September 2003

I've been meaning to add code corners to the last few newsletters but hadn't found the time. So this newsletter will be all about building the types of AI systems discussed in the last few issues. In other words, this is an issue for those who want to roll up their sleeves and create their own AI systems.

The two types of AI application studied are chat bots (July 2003 newsletter) and ontologies (June 2003 newsletter).

Both are perfect examples of the ideas expressed in the first newsletter (November 2002) about AI applications in general. Both require a knowledge representation language and a reasoning engine that knows how to apply that knowledge.

In the case of chat bots, the knowledge is patterns of inputs and responses that drive a dialog. In the case of ontologies, the knowledge is relationships and attributes of words and concepts.

A chat bot tool provides a way to represent the knowledge and a way to deploy it. An ontology tool does the same.

There are chat bot and ontology tools available, and good ones at that, but they suffer the problems of all AI tools. They make assumptions about how and what types of knowledge can be represented, and further, on how that knowledge is deployed.

If the assumptions fit an application area, that's great. But if not... Well that's a very practical line of reasoning. The simple truth is, its a lot more fun to play with your own knowledge representation and reasoning engine, and, you can get it to do exactly what you want.

The programs can be written in any programming language, but we'll use Prolog in the newsletter because it is particularly versatile for a rapid prototyping and experimentation approach to AI concepts, as well as being very compact and fitting better in a newsletter.

Logic programs have the added benefit that they read much like program specifications and can be used for the design of applications intended for implementation in other languages. So if you don't like Prolog, read along and code in Lisp, Mozart, Java or the language of your choice. For those that don't know Prolog, there are boxes with explanatory notes.

One final note. Design decisions for the programs were always made in favor of clarity, sometimes at the expense of performance, although this will not be an issue until the knowledge base for either program grows to a large size.

As always I like to read your feedback, and for any who request it I can e-mail back the source code developed in this article.

Writing a Chat Bot

Chat bot is the generic name for programs derived from Weizenbaum's original Eliza program, which mimicked a therapist. Eliza was written almost as a joke, to show how mindless pattern-matching rules can be used to generate realistic conversations, thus proving the software was NOT intelligent. Yet chat bots have come closer to passing the Turing test than other types of programs, and Richard Wallace hypothesizes that that is because much human dialog is, in fact, mindless pattern-matching.

Pronoun Reversal

One of the main insights in Eliza was how much can be done simply by changing the grammatical person of a sentence, like 'I' to 'you' and 'mine' to 'yours'. This is called pronoun reversal. We'll start with a very simple, and extremely annoying chat bot that does only that, and adds a question mark at the end.

To represent the knowledge of which words to reverse, we'll use Prolog facts, which are logically the same as relations in a relational database. Here's the few we'll use for testing, in a relation call me_you/2, the /2 indicating there are two arguments.

me_you(me,you).
me_you(i,you).
me_you(my,your).
me_you(am,are).

The choice of the me_you/2 predicate was the first design decision about knowledge representation.

Input/Output

The reasoning engine needs an interface to the user or caller. For the chat bot all that is required is a predicate (like a function) that takes an input string of user text and puts out a response string. A predicate respond with two arguments, the input and the response, provides that service. It is refered to in Prolog as respond/2.

Any application that can call Prolog could be used to provide the dialog interface with the user, and Prolog is certainly a language that can call Prolog. Here's a simple loop for a console version of the program that can also be run in a Prolog listener (interpreter).

main :-
    write('DDJ AI Expert ChatBot Sample'), nl,
    repeat,
    write('> '), read_string(InputString),
    respond(InputString, ResponseString),
    write(ResponseString), nl,
    InputString == `quit`.

A Prolog logic base, or program, is made up of logical clauses that define predicates. This one is simple. The predicate is called main/0 (the /0 indicating no arguments) and has only one clause. After the neck :- symbol are a comma separated list of goals, which typically call other predicates. The most important one here is respond/2, which will do the real work. Logical variables are indicated by an initial upper case letter, such as InputString.

Core Pattern Matcher

How should a sentence be represented internally by the program? Thinking of the sentence as a list of words seems like a good starting point. Proceeding in a top-down manner, here is the implementation of respond/2.

respond(InputString, ResponseString) :-
    string_to_list(InputString, InputWordList),
    swap_person(InputWordList, SwappedWordList),
    form_response(SwappedWordList, ResponseWordList),
    list_to_string(ResponseWordList, ResponseString),
    !.

The !, called cut, is used to turn off Prolog's automatic backtracking search for multiple solutions for a given rule. In this case, we only want one response for each user input.

The first and last goal of respond/2 perform the service of converting strings to and from lists of words. For experimentation, those steps could be simply eliminated, using lists directly for input and output. For final use, those steps should be quite nice, dealing with punctuation and capitalization. For now, we'll make use of simple built-in predicates that do the conversion for us but require everything in lower case. Later nicer versions of these can be implemented.

string_to_list(String, List) :-
    string_tokens(String, List).

list_to_string(List, String) :-
    stringlist_concat(List, ` `, String).

The utility predicates for dealing with strings and lists are not part of the ISO standard for Prolog, so different implementations support different ones. They are all relatively easy to write in Prolog. The two used here, string_tokens/2 and stringlist_concat/3 are part of Amzi! Prolog. You might have to find equivalents for other Prologs, or, simply skip this step and use lists directly for input and output.
Next to implement is the real guts of the program where the me_you/2 knowledge is applied in the first step of creating a response. Recursion is typically used to process lists in Prolog. The recursive rules to replace the pronouns in a list of words are:

boundary condition - The list is empty, we're done.
recursive condition - Take the first word in the list and call swap_word/2 to see what the replacement word is. Put that word in the output list and recursively call the rest of input list to generate the rest of the output list.

swap_person([], []).
swap_person([X|Xs], [Y|Ys]) :-
    swap_word(X,Y),
    !, swap_person(Xs,Ys).

[H|T] is Prolog notation for separating a list into its head (first element), and tail (remainder of the list). This allows for the easy writing of recursive rules where some operation is performed on the head, and the tail is then used as the argument to the next recursion. The empty list is represented by [] and usually signifies the boundary condition for a recursion.

So, for example, when swap_person/2 is called with the input list [i,want,pizza], the pattern [X|Xs] is mapped to it in a process called unification. X becomes i, Xs becomes [want,pizza]. swap_word/2 will unify Y with you and Ys will become what ever is returned by calling swap_person/2 with [want,pizza].

The swap_word/2 rules look in the knowledge base portion of the program to see if the word is one mentioned in the me_you/2 relation. If its not, the word is replaced with itself.

swap_word(X,Y) :- me_you(X,Y).
swap_word(X,Y) :- me_you(Y,X).
swap_word(W,W).

One last bit and we're ready to run. We form the response, for now, by simply adding a question mark on the end.

form_response(In,Out) :-
    append(In, [?], Out).

append/3 is a common list utility in Prolog that appends one list to another. In this case we're adding the list with the single element, ?, to the end of our output list of words. To use a library of common list utilities we need these two lines at the front of the program. We'll use other list utilities later as well. (If you don't have a library, the implementations of various list utility predicates can be readily found and coded directly in the program.)

:- load(list).
:- import(list).

We're now ready to load the program into a Prolog listener and see how it runs.

?- consult(chatbot).
yes
?- main.
DDJ AI Expert ChatBot Sample
> i have problems with my computer
you have problems with your computer ?
> yes
yes ?
> will you help me
will i help you ?
> quit
quit ?
yes

It's already sounding like some people I know.

Adding Response Patterns

The "intelligence" of a chat bot lies in the patterns it is told to recognize. Lists of words and variables can represent a pattern to look for, and lists that have those same variables in them can be used for various possible responses.

