Badanswer.aiml

April 24, 2005
Dr. Richard S. Wallace


Pandorabots has developed an approach to "learning" new AIML
categories. Their solution is borrowed from the backquoting and macro
mechanisms of, wouldn't you know, Lisp.

For a demo check out the AAA bot at

http://www.alicebot.org/aaa.html

If you don't like the bot's response type "bad answer" and it will
dynamically create a new category (visible only to the individual
client).

We are releasing some AIML so you can experiment with this:


http://www.alicebot.org/aiml/aaa/Badanswer.aiml

This <learn> is different than any previous AIML implementation of
<learn> although it may have some properties in common with some of
them.

These <learn> and <eval> tags are specific to Pandorabots AIML only,
not part of the AIML standard nor any other AIML interpreter.

The <learn> tag contains a complete AIML category. The <category>
contains the usual <pattern>, <template> and optional <that>. The way
the learn tag is defined, nothing inside the new category has to be
quoted, or marked as CDATA or marked up in any way differently than an
ordinary AIML category.

The exception to the rule is provided by the <eval> tag. If you do
want to process something inside the new category, use the <eval> tag.
Notice in the examples given the botmaster used <eval> to create a new
uppercase AIML pattern from a stored predicate value:

<eval>
<uppercase>
<get name="badanswer-input" />
</uppercase>
</eval>


Otherwise, outside of the <eval> tag, everthing inside the <learn> tag
is not evaluated, and the result is stored as a new AIML category.

These features are available for any bots on Pandorabots but obviously
not part of the AIML spec or standard.

Things that are missing are:


The added AIML isn't persisted in the db for a particular customer.
(For that matter neither are predicates.)
The botmaster can't examine the added AIML.
It's not clear how well this will scale with large numbers of clients
(each gets there own mini-graph).

Anyway, feel free to experiment with it.


-----

Updated June, 2005

by John Campbell

---

Badanswer.aiml makes use of the Pandorabots new
<learn> and <eval> tags. 

The version as previously published here cannot
handle multi-sentence inputs, and  
breaks badly on them. My experience has been
that visitors who have used BAD  
ANSWER seem to want to input multi-sentence
answers. 

After some thought I've come up with a modified
version. This makes use of  
Pandorabots extended tags "response" and
"request" to capture multi-sentence inputs  
and responses.  

Once the topic has been set to BADANSWER or
BADANSWER2, all inputs are  captured 
and processed. We need to detect whether the
current input sentence is part  of the 
corrected answer (whether one sentence or more).
We do this by checking the  value 
of the "current bot response" (<response index =
"1" />) to see if the user is still  
responding to "OK, what should I have said". If
so, no processing is done to that  
sentence.  

We then check the value of the "previous bot
response" (<response index = "2" />) to  
see if we have moved on to the next input. If
so, we can pick up previous input in 
full,  whether multi-sentence or not, using
<request index = "1" />. 



 


If you don't want to ask if the response depends
on the context, you can just change  
the line 


<li name="badanswer-that" value="*"> 

to 

<li name="badanswer-that" value="nevercomehere"> 

though there are far better ways to do that. 


This seems to work fine both in training mode
and when published.  

I had previously written a series of categories
for this that worked perfectly in 
training  but not when published! There's
something odd about srais in topics that I 
can't  fathom. If, for example, I change the line  

<li>Yes or No?</li> 

to 

<li>I'll take that as No, then.
<srai>No</srai></li> 

the bot output is the generic substitution for
"No", as if the topic was of no effect, 
but  the learning still happens, so obviously
the processing has taken place within 
the topic. 



John Campbell 
http://www.jesus.org.uk/bot 

For the multi-sentence handling abilities of
this extended BADANSWER, we use a rather unusual
approach to testing the inputs.

First, the conversation:

    Human: hi
    Sylvie: Hello there. 

    Human:  Who is Rich Wallace
    Sylvie: I remember faces better than names.
I would search with Google for Rich Wallace.
 
 
    Human: that's not right. try again.
    Sylvie: Would you like to teach me a new
answer to " Who is Rich Wallace"? Yes or
No?

    Human: yes please
    Sylvie: OK, what should I have said?  

    Human: Rich is a good friend of mine who
tries to please. He is a famous computer
scientist.
    Sylvie: " Rich is a good friend of mine who
tries to please...."? Does this depend on me
having just said, " Hello there."?  

    Human: no
    Sylvie: I'll try to remember to respond, "
Rich is a good friend of mine who tries to
please. He is a famous computer scientist."
when you say, " Who is Rich Wallace"  
        
    Human: good
    Sylvie: All right then.  

    Human: who is rich wallace
    Sylvie: Rich is a good friend of mine who
tries to please. He is a famous computer
scientist.  

This is what we want to achieve. It doesn't
appear quite like this in the log, as the inputs
appear sentence by sentence. You can check it
out by asking Sylvie at
http://www.jesus.org.uk/bot

And now with my comments:

    Human: hi
    Sylvie: Hello there. 

