Life in the startup lane

While there have been many wishes and predictions for 2009, mine is simple, it’s HTML 5.  The adoption of HTML 5 specs by browsers, rendering engines and content publishers in this coming year will make 2009 a good year for me. As I have written in the past (here and here), content extraction is an often overlooked challenge that gets in the way of deriving web content semantics. This is an issue that often gets overlooked but for those of us who are passionate about extracting web content semantics, we understand how much it gets in the way of making much of the good work being done now even better. As we have seen recently, this is not an issue that is challenging only the small players, some of the major applications that rely on content extraction such a Google Alerts are seeing a degradation in content quality as they provide articles that have keyword hits in the navigations bars, ads and other non-content related text on web pages.

HTML 5 had taken steps in specifying how web content (e.g. news story, blog entry) should be represented in a page. The specification has attempted to structure a web page by separating different parts of a web page such as headers, footers, navigation, content etc.  The elements of HTML 5 that will help with content extraction are <section> and <article>.

The <section> element is described in the HTML 5 specification as follows,

“The section element represents a generic document or application section. A section, in this context, is a thematic grouping of content, typically with a header, possibly with a footer.

Examples of sections would be chapters, the various tabbed pages in a tabbed dialog box, or the numbered sections of a thesis. A Web site’s home page could be split into sections for an introduction, news items, contact information.”

Having an HTML element that groups content is very welcome.  The <section> element can be used to contain content such as a news article.  HTML 5 has gone one step further to make this possible by introducing the <article> element which the specification described as follows,

“The article element represents a section of a page that consists of a composition that forms an independent part of a document, page, or site. This could be a forum post, a magazine or newspaper article, a Web log entry, a user-submitted comment, or any other independent item of content.

An article element is “independent” in that its contents could stand alone, for example in syndication. However, the element is still associated with its ancestors; for instance, contact information that applies to a parent body element still covers the article as well.”

 A structured implementation of the <section> and <article> elements by content publishers will go a long way in making content extraction simpler thereby providing for a small step in making web content semantic analysis easier.

The HTML 5 specification has been out there for some time, its time for rendering engines to start implementing some of the new semantic oriented elements in the specification (some rendering engines have already started implementing parts of the specification). 2009 sounds like a good year for rendering engines, content publishers and content generation software to come together and help chart the course for web semantic analysis-based applications.

NOTE: HTML 5 contains other descriptive elements that help with the expression of semantics of textual data. I will get to those in future posts.

A couple of weeks ago, I attended the Yahoo Open Hack Day at the Yahoo Campus in Sunnyvale, CA.  At Open Hack Day, Yahoo opened up all their technologies for a few chosen hackers to play with and evaluate for a weekend.  The technology that I was most interested in was BOSS (Build your Own Search Service). BOSS is “Yahoo!’s open search web services platform”.  Simply put, this means Yahoo has opened up its web index for anyone to use using the BOSS API.  This is unprecedented and opens up a ton of opportunities to advance some of the topics that I have discussed on this blog, primarily NLP: Unstructured thinking for unstructured data and 2008 Web Search is still in 1979.

As I have said in the past, the goal of the semantic web is still a long ways to be realized. However, rather than wait for every website owner to build semantic web conforming website (or retrofit their past content to be semantic web compliant), we should seek to derive web semantics at the application level using a whole new set of applications, web semantic analysis-based applications. Yahoo’s BOSS can be one of the missing components that pushes the ball forward towards this goal.

Surprisingly, one of the challenges of deriving web semantics is as simple as programmatically identifying and extracting the content from a web page (I have talked about this in a previous post: A case for standardizing blog templates).  Before semantic analysis can be performed on a web page, the proper content must be extracted fom the web page first. As humans, when we look at a web page, we can readily distinguish the “main content” of a web page from  navigation bar, header, links or ads.  This is not so easy for computer programs to accomplish.  At Filtrbox, we have developed algorithms to accomplish this with a very high success rate only because we have devoted time and resources into the algorithms because they are core to our business.  Other application developers wishing to leverage web content semantics may not have the time and resources to build such algorithms because that is not core to their business. This is where Yahoo BOSS comes into the picture. We know that Yahoo has built its massive Web index by indexing the “main content” extracted from web pages.  Yahoo has  invested time and resources to solve the content extraction problem. In addition, they have built a massive infrastructure to index and store web content.  Therefore, instead of re-inventing the wheel, developers of applications that leverage web semantics can take advantage of Yahoo’s content extraction through the Yahoo BOSS API. However, Yahoo needs to open up a little more for this to be possible.