There are two types of logical variables in Prolog. Named ones, beginning with an upper case letter, and anonymous ones, represented by an underscore. The named variables can be used to echo parts of the input back, and anonymous ones for indicating we don't care about parts of the input.
Here's two responses for any input that has the adjacent words "blue screen" in it.

response( [_,blue,screen,_],
    [ [why,do,you,use,windows,?],
        [wipe,your,computer,clean,and,install,linux,'.'] ]).

Here's a response pattern for any input with the word "linux" in it.

response( [_,linux,_],
    [ [nobody,supports,linux,'.'] ]).

That should drive them nuts.

Here's a few more response patterns that might come in handy to keep the conversation going. The last one is a catch all that is used if none of the other patterns match. It does the job of echoing the input with a question mark. Remember that the input word lists have already had the pronouns swapped before they are used in these patterns.

response( [yes,_],
    [ [why,?] ]).
response( [no,_],
    [ [why,not,?] ]).
response( [_,you,like,X],
    [ [how,can,you,like,X,?],
    [is,it,strange,to,like,X,?] ]).
response( [_,want,to,X],
    [ [why,would,you,want,to,X,?],
    [you,can,not,X,'.'],
    [is,it,dangerous,to,X,?] ]).
response( [X],
    [ [X,?] ]).

When a named variable appears multiple times in a rule, like X in the above patterns gets a value, all occurances of X automatically have the same value. So when the pattern matcher we will write determines a value of X from the input pattern, then that value will be echoed in the output patterns.

Before looking at the code, here's how these response rules will drive a dialog with the chat bot.

?- main.
DDJ AI Expert ChatBot Sample
> i get a blue screen on my computer
wipe your computer clean and install linux .
> ok, i did that but now i cannot get linux to work.
nobody supports linux .

When there are multiple response patterns to pick from, the engine will pick one at random. Here's another run with the same rules, but a different result.

> i get a blue screen on my computer
why do you use windows ?
> i like the windows operating system
is it strange to like the windows operating system ?
> no
why not ?
> lots of people like windows
lots of people like windows ?
> yes
why ?
> it is easy to use
it is easy to use ?

The Response Pattern Matcher

The tricky part of the pattern matcher is setting a variable to possibly multiple words in the input. For example, in the response pattern [_,want,to,X] X might match a phrase. That phrase is best represented as a list of words as well, so we will have an output list with embedded lists in it. A list utility, flatten/2, will fix that problem.

For example if the input is [i,want,to,learn,to,fly] X will be bound to [learn,to,fly] and the outputs will look like [why,would,you,want,to,[learn,to,fly]]. flatten/2 converts that to [why,would,you,want,to,learn,to,fly].

Here is a new version of form_response/2 that uses the response/2 rules in the knowledge base portion of the application.

form_response(Input,Response) :-
    response(InputPattern, ResponsePatterns),
    match(InputPattern, Input),
    random_elem(ResponsePatterns, ResponsePattern),
    flatten(ResponsePattern, Response).

This rule makes use of Prolog automatic backtracking search to find the right response rule. It first picks a response/2 and then uses match/2 to see if its pattern matches the input. If it doesn't, then Prolog backtracks and tries the next response/2 rule. When it finds one that matches, it picks one of the possible response outputs, flattens the list and returns the response.

Here's the recursive rules that try to match the input word list with a pattern in one of the response rules. It will be called with the input pattern from a rule and its job will be to see if that pattern can be matched with the user's input. Further, if there are variables in the input pattern, then the binding for those variables will be determined as well.

There are two boundary conditions.

  • One is when the the word list for the input pattern is empty, meaning the pattern matcher succeeded.
  • The other is when the only element left in the input pattern is a variable, in which case the pattern also succeeds and the variable is bound to whatever is left in the user input.

There are two recursive cases.

  • When the next element in the input pattern list is a variable, a separate predicate is called to gather words from the user input up until the next specific word in the input pattern is encountered.
  • When the next element is a specific word, then if it matches the user input, recursion continues.

