Buzz’s Blog: On Web 3.0 and the Semantic Web

Apr 2 2009   5:59AM GMT

Full Text searching: cleaver heuristics for managing large web-based document collections.

Roger King Roger King Profile: Roger King

There is an explosion of technology for supporting sophisticated forms of media on websites and in web applications. In our continuing series on advanced web applications (in particular, as they pertains to the Semantic Web and Web 2.0/3.0), we’ve looked at continuous media, in particular, video and multimedia presentations. But there is a very old form of continuous media, something that is perhaps the dominant media on the Web, and that’s text.

It’s becoming a very major issue in web development.


In this blog entry, we’ll be looking at a particular form of text, called “full text”.

But just what is text to begin with? It’s character-based data, anything we can read.

And what will we want to do with it in next-generation web applications? It’s important to note that more and more vast libraries of documents are being put online. Web applications need to provide far faster and more accurate searches of documents than what we can perform with Google.

Interestingly, a successful technology, called “full text retrieval”, is already in place in the relational database systems that underlie modern web applications. It’s there working for us, and we are likely to not be aware of how clever it is.

It’s also something that should be used much more heavily by web application developers.

Let’s step back and consider three different – and increasingly more sophisticated – ways of managing character data.

Atomic Character Attributes.

First, there is the traditional relational database approach, whereby data is stored as tables made of rows of atomic, fixed sized attributes. By atomic, we mean that each attribute has no internal structure. So, a table of insurance claims might have rows with the following attributes: Claims-ID (an integer), Amount (an integer), Medical_Problem (a fixed length character string), and Subscriber_Name (a fixed length character string). Using SQL, the universal database “query” language, we might look for all rows that contain the name “Fred Jones”. Or, we might search for all rows that have claim numbers that are between 110 and 115.

Essentially, this approach limits us to comparing small strings of data to each other or to fixed values. There are some common extensions that we find in relational databases, such as being able to ask the question to find all rows where the Medical_Problem is something like “broken leg”. Then if a row actually has the value “broken legs”, we would most likely see this row in our results.

Full Text.

Second, there is the ability to search pieces of text according to their natural language (in this case, English) meaning. In this case, we consider the character data to have internal structure, and the values are not considered atomic. Often, these pieces of text are long and of variable length from one row to the next.

It is actually an extension of – but a very dramatic one – of the like operator in SQL.

It is what we call “full text” management or retrieval, and modern relational database management systems like MySQL and Microsoft SQL Server support this. This was seen long ago as a critical extension to relational database technology. Thus, we might rename our Medical_Problem field to Doctor’s_Diagnosis, and allow free form English text in this attribute, as well as allowing the value to be quite long. Then we might search for all rows where the doctor describes “fractures of the lower limbs”. Notice that none of these words might actually appear in the attribute, which might simply refer to “broken legs”.

Natural Language Processing.

This capability would clearly be very powerful, if we could do it right. The problem is that to support it fully, we would need to use highly advanced natural language processing techniques, which are very time consuming to execute, especially on huge databases of large documents. The full text approach tries to simulate true natural language searching in a far less expensive way. The real thing, by the way, might not be all that accurate anyway. Natural language is naturally ambiguous and very subtle.

True natural language searching would be our third way of processing character-based data, by the way. It is not a fully developed technology. And importantly, we usually don’t need anything that fancy.

The Clever Compromise.

So, our middle option, full text searching, is what dominates today – and it is a surprisingly accurate, and efficient, technique that operates on a small set of heuristics. It can transform a dumb webpage where we can only search for small, fixed character strings, to a rich next-generation webpage that can effectively be searched according to its meaning. It allows us to manage very large text documents in web applications – and get us surprisingly close to the semantic power of true natural language searching.

We’re not going to go into a lot of detail here, but here are some of the heuristics that are used in full text search. First, “stemming” and related techniques are used; they conjugate verbs, detect plurals of nouns, and remove prefixes and suffixes. Another technique is to use a “stop list” that lists words that should be ignored, like “the”. The system might also let us specify the “proximity” of words; this refers to how closely specific words should appear in a document. It can also be powerful to include a synonym checker. And the ability to allow for “wild cards”, in particular, letters that may vary in a passage without changing its meaning, can be quite useful. Dictionaries of technical words that pertain to specific domains (like medicine or law) are very useful. We might also provide a feedback capability, whereby users can train full text search engines to be more accurate.

This clearly doesn’t come anywhere near true natural language processing – but it is fast. It will be a growing technology on the new web, with a lot of hidden development, making this heuristic-based technique more and more effective.


We should note that there is a significant up front cost in preparing a document for full text searching: we need to build an index with an entry for every (non-stop) word in the text. Then, when a query is executed, we can look for words in the document by searching the index, instead of searching the full text. If there were no index, the search would be extremely time-consuming.

The Future.

As more and more governmental, educational, medical, and other complex documents become available on the web, advanced full text searching will enable us to search vast databases in a tractable fashion. Even more clever full text retrieval engines will turn dumb, “gotta Google them” document portals into true Web 3.0 and Semantic Web applications.

1  Comment on this Post

There was an error processing your information. Please try again later.
Thanks. We'll let you know when a new response is added.
Send me notifications when other members comment.
  • metaphic
    "But just what is text to begin with? It’s character-based data, anything we can read."
    While text can be understood as a bunch of words, a word cannot be understood as a bunch of letters. A word is a meaningful unit with analogous values, and should not be treated as a number. imho.

    10 pointsBadges:

Forgot Password

No problem! Submit your e-mail address below. We'll send you an e-mail containing your password.

Your password has been sent to:

Share this item with your network: