ChatScript Training

by Bruce Wilcox

Agenda - What and Who of ChatScript

Section 0 – Introduction

What and Who of ChatScript
Section 1 – Fundamentals of Scripting
- Overview of Meaning / ChatScript
- Patterns in depth
- Topics, Output, Data, Functions
Section 2 – How does the engine work

What is ChatScript?

My personal NL research project (Open Source)
NL toolkit pipeline
Dialog manager
Expert system for understanding meaning
Expert system for planning

Who uses ChatScript

Hundreds of people in 25+ countries
Basic of NL tech for some companies:
Sensory, Inc (device control)
YourMd (medical diagnosis)
ToyTalk (own rewrite for children's software/toys)
AISoy & Engineered Arts (robot control & personality)
Kore, Inc (business bots)

What have we built with CS?

Conversational chatbots
Ben + Meg (ESL conversation)
K9 (for a robotic dog toy)
Angela 18-year old female cat (60M+ downloads)
Bodo (Star Wars alien demo)
Big Bad Wolf & Pig (interactive stage play demo for kids with asthma)
Kora (control over business bots)
Morgan (Director Morgan Spurlock' clone)
Rose + Suzette ( Loebner winning chatbots)
Dark Design ARG (multiple human bots + 1 parrot)
Pearl (receptionist to experts)

What else have we/others built

Amazon improved product search demo
Wiki demo indexing sentences in Wikipedia
Omni command statistics analysis
Genie commanded basic phone abilities
NL controlled data retrieval from Dow Jones Database
Patient simulator for doctor training (Ohio State University)
Medical diagnosis app (YourMd)

Agenda - Overview of Meaning / ChatScript

Section 0 – Introduction
Section 1 Fundamentals of Scripting
- Overview of Meaning / ChatScript
- Patterns in depth
- Topics, Output, Data, Functions
Section 2 How does the engine work

Fundmentals of Meaning

No intrinsic meaning. Language is agreement.
Consists of:
- Words, grammar, punctuation
- Context
- topic of conversation – the bat flew out of his hands
- prior sentence in volley: I like apples. Do you?
- Prior sentence in prior volley – Do you like apples? Yes.
- Words in nearby sentences – pronoun resolution
- Idioms/Phrases - :) what do you do?

Fundamentals of ChatScript Script

Rules
Topics
Keywords
Concepts
Comments
- # for user comments
- #! for ChatScript directive comments
- ##<< and ##>> for block comments

Rules

The basis for all meaning detection and action.
Fundamentally an IF/THEN statement.
4 parts:
- kind – what types of sentences does it reply to
- label – way for other rules to refer to it
- pattern – IF test for allowing output to execute
- output – THEN part for taking action

u: MYRULE ( I love meat) So do I.

Kinds of rules

Responders – react to user input
- ?: - responds to questions from user
- s: - responds to statements from user
- u: - responds to union of questions and statements
Gambits – program says when it has sente

  t: Have you ever had a pet?

Rejoinders – based on what we just said to user

a: (yes) What kind of pet?
  b: (dog) I like dogs.
  b: (cat) I prefer dogs.
a: (no) Pets are fun. You should try having one.

Questions about sentence kinds

What if user omits punctuation?
System infers it from sentence structure, words.
What is a command?
By default a statement. But script can remap them
to be questions: Describe an elephant is like What is an elephant?
What about exclamations?
They are statements. You can match ! in patterns.

Patterns

Patterns detect meaning.
Constrains when the output of a rule can fire.
Patterns are like regular expressions, but for meaning instead of characters.
Patterns consist of tokens separated by spaces rather than letters all jammed together.

Output

Executed if a rule's pattern matches
- Can be simple text

?: ( << you ~like pizza >>) I love pizza.

Can be random text

?: ( << you ~like pizza >>) [Of course][Absolutely] I [love it.][ adore it.]

Text autoformats
- Can be code and text intermix or code only

?: ( << what you age >>) I am ^compute(%year – 1989) years old.

Rule output is discarded if code fails in a rule

Agenda - Patterns in depth

Section 0 – Introduction
Section 1 Fundamentals of Scripting
- Overview of Meaning / ChatScript
- Patterns in depth
- Topics, Output, Data, Functions
Section 2 How does the engine work

Patterns in Depth

Generally done on the current sentence
Tokens are instructions to process, left to right.

#! what is my name
?: ( what is my name) - simple pattern

grab indicated letter if we have a phone number

u: LETTERS($$phone @_10+ _*1 ) - esoteric pattern

Simple Pattern Match

Match words in sentence

u: (I love pizza)

match: I love pizza, I will love pizza

Pattern words cased as they should be

u: ( I love pizza from Pizza Hut)

match: i love pizza from pizza hut, I LOVE PIZZA FROM PIZZA HUT, i LOVE pizza From pizza Hut

Pattern Problem

Linear sequences not scalable.
AIML improved this by:
- eliminate case – all upper case (losing useful data)
- eliminate punctuation (losing useful data)
- adding wildcard * to match 1 or more words

I * LOVE * YOU

AIML Problem

Requires match of entire input sentence.

Needs 4 patterns to match.