Here is where Yahoo needs to open up: Although Yahoo currently performs content extraction and content indexing, unfortunately the Yahoo BOSS API is not geared towards applications that analyze web data semantics.  The Yahoo BOSS API in its current form is geared towards web searches.  It is keyword query-based and returns at least TITLE, URL and ABSTRACT/EXCERPT.  Unfortunately, to move towards web semantic analysis-based applications, the ABSTRACT/EXCERPT alone is not enough.  Instead, the Yahoo BOSS API should return the WHOLE “main content” (not links,ads and navigation etc) of a web page.  Returning the whole content enables applications to perform semantic analysis on the data from millions of web pages that is stored in Yahoo’s web index, thereby adding value to the data and moving the ball forward towards unlocking the hidden value in web data using web semantic analysis-based applications.

(NOTE: This article was originally published at my old blog)

One of the questions that I am often asked is how Filtrbox is different from traditional RSS readers and aggregators.  The following are the major differences:

Closed Search Domain vs. Open Search Domain

When using traditional RSS aggregators, the user supplies the list of RSS feeds. This means that the domain of information gathered by a traditional RSS reader/aggregator is limited to the RSS feeds that are known to the user.  I call this a closed search domain. However, in an environment such the one we have today where thousands of new content sources are being created on a daily basis and anyone can potentially become a publisher, it is unrealistic to put the burden on the user to keep up with the thousands of new content sources that are sprouting up each day.  Filtrbox takes this burdensome responsibility away from the user and discovers the new content sources for the user because Filtrbox’s search domain covers all the new content sources. I call this an open search domain. The user can also add RSS feeds to the search domain, thereby guaranteeing that their RSS feeds of interest are searched. This approach leads to the user discovering new content sources.

Publisher centric vs. Content centric

Traditional RSS readers/aggregators present to the user all the content that is published by a specific publisher regardless of whether the user is interested in the content or not. Thus, the traditional RSS readers/aggregators implement a publisher centric information consumption model. On the other hand, Filtrbox implements a content centric information consumption model.  Rather than deliver to the user all the content published by a specific publisher, whether its relevant or not, Filtrbox allows the user to filter for the content that they are interested in from ANY publisher by providing contextual keywords. The content centric model implemented by Filtrbox greatly reduces information overload because each piece of content is examined and filtered for contextual relevance before it is delivered to the user.

No filtering vs. Contextual relevance filtering

As indicated above, traditional RSS aggregators do not filter the content.  All content published by a publisher in the user’s closed search domain is delivered to the user regardless of whether it is relevant or not.  Filtrbox applies algorithms that filter content from an open search domain of publishers for contextual relevance.  Filtrbox uses multiple factors to determine the contextual relevance of content and assigns a score called FiltrRank.  The most important feature of the algorithm is that the contextual relevance algorithm learns from a Filtrbox user’s implicit interests and applies the implicit interest to future contextual relevance filtering. This means that the content delivered to the user is content that that specific user is interested in and not content other people are interested in.  Contextual relevance filtering plays a large part in the reduction of information overload.

Beyond RSS

 
One of our favorite past times at Filtrbox is figuring out fun but useful things to do with our technology.  So in an effort to showcase our robust content filtering technology, we decided to put together FREE widgets that can be used to track news on each member of Team USA as well as the individual sports in which Team USA is competing.  

For example, if you only want to follow Lopez Lomong, you can set up your widget to show news about Lopez only or if you care about Women Gymnastics only, you can set up your widget to show you news about Women Gymnastics only.   In addition, you can set up different combinations of athletes and sports.  Most olympic content consist of every bit of news about the Olympics and you have to do the filtering for the news that you are interested in. At Filtrbox, we do the filtering for you.

We have created two types of widgets, a blog widget to embed on your blog and a desktop widget that runs on your desktop.  Both widgets can be found here.

My favorite widget is the desktop widget and because its my favorite, it has an additional special page for itself here.

Download the widgets and enjoy the Olympics.

Java + LAMP Developer

Join a dynamic, growing software company in Boulder, Colorado
Basic requirements are:

* Solid experience with Java and LAMP
* System administration skills (a plus)
* Working knowledge of Information Retrieval and/or NLP (a plus)
* Must be energetic, motivated and creative

Please send your resume to TOM AT FILTRBOX DOT COM

NOTE: Prima Donna, high maintenance Rockstar developers, please do NOT bother sending your resumes!!!!

This is a pic of the new “Hop 2 Chautauqua” route map that I took this morning at the bus stop on Pearl and 23rd.

