Natural Language Part 2 - Automated Email Response

As we saw in the last edition, an important current area of natural language work in the "real world" is in automated email response by companies such as Brightware, Kana, and eGain. What systems built with these tools do is provide automated response to email within a very short period of time. They go beyond simply replying with an auto-reply promising that a "real person" will get to the message in short order. Instead, they attempt to categorize the message based upon its content, and provide a reasonable response based upon this content. They also route the message to the appropriate human in the organization for responding to the particular category of message. The present state of the art in this form of natural language understanding is essentially to categorize messages, with a high but not perfect degree of accuracy, into one of a large (but still finite) number of categories, and provide a stock response. The envelope in this area of AI is likely to be pushed in the upcoming years, as systems gain the ability to do detailed understanding of messages and provide a tailored, not a stock, response, and accuracy improves to near-perfect. We'll talk about what lies ahead in Natural Language Part 4. The purpose of this edition is to talk about the state of the art at present.

The simplest form of natural language understanding is simply to categorize messages based upon keyword lookup. The automated response engine starts by doing what the search engines do and provides automated response based upon the presence of one of a set of keywords in the email. Unlike with the search engines, however, the keywords are defined not by the user but by the webmaster when the site is set up for automated email response. For example, if a site is providing general information about world history, and it gets an email including the phrase "Otto von Bismarck", then an automated reply is generated directing people to resources about Germany and German history. This fairly simple approach has certain obvious disadvantages. For example, in the case of "Otto von Bismarck", the appearance of that phrase might not indicate an interest in Germany at all but could refer to historical information about retirement plans (Otto von Bismarck introduced the retirement age of 65) or North Dakota (Bismarck is the capital of North Dakota), to cite a couple of examples. However, this type of approach does have one significant advantage, and that is that it is easy to set up: you can just identify a large number of keywords that are important to your organization, and easily categorize emails based upon the presence of one of those keywords. And this is likely to be fairly accurate, at least in as far as it goes.

The next step in developing an automated response tool is to do a more in-depth analysis of the incoming messages. Two approaches for doing so stand out: case-based reasoning and Bayesian networks. The general idea of case-based reasoning is that we put together a series of examples, called cases, of how a particular problem was solved previously, and then when presented with a new problem to solve, we find the "case" from the past which comes closest to the current problem, and use the known solution to that problem to solve the current problem. For automated email response, what we would do is take a large number of previous (or artificial "use case") emails together with the information about how each one was categorized. Then with each new email that comes in, we would categorize it based upon which of the old emails it is most like. To do this, it is necessary to identify as many characteristics of the emails as possible. What are these characteristics? Certainly, keywords as mentioned previously will play an important role. However, unlike with pure keyword lookup, case based reasoning allows us to look at a set of keywords in any given email and to draw conclusions based upon how they are related. Other characteristics of an email could include the sender, the length of the email, whether it has any attachments, whether it appears to be a reply to a previous email, the time of day it was sent, the type of organization it was sent from, etc. One can also look not just at individual keywords, but also whether pairs of keywords appear near each other, and the presence of certain words such as "urgent", "important", and so on. Bayesian networks are more of a probabilistic tool.

'Bayesian networks also look at various characteristics of the email, but rather than categorize emails based upon which old email they look most like, a Bayesian network will attempt to determine the probability that a given new email falls into a given category based upon the experience with old emails. Bayesian networks, in a way, are similar to neural networks in that both types of networks do learning based upon old data. However, whereas neural networks are based upon modeling the human brain, Bayesian networks are based upon probability theory. Bayesian networks will use a lot of the same characteristics that case-based reasoning does in order to categorize emails.

Each approach has its advantages and disadvantages. The major advantage to the case-based reasoning approach is that it is not necessary to have a lot of training data-a human being can get it set up just by providing a number of use cases. The disadvantage is that it doesn't learn so easily. The advantage to the Bayesian networks approach is that it can learn and continue to improve over time, although it requires a number of months worth of sample data to get set up.

Obviously, automated email response is in its infancy, and it is going to improve considerably over the next few years. In Part 4, we will look at the future of automated email response, and natural language in general, in more detail. But first, in Part 3, we will look at voice recognition systems.

More information about case-based reasoning may be found in Case-Based Reasoning. More information about Bayesian networks may be found in Bayesian Theory and Theory and Practice of Bayesian Belief Networks. These resources are general resources on case-based reasoning and Bayesian networks and are not necessarily specific to natural language understanding.

Next edition: Natural Language Part 3 - Voice Recognition.


Home: Ramalila.NET



All copyrights are maintained by respective contributors and may not be reused without permission. Graphics and scripts may not be directly linked to. Site assets copyright © 2000 RamaLila.com and respective authors.
By using this site, you agree to relinquish all liabilities and claims financial or otherwise against RamaLila and its contributors. Visit this site at your own risk.