Last year I took a course in Natural Language Processing, and was surprised to find that much of the research in the field still uses Lisp. Lisp is a fine language, but Perl can do anything Lisp can do, and Perl source code is much easier to read. I searched the Web for Eliza clones, but I couldn't find any written in Perl. So I wrote one.
The Chatbot::Eliza module is a faithful clone of Weizenbaum's Eliza algorithm. It encapsulates Eliza's behavior within an object.
You can install Chatbot::Eliza just like any other Perl module. Once installed, this little bit of code is all you need to start an interactive Eliza session:
use Chatbot::Eliza; $mybot = new Chatbot::Eliza; $mybot->command_interface;
Let's see what this looks like. If you install the module from the CPAN, save the three lines of code above to a file, and then execute it, here's what the output looks like:
Eliza: Please tell me what's been bothering you. you:
This is an interactive session; type your reply to Eliza after the you: prompt. Listing 1. shows a transcript of a sample run. You can set a few parameters of your Eliza object, such as its name, or a configuration file for it to read:
$myotherbot = new Chatbot::Eliza "Brian", "myscript.txt"; $myotherbot->command_interface;
In this way, you can customize what the chatterbot says by providing your own configuration file. This consists of a list of keywords, decomposition rules, and reassemble rules. If you don't like Eliza's default rules, you can write your own.
For instance, the following lines in myscript.txt would have Eliza (or Brian, as we've called it above) begin with one of two otherworldly greetings, chosen at random:
initial: Greetings Earthling! initial: Take me to your leader!
Chatbot::Eliza contains a default configuration file with default greetings, salutations, 'quit'-equivalents, and rules for determining how Eliza should converse. If you want to want to watch Eliza think, you can turn on the debugging output before you launch your session:
Listing 2 shows part of the same session as Listing 1, with the debugging output turned on.
The Eliza algorithm is actually relatively straightforward. It consists of three steps:
So, in Listing 2, Eliza read the input string He says I am too lazy and found the keyword i. It ran through its entire list of keywords, but i is the only one that matched. The keywords are ranked; if more than one keyword matches, it picks the most salient.
Next, it applied all the decomposition rules for the keyword i, (* i was *, * i am* @happy *, and so on) to see if any matched. One rule did: * i am *. Using this rule, we isolate parts of the input string around i am: the two phrases He says and too lazy.
Next we randomly select a reassemble rule: Is it because you are (2) that you came to me. We use this rule to construct the reply. We replace (2) with the text that matched the second asterisk in the decomposition rule - in our example, the string too lazy. Finally, Eliza replies with, Is it because you are too lazy that you came to me?
The Eliza algorithm has pre- and post-processing steps as well. These handle the transformation of words like I and you; you can read the documentation embedded in the module to learn more.
You can also access all of the module's internal functions from your program. For example, using the transform() method, you can feed a string to Eliza and fetch its response:
$string = "I'm sad."; $response = $mybot->transform( $string );
The Eliza bot is an object, and its configuration data is encapsulated,which means that you can instantiate other Eliza bots, each with their own distinct configurations.
In Listing 3, we create two bots and make them talk to one another. Two bots conversing only produces interesting results if we have clever scripts. Listing 4 shows sample output from the program, where both bots use the default Eliza script. In general, the default Eliza script does not produce any sensible conversation when interacting with itself. (In fairness, people who talk to themselves often don't make much sense either.)
One of the reasons the original Eliza was so successful was Weizenbaum's clever rules and limited domain; the role of a mock Rogerian psychotherapist doesn't require much intelligence, and a relatively simple algorithm can pull it off. But as Weizenbaum found, even this relatively simple program fooled naive users into believing that it understood what they were saying. Scarier still, some people actually felt more comfortable talking to Eliza than a flesh-and-blood psychiatrist.
Listing 5 shows an interactive session, this time with a human deliberately trying to expose Eliza's weaknesses.
The CPAN includes many modules which allow a script to interact easily with resources on the Internet. For example, it is straightforward to write scripts which combine the Chatbot::Eliza module with the Net::IRC module or the CGI module. (The CPAN distribution of Chatbot::Eliza includes a sample CGI script - a Web-based Eliza.)
I am currently working on alternative rulesets. If I come up with one that works reasonably well, I'll include it in future releases of the module.
_ _END_ _
The CMU Artificial Intelligence Repository: http://www.cs.cmu.edu/afs/cs/project/ai-repository/ai/areas/classics/0.html.
Hayden's Java Eliza: http://members.home.com/chayden/eliza/Eliza.html.
Goerlich's Java Eliza: http://philly.cyberloft.com/bgoerlic/eliza.htm.
The white paper on the Loebner Prize competition: http://www.vperson.com/mlm/aaai94.html.
Julia, an Eliza-like chatterbot which roams on TinyMUDs: http://www.vperson.com/mlm/julia.html.
The CYC Project: http://www.cyc.com/.
UMBC AgentWeb: http://www.cs.umbc.edu/agents/.