On Thursday (04/24/2008 ) last week, I had the privilege of talking to Dr. Jim Martin’s Natural Language Processing (NLP) graduate class, at the University of Colorado at Boulder, about the work that we are  doing at Filtrbox and the role that current NLP students will play in the future of information technology.  This blog post is the basis of my message to the class.

As I have written before, the problem that we face today is how to harness the data that is available on the web so that we can apply meaningful interpretation to it using applications.  This problem is rooted in the assumption that the data that is stored on the web is “unstructured”.  Unlike the majority of the data processed by applications today which is stored in some form of a structure e.g. a relational database, the data on the web is not so, as its is perceived as discrete pieces of data scattered all over the web.

I told the class that part of what I am doing at Filtrbox is an attempt to prove that the data on the web is not as “unstructured” as we may think today.  Within that data, there is a lot of structure, relationship and general interconnectedness no matter how “discrete” we may think it is.  With effective mining of the data and good applications, we can apply interpretation to the data and produce meaningful information.  However, we are still far from applications that can apply effective interpretive meaning on this data.  The reason for this is that we have to address the problem of information retrieval (IR) first before we can get to the writing of applications. 

To recognize where we are today on the continuum of web data information retreival and applications; a look at the evolution of enterprise applications gives us a great analogy:

Enterprise applications are where they are today primarily because they have a structured data storage model (Relational Database or RDB) and a standard access model (Structured Query Language or SQL).  Before there were enterprise applications that we know today, there were only RDBs and SQL.  While RDB work dates back to the 1960s, the RDBs that the majority is familiar with today had their beginnings in the 1970s.  The first (or widely believed to be) commercially available implementation of RDB+SQL was Oracle, then known as Relational Software, in 1979. This provided the ability to query an RDB for data using SQL but no applications as we know them today.  Analogizing this with the web, this is where we are today. We can go on Google or our favorite RSS readers (RDB analogy) and query for web data using a weak REST API or search form (SQL analogy) but we have no applications comparative to what is in enterprise today to interpret that data.  So simply put, today we are where enterprise applications were in 1979.

My message to the class was that applications like Filtrbox are starting to barely scratch the surface with respect to the implementing of applications on top of web data.  That is because, although its 2008, we are still in 1979.  The stumbling block is the perception of the “unstructured” nature of web data. Today’s NLP students will play a large role tomorrow in identifying and establishing structure in the “unstructured” web data in order to move us beyond 1979.

At Filtrbox we are looking for a Flex developer to join our team.  If you meet the requirements below, send a resume to: TOM AT FILTRBOX DOT COM.

*Working knowledge of Flex 
*Solid experience creating Web 2.0 UIs 
*Actionscript 2 or 3 (must be recent) 
*Flash 8 or 9 (must be recent) 
*RoR/PHP/CSS/javascript/java skills a plus 
*Must be energetic, motivated and creative 

Last night, Filtrbox commandeered “The Bunker” at Techstars (thanks to David Cohen) for the first Filtrbox Pizzabox Bug-Bash.  We invited Boulder locals to come in and help us test Filtrbox as well as provide us feedback on the product thus far. The event was a success and I would like to thank all those who were in attendance.  The feedback that we received from testers was great.   Look out for more information about this event on the Filtrbox blog. We had an awesome evening of fun, pizza and beer; here are some pictures from last night:

(I took copious notes during TechStars 2007. I am opening up my notebook and sharing them with aspiring entrepreneurs.   I am going to serialize my notes on this blog.  These are my RAW notes, so sometimes people spoke too fast or were inaudible but I tried to get the gist of what they were saying. There is very little editing to these notes.)

The following questions were addressed during one of the early TechStars panels:

1) What kills most startups?

• Surprise!! Surprise!! Not making money is not usually the big issue that causes failures unless you don’t have a vision
• Team dynamic issues – startup failures are mostly caused by founding team friction
• Companies fail due to execution failures. Execution failures are still team issues that can be categorized as follows: 

1.  Team dysfunction issues
2.  Team poor performance issues

1 and 2. are a “chicken and egg” situation

• Do not be afraid to address team issues head on, solve them and remove the problem
• Once you have a team issue problem that threatens your startup, re-adjust what you are doing or join another team (None of the original TechStars team members changed teams)
• Beware of meandering, where after several weeks you are not getting anywhere. Address and re-adjust immediately because you risk team members losing passion because you are not getting anywhere

 2)  “Getting acquired” as a business model

• Getting acquired is not a business model.  It’s a WISH!!!
• Concentrate on building a business that has compelling value

 3) The “style” of a startup

• You have the permission to create your own identity
• Have an attitude
• Have a style
• Develop a style for your startup and yourself and work it (America’s Next Top Startup, anyone???) all the way through

 MORE TO COME …..