Just usual introductory nonsense. The human now
asks the bot the critical question.

    Human:  Who is Rich Wallace
    Sylvie: I remember faces better than names.
I would search with Google for Rich Wallace.
 

Human doesn't like the answer.
 
    Human: that's not right. try again. 
the first sentence srais to BAD ANSWER, and  we
set the following:
--        <set name="badanswer-input"><input index="2"
/></set> -- Who is Rich Wallace
--         <set name="badanswer-that"><that index="2,1"
/></set> -- hello there
--         <set name="badanswer-that-full"><that
index="2" /></set> -- hello there
--         <set name="topic">badanswer-prelim</set>

and Sylvie responds:
    Sylvie: Would you like to teach me a new
answer to " Who is Rich Wallace"? 

The second sentence ("try again") is unwanted
and is trapped by the badanswer-prelim topic,
appending the answer:
    Sylvie: Yes or No?

Normally this response only appears if the user
fails to reply "Yes" or "No" to the previous
question.

    Human: yes please

This matches <pattern>_ _</pattern> in the
badanswer-prelim topic. This double wildcard
overrules any other match such as <pattern>_
PLEASE</pattern> which would otherwise match
regardless of the topic and confuse the output.
We could process this separately, but as we only
want a response of Yes or No, we just take the
first word. We then change the topic, according
to whether there is a context to the original
question. 

    Sylvie: OK, what should I have said?  
    Human: Rich is a good friend of mine who
tries to please. He is a famous computer
scientist.

The two sentences are processed one-by-one. We
keep track of where we are by looking at the two
previous bot responses.

The first sentence (Rich is a good friend of
mine who tries to please) matches:
-- <pattern>_ _</pattern>
-- <that>OK WHAT SHOULD I HAVE SAID</that>

and reduces to the first word ("Rich") in an
attempt to avoid other pattern matching which
would cause (in this case) "Your polite style is
very nice" to appear in the output.

This then matches:
-- <pattern>_</pattern>
-- <that>OK WHAT SHOULD I HAVE SAID</that>
and to prompt the user we echo:
-- "<set name="badanswer-newresp"><input
index="1" /></set>..."?
    Sylvie: " Rich is a good friend of mine who
tries to please...."? 
 
As badanswer-that is not null, we match:
-- <li name="badanswer-that" value="*">
and add on:
    Sylvie: Does this depend on me having just
said, " Hello there."?  
and set the topic to BADANSWER. Every input will
now be matched by the catch all "_" pattern.

The second sentence (He is a famous computer
scientist) is now processed.

This matches;
-- <topic name="BADANSWER">
-- <pattern>_</pattern>

to help with the tests in this section, we set:
-- <set name="star"><star /></set> -- He is a
famous computer scientist
-- <set name="response1"><response index = "1"
/></set> -- OK WHAT SHOULD I HAVE SAID
-- <set name="response2"><response index = "2"
/></set> -- I REMEMBER FACES BETTER THAN NAMES.
I WOULD SEARCH WITH GOOGLE FOR RICH WALLACE.

This matches the condition:
-- <li name="response1" value="OK, what should I
have said"></li>
which tells us that we are still processing the
answer to that question, and outputs nothing(!).
This could continue for several sentences in
response to one single question, for example,
"What is the Gettysburg address?".


The output for the two sentences taken together,
therefore still looks like:
    Sylvie: Does this depend on me having just
said, " Hello there."?  


Human now responds:
    Human: no

This also matches:
-- <topic name="BADANSWER">
-- <pattern>_</pattern>

we set:
-- <set name="star"><star /></set> -- no
-- <set name="response1"><response index = "1"
/></set> -- DOES THIS DEPEND ON ME HAVING JUST
SAID HELLO THERE
-- <set name="response2"><response index = "2"
/></set> -- OK WHAT SHOULD I HAVE SAID

this matches the condition:
-- <li name="response2" value="OK, what should I
have said">
which tells us that we have finished processing
the responses to that question. We can now set
the correct answer to the previous request,
which was a multi-sentence input:
-- <set name="badanswer-newresp"><request index
= "1" /></set> --  Rich is a good friend of
mine. He is a famous computer scientist.

We then meet a separate condition test:
-- <li name="star" value="no">
and we are virtually finished, having decided
that the "that" is not relevant.


    Sylvie: I'll try to remember to respond, "
Rich is a good friend of mine who tries to
please. He is a famous computer scientist."
when you say, " who is Rich Wallace"  

And proceed to the <learn> section.
After which:
        
    Human: good
    Sylvie: All right then.  

    Human: who is rich wallace
    Sylvie: Rich is a good friend of mine who
tries to please. He is a famous computer
scientist.  

Phew.

John
www.jesus.org.uk/bot

Jesus Fellowship/Jesus Army    "Jesus
People--Loving People"
tel: +44 (0)1327 344500              fax: +44
(0)1327 344521
email: info@jesus.org.uk            
http://www.jesus.org.uk