If the input doesn't match, none of the four match/2 rules will succeed and the call to match/2 fails, triggering backtracking search for another response/2.

match([], []).
match([Xs], IWs) :-
    var(Xs),
    !,
    Xs = IWs.
match([Xs,W|PWs], IWs) :-
    var(Xs),
    !,
    fill_var(Xs, W, IWs, IWsLeft),
    match([W|PWs], IWsLeft).
match([W|PWs], [W|IWs]) :-
    !,
    match(PWs, IWs).

fill_var/4 is another recursive predicate that walks the input list looking for the first occurance of a word that matches the word after the variable. It has four arguments:

  • the output list of words, returned to caller;

  • the next word in the input pattern that cause recursion to stop when its found;

  • the user input list being walked, looking for the stop word;

  • the remainder of the input list, returned for the caller for further processing.

    fill_var([], W, [W|IWs], [W|IWs]) :-
    !.
    fill_var([X|Xs], W, [X|IWs], IWsLeft) :-
    fill_var(Xs, W, IWs, IWsLeft).

It is easy in Prolog to test individual predicates such as this one. When the program is loaded in a Prolog listener, any of the predicates can be queried directly.

?- match([a,b,c], [a,b,c]).
yes

?- match([a,X,d], [a,b,c,d]).
X = [b, c] 
yes

?- match([_,like,X], [you,like,to,fly,planes]).
X = [to, fly, planes] 
yes

?- match([_,like,X], [you,want,to,fly,planes]).
no

?- match([a,b,c], [a,r,c]).
no

If you are new to Prolog, the best way to learn to understand code like this is to run it in a debugger, which will provide a trace of Prolog execution and variable bindings. Here is a trace of form_response/2 from one type of Prolog debugger.

DDJ AI Expert ChatBot Sample
> i like pizza
'CALL':user:form_response([you, like, pizza], H132)
'CALL':user:response(H321, H322)
'EXIT':user:response([H388, blue, screen, H394],
[[why, do, you, use, windows, ?],
[wipe, your, computer, clean, and, install, linux, .]])
'CALL':user:match([H388, blue, screen, H394], [you, like, pizza])
'FAIL':user:match([H388, blue, screen, H394], [you, like, pizza])

'REDO':user:response([H388, blue, screen, H394], 
    [[why, do, you, use, windows, ?], 
    [wipe, your, computer, clean, and, install, linux, .]])
'EXIT':user:response([yes, H390], [[why, ?]])
'CALL':user:match([yes, H390], [you, like, pizza])
'FAIL':user:match([yes, H390], [you, like, pizza])

'REDO':user:response([yes, H390], [[why, ?]])
'EXIT':user:response([no, H390], [[why, not, ?]])
'CALL':user:match([no, H390], [you, like, pizza])
'FAIL':user:match([no, H390], [you, like, pizza])

'REDO':user:response([no, H390], [[why, not, ?]])
'EXIT':user:response([H388, you, like, H394], 
    [[how, can, you, like, H394, ?], 
    [is, it, strange, to, like, H394, ?]])
'CALL':user:match([H388, you, like, H394], [you, like, pizza])
'EXIT':user:match(['[]', you, like, [pizza]], [you, like, pizza])
'CALL':user:random_elem([[how, can, you, like, [pizza], ?], 
    [is, it, strange, to, like, [pizza], ?]], H348)
'EXIT':user:random_elem([[how, can, you, like, [pizza], ?], 
    [is, it, strange, to, like, [pizza], ?]], 
    [how, can, you, like, [pizza], ?])
'CALL':user:flatten([how, can, you, like, [pizza], ?], H132)
'EXIT':user:flatten([how, can, you, like, [pizza], ?], 
    [how, can, you, like, pizza, ?])
'EXIT':user:form_response([you, like, pizza], 
    [how, can, you, like, pizza, ?])
how can you like pizza ?
> 

This trace illustrates that the rules are tried in the order they appear. So if a user input might match multiple response patterns, the first one will have priority.

