) — you get a very “quick & dirty” introduction to a number of natural language processing concepts and ways to implement and play around with them. Although Perl has many natural language processing tools, the Cozens book cuts to the chase, explains which are the easiest tools to use, and shows you how to use them.
  • “Coherent Keyphrase Extraction via Web Mining” discusses four variants of the KEA key-phrase extraction algorithms – two baseline and two enhanced – and evaluates their performance on a set of computer science and physics research papers. The enhanced versions use statistical association based on web term/phrase frequencies to “weed out” key-phrases that are unlikely to be associated with the text, and they improve on the baseline by generating a higher percentage of key-phrases that match author-generated key-phrases.
  • “Narrative Text Classification for Automatic Key Phrase Extraction in Web Document Corpora” compares three key-phrase extraction algorithms (TF.IDF, KEA, and Keyterm) for summarizing web-pages on websites. They find that Keyterm performs best in tests with users. KEA is a keyphrase extraction package, and another one is Lingpipe.
  • Paper from MSFT Research for clustering algorithms to find user-verifable “related” phrases within search results — could in theory be applied to any list of documents, “Learning to Cluster Web Search Results”
  • Autonomy’s Clustering for Enterprise.

Posted in Artificial Intelligence, Internet, Neuroscience, People, Science, Technology | 6 Comments »

Human-Like Memory Capabilities

June 18, 2008

Human-Like Memory Capabilities by Scott Fahlman, June 17, 2008

My interpretation of what he is saying is that he is looking to build an artificial memory system that can

  1. build-up new complex concepts/facts from incoming knowledge/information
  2. cross-check any given input against known facts
  3. “route” to the relevant fact(s) in response to any new situation — (i’ve always wondered if there is a connection to routing on a graph).

All this happening automatically and rapidly in real-time by taking advantage of massive parallelism built up from millisecond circuits, just like the human brain does, not needing the GHz circuits of today’s microprocessors.

A friend of mine asked me, but isn’t this what exactly what Google is?

Maybe Google includes a subset of this list. It indexes incoming knowledge (facts) and makes them searchable in response to a human-defined query. Still i see some differences which I outline below… See a related blog post “So What’s the Google End-Game?“about Google and artificial intelligence that quotes the Atlantic Monthly article “Is Google Making Us Stupid?”

First, the ability to specify the query in real-time, in real-life situations. Google or machines can’t do that, only humans can at this point. Second is low search efficiency relative to human memory. Although Google may be the most comprehensive and best search engine in the world today, it still requires a lot of human interpretation to use it and refine queries through multiple searches based on initial search results returned — as an example, I’m picturing all the effort needed to do searches for scientific papers and content. Since we end up having to do many many searches the “search” efficiency is not very high compared to human thought which appears to be near-instantaneous among our store of facts — that too it uses millisecond circuitry compared to GHz microprocessing.

Google search may be a machine, but at the heart of it all are associations and judgments originally created by humans in at least two ways. PageRank uses number and prominence of hyperlinks that point to pages as its metric (collaborative filtering) — the more the better. See “On the Origins of Google”

… the act of linking one page to another required conscious effort, which in turn was evidence of human judgment about the link’s destination.

Another area is Bayesian association of “related” keywords (ex. “nuclear” is related to “radioactive”) based on mining human-generated content. See “Using large data sets”. These associations are input by humans on the web, and merely computed/indexed by Google. Like Google, to some degree it’s possible that people communicate with each other to learn and form their own relevance/judgement. I don’t think that explains 100% of how human memory works.

There must be something else based on a human personal experience with the world — like the way babies learn by putting everything in their mouth — that can bootstrap human memory to turn it into what it ends up becoming. Is it logic, association, or something else? I think that what’s missing in today’s machines memories — Google included.

This sums it up… See page 149, “Advanced Perl Programming,” by Simon Cozens

“Sean Burke, author of Perl and LWP and a professional linguist, once described artificial intelligence as the study of programming situations where you either don’t know what you want to don’t know how to get it.”

Posted in Artificial Intelligence, Internet, Neuroscience, Science, Technology, Uncategorized | 1 Comment »