Natural Language Processing


This brings us to natural language processing (NLP), which has been on computer science's Top Ten List of Holy Grails for more than 40 years. It is a difficult problem, but much progress has been made.

Scale Problems

One of the things that has dogged NLP since the beginning is the problem of scale. The problem is not amenable to demos, prototypes, or even incremental approaches (at least not yet). To get even the most rudimentary behavior out of a system it needs to have a vast amount of information organized in a way that will overcome the chaos that we deal with every day as we process natural language.

Common Sense

One of the less obvious problems has been how hard it is to represent "common sense." After making great strides with artificial intelligence (AI) in the 1970s and 1980s on "hard problems" such as specialist diagnostics and complex scheduling, we expected that some of the "easy" problems were just a "trickle down" away. As chess masters fell to the ever increasing power of computational intensity, wouldn't it be just a matter of time until a computer could engage in conversational English?

Doug Lenat rose to this challenge in the mid 1980s with a project that became known as Cyc (from encyclopedia).[31] The aim of Cyc was to organize a body of knowledge such that a program could reason with common sense comparable to that of a human. The Cyc knowledge base contains more than 1 million concepts (assertions, facts, etc.), which gives an idea of the scale of the problem of achieving interpretation at the level of common sense.

Keywords

Keywords have been the brute force approach to mining information from natural language, but there is no semantic analysis occurring with keywords. Words are notoriously polysemous, which means that they are characterized by having more than one meaning. This plays havoc with keyword searches in general and natural language algorithms that are based on keywords, because there is no way to separate the references to the alternative meanings from the search.

Statistics

One of the promising areas over the last decade has been the use of statistical methods to attack machine interpretation of text. Various approaches (e.g., Bayesian statistics, Markovian analysis, and analysis of word distances and frequency) have helped greatly in reducing ambiguity in language interpretation by paring down the universe of possible meanings into a few that have been found in a reference set. Although this approach has made great strides, it seems to be limited because in the end there is no attempt to determine the semantics of the text.

Symbolic Logic

AI uses what is called symbolic logic. Symbolic logic is based on constructing databases of assertions, which have symbolically substitutable parameters, which are resolved at run time to make inferences. For example, we might say that any physical object x cannot be in the exact same location as another physical object y. We could say that my keyboard is a physical object. So is my computer mouse. Through symbolic logic, we could conclude that my keyboard and mouse cannot be in the exact same location.

Bottom Up

At the same time there has been a great deal of productive work from the bottom up, so to speak. Neural networks function on the principle of getting many small components to collectively work on a problem and generate solutions. However, there are problems with trying to direct this activity, and once directed it is not always apparent how the solution was achieved. Neural networks "grow" a solution of intermediate, interacting nodes that solve a particular problem, without necessarily giving any insight into the mechanisms used to solve the problem.

Agents

Some projects are based on the theories of Marvin Minsky and others, who hypothesize that complex intelligent behavior can be explained by the interaction of a limited number of not very complex "agents" if they are allowed to interact in a sufficiently rich fashion.

We have made considerable progress in each of the aforementioned areas. We have large-scale bodies of knowledge, terms, and concepts that are a necessary prerequisite. We have worked out many of the issues with statistical and bottom-up approaches to solving at least part of the problem, and we are beginning to see agent-based systems that are bringing all this together.

I believe we will soon create systems that can parse English (or any other human language) text and extract a semantically rich understanding of the meaning of the text digested. The understanding may or may not be expressed in a form we would recognize, but if the system can ask and respond to questions about the content of the text in a way that would convince an intelligent person that the system was intelligent (the Turing test), we can say that the system "understood" what it read (at least to some level).

[31]See http://www.cyc.com/ for further information.




Semantics in Business Systems(c) The Savvy Manager's Guide
Semantics in Business Systems: The Savvy Managers Guide (The Savvy Managers Guides)
ISBN: 1558609172
EAN: 2147483647
Year: 2005
Pages: 184
Authors: Dave McComb

flylib.com © 2008-2017.
If you may any questions please contact us: flylib@qtcs.net