At this point, all the code is here that is needed to run the program and get the results displayed a few paragraphs back.

Enhancements

There are a number of enhancements that can be made to the sample.

Implement better formatting of input and output strings.

Put the knowledge in a separate file, so different sets of pattern matching rules can be used. Let the user can pick from a menu of rule files.

Indexing of response rules by key words in the patterns can be added. So for example, rules that match the word 'can' might be indexed on 'can'. This can be used for greater efficiency if a very large number of patterns are coded, and it can be used to easily let one pattern refer to another for completion.

Memory of past variable bindings can be kept and popped back in from time to time. So we might remember the user liked pizza and if nothing else matches, have a response that uses the memory to ask a question about pizza.

Add facts similar to the me_you/2 facts that can be used to map a variety of input words into a single word used for pattern matching. For example, 'need' might map to 'want' so the rules matching 'want' will also work if the user types 'need'. Also it would be nice if the these facts used lists of words, rather than being restricted to single words.

And of course creating more and more response patterns to make the chat bot more interesting.

Ontology

Tools for ontologies are easy to create and play with in Prolog. As always, first a knowledge representation for the ontology is determined, and then tools to use it are developed.

Knowledge Representation

The entries for the ontology can be stored as property lists, where each word has a property list associated with it. A property list is just a list whose elements are Property = Value. Using this idea for representing ontological knowledge, here's some entries we might want to support an application about food.

word(pizza, [
    kind_of = food,
    contains = cheese,
    contains = tomato ]).
word(broccoli, [
    kind_of = food,
    color = green ]).
word(donut, [
    kind_of = food,
    contains = sugar ]).
word(spaghetti_sauce, [
    kind_of = food,
    contains = tomato ]).

Here's some more entries about people and allergies.

word(joe, [
    instance_of = person,
    allergy = tomato ]).
word(jill, [
    instance_of = person,
    allergy = cheese ]).

Reasoning Engine

Now, armed with typical list utilities we can write a predicate that can answer questions about the properties of words in our ontology.

property(Word, Property, Value) :-
    word(Word, PropertyList),
    member(Property = Value, PropertyList).

Because Prolog is just matching patterns itself, the property/3 predicate works no matter which variables are bound. So it can find properties of a particular word, or words with a particular property.

Here's some things we can do with property/3 in the listener. (The semicolon tells Prolog to look for another solution, so in these queries we find all solutions.)

?- property(pizza, contains, X).
X = cheese ;
X = tomato ;
no

?- property(X, kind_of, food).
X = pizza ;
X = broccoli ;
X = donut ;
X = spaghetti_sauce ;
no

?- property(X, kind_of, food), property(X, color, green).
X = broccoli ;
no

?- findall(X, property(pizza, contains, X), L).
X = H19
L = [cheese, tomato] 
yes

Applications of the Ontology

The ontology we developed can be used to support other applications. The types of information we stored in the small sample can be used, for example, in an expert system that warns people about what foods they can and cannot eat based on allergies.

The rules of that expert system might be written directly in Prolog. Each of these rules finds a person word and a food word and compares the allergy of the person with what the food contains.

can_eat(Person, Food) :-
    property(Person, instance_of, person),
    property(Person, allergy, Allergy),
    property(Food, kind_of, food),
    not property(Food, contains, Allergy).

cannot_eat(Person, Food) :-
    property(Person, instance_of, person),
    property(Person, allergy, Allergy),
    property(Food, kind_of, food),
    property(Food, contains, Allergy).

Note that the ontology has changed the knowledge engineering task for this allergy expert system. Without the ontology, the knowledge would all be stored in the rules. With the ontology, the rules become simple general purpose rules and the knowledge is stored in the ontology.

Here's how these two rules can be used.

?- can_eat(joe, pizza).
no

?- can_eat(joe, donut).
yes

?- can_eat(joe, X).
X = broccoli ;
X = donut ;
no

?- cannot_eat(joe, X).
X = pizza ;
X = spaghetti_sauce ;
no

?- can_eat(X, broccoli).
X = joe ;
X = jill ;
no

Transitivity

Transitivity is an easy enhancement to add to the ontology engine. Transitivity applies for certain properties, such as kind_of. It means that if X kind_of Y and Y kind_of Z then X kind_of Z. We can define certain properties as transitive by specifying it in the knowledge base.

transitive(kind_of).
transitive(contains).

Now if we're looking for a value for a property that is transitive, our engine can recursively call itself.

property(Word, Property, Value) :-
    word(Word, PropertyList),
    member(Property = Value, PropertyList).
property(Word, Property, Value) :-
    transitive(Property),
    word(Word, PropertyList),
    member(Property = Word2, PropertyList),
    property(Word2, Property, Value).

To test it make one small change to the ontology, have pizza contain spaghetti_sauce instead of tomato.

word(pizza, [
    kind_of = food,
    contains = cheese,
    contains = spaghetti_sauce ]).

The test will be to see if joe is still not allowed to eat pizza, as the pizza does not directly contain tomato.

?- property(pizza, contains, X).
X = cheese ;
X = spaghetti_sauce ;
X = tomato ;
no

?- can_eat(joe, pizza).
no

?- cannot_eat(joe, pizza).
yes

Chat Bot with Ontology

We can use the ontology we developed to add more intelligence to the patterns in the chat bot rules. In particular we can specify properties to include in the pattern match. Here's a pattern that recognizes when the user is talking about food.

response( [_,property(W,kind_of,food),_],
    [ [are,you,talking,about,W,because,you,are,hungry,'?'],
    [when,did,you,last,eat,W,'?'] ]).

And here's some to add to the annoying behavior of our technical support.

response( [_,property(W,kind_of,windows),_],
    [ [you,should,dump,W,and,switch,to,unix,'.'] ]).
response( [_,property(W,kind_of,unix),_],
    [ [you,should,dump,W,and,switch,to,windows,'.'] ]).

These last two require the addition of some more information in the ontology.

word(windows, [
    kind_of = operating_system ]).
word(unix, [
    kind_of = operating_system ]).
word(xp, [
    kind_of = windows ]).
word(w2000, [
    kind_of = windows ]).
word(linux, [
    kind_of = unix ]).

And we have to make some small changes to the basic pattern matcher to recognize and handle these property patterns in the rules. The new clauses are indicated in red.

match([], []).
match([Xs], IWs) :-
    var(Xs),
    !,
    Xs = IWs.
match([Xs,W|PWs], IWs) :-
    var(Xs),
    !,
    fill_var(Xs, W, IWs, IWsLeft),
    match([W|PWs], IWsLeft).
match([property(W,P,V)|PWs], [W|IWs]) :-
    property(W,P,V),
    !,
    match(PWs,IWs).
match([W|PWs], [W|IWs]) :-
    !,
    match(PWs, IWs).

fill_var([], W, [W|IWs], [W|IWs]) :-
    !.
fill_var([], property(W,P,V), [W|IWs], [W|IWs]) :-
    property(W,P,V),
    !.
fill_var([X|Xs], W, [X|IWs], IWsLeft) :-
    fill_var(Xs, W, IWs, IWsLeft).

Trying the new pattern rules we get a dialog like this.

?- main.
DDJ AI Expert ChatBot Sample
> i have problmes with xp
you should dump xp and switch to unix .
> ok i just installed linux
you should dump linux and switch to windows .
> you are going in circles
i am going in circles ?
> do you wnat to get a pizza
when did you last eat pizza ?
> yesterday
yesterday ?
> yes i ate pizza yesterday
are you talking about pizza because you are hungry ?

Conferences

The 5th IFAC/CIGR Workshop on Artificial Intelligence in Agriculture
will be held in Cairo on March 8-10 2004. Deadline for submitting
extended abstract is Sept. 30,2003. More Information can be found at
www.claes.sci.eg/aia04