  WORD . WORD
  * WORD ... WORD
  WORD ... WORD *
  WORD ... WORD *

2015 AIML's 1 st revision in 20 years. Provides wildcard matching 0 or more words.
```
* WORD ... WORD *
```

Patterns- CS canonical forms

Input creates 2 sentences - original + canonical
Matching looks at both
- nouns: bicycles -> bicycle
- verbs: were -> be
- numbers: two -> 2 2.0 -> 2 two thousand and fifty -> 2050
- determiners: the -> a
- personal pronouns: my -> I him -> he whom -> who

u: (I love pizza)

matches: I loved pizza, me loves pizzas

Patterns - Canonical forms

Can require original form only using '

u: ( 'I 'love 'pizza)

Can require original form if not canonical

u: ( I loved pizzas)

matches: me love pizza, me loved pizzas

Patterns – bags of words

[ ] - find one

u: ( I love [ pizza bacon sausage hamburger])

{ } - maybe find one

u: ( I love {the my} moon )

<< >> - find all in any order anywhere

u: ( << pizza I love >> )

Particularly important. Breaks the straight-jacket of sequence.

Patterns – composed bags

u: ( << I [like love adore] [pizza bacon] >>) So do I.

matches: I love bacon, Pizza is what I adore, I hate pizza but love rocks

u: ( I love [ meat << flesh animal >> ] ) So do I.

matches: I love meat, I love animal flesh, I love the flesh of an animal

Pattern - wildcards

Keeping sequence but not requiring contiguity
* - any number of words (including 0)

u: ( I love * pizza) - matches "I love spicy pepperoni pizza"

but matches I love you and hate pizza – still a flaw with AIML

*1 - 1 word exactly *2, *3 ..

u: ( I love *1 pizza)

matches I love pepperoni pizza

*~1 - 0 or 1 word *~2, *~3 ..

u: ( I *~2 love *~2 pizza) ?? examples?

But I do not love pizza is a problem.

Patterns - exclusions

! - make sure this not seen later in sentence

u: (!not I *~2 love *~2 pizza)
u: (!not !pepperoni I *~2 love *~2 pizza)
u: (![not pepperoni] I *~2 love *~2 pizza)

Patterns- Phrases

Phrase = sequence of 1-5 words in " "

u: ( I * work * "International Business Machines")

How is this different from

u: (I * work * International Business Machines )

The phrase is a single idea, an entity. Your pattern should reflect that. Humans will understand your pattern better. And CS will handle it correctly against a range of tokenization options. Particularly if it involves periods and commas

u: (I * live * “Raleigh, North Carolina”)

Patterns - Concepts

A concept set is a bag of words.

concept: ~like (like adore love lust "be big on" )

Can be used in patterns in place of words. Same canonical rules as patterns, ' or not.

u: (!~negativeWords << I ~like ~nutrient >>)

This rule captures a probable meaning I like some kind of food or drink Can still go wrong. Rules are not perfect.

Patterns - memorization

_ - memorize word of input Auto assigned match variable numbering.

u: ( I * love * _~nutrient) I love _0 too.

Stores original, canonical, and position.

u: ( I * _~like * _~nutrient) I love '_1 too.
u: ( I like _{~pizzaToppings} pizza)

I like pizza -> _0 is null, I like pepperoni pizza -> _0 is pepperoni

Patterns - comparisons

Can test relationships > < == !=

u: ( _~number _0==3) # relatively pointless since u: (3) is equivalent
u: ( _~number _0>0 _0<10)

Can test concept membership ? !?

u: ( _~animals _0?~pets)

Can test bit patterns & ^

u: (_~number _0&1)

Can test existence (not null)

u: ( I love _{pepperoni} pizza _0) – but a pointless pattern here

Patterns - tag-along patterns

Output requests more matching via ^refine(). Acts like a switch statement, only 1 can match.

u: ( I ~like _~nutrient) ^refine()
  a: (_0?~dessert) I love brownies.
  a: (_0?~meat) I like steak.
  a: (_0?~beverage) I drink tea.
  a: () All food is good.

Patterns - esoteric

Be aware:

You can match forwards and/or backwards
You can jump around in the sentence.
You can match partial spellings of words.

u: ( my sign is Sagitt* )

You can hide or add concept meanings

u: ( _coffee ~nutrient) ^unmark(~color _0)

I like coffee ice cream –> coffee is not a color

Force alignment to start or end of sentence

u: ( < ice is nice > ) ice is 1 st word of sentence, nice is last word

Agenda - Topics, Output, Data, Functions

Section 0 - Introduction
Section 1 Fundamentals of Scripting
- Overview of Meaning / ChatScript
- Patterns in depth
- Topics, Output, Data, Functions
Section 2 How does the engine work

Topics

A named collection of rules tried in order Can have keywords to find or invoked by name Has properties like don't erase, allow repeat

topic: ~funerals repeat (die death burial funeral)
?: ( will you ever die ) Not soon I hope
t: have you ever been to a funeral?
  a: (yes) Did you enjoy it?
t: I have. It was boring.

Use relevant keyword (! prepositions/pronouns)

Data - variables

Match variables from memorization like _0

User variables like $var

$var = _0 -> single $ are permanent
$$tmp = 20 -> double $$ last for the volley only

Read-only system variables like %time

Data

All data is text.
Digit text autoconverts to numbers as needed.
Active Strings ^"The answer is $answer."

Convert on use into simple text Is not autoformatted

Data - facts

A triple of relationship information

(you age 23)
("The Wind and the Willows" author "Kenneth Grahame")

Represents internals using MEMBER and IS

(black member ~colors) - CS representation of a concept
(Laborador~1 is dog~2) - Wordnet relationship of word meanings

Can query for facts given one or more fields

Facts- Tables

Can define tables of data to process into fact.
Processing happens at build time
Efficient scripting- all elaborate patterns in common topic

table: ^favetable(~foods)
chocolate candy    "I love Cadbury's Flake chocolate bar."
Easter    candy    "I love marshmallow bunnies."
[salad _] dressing "I prefer sweet dressings to sour ones."
pizza     topping  "My favorite pizza topping is pepperoni because I like meat."
comfort   food     "My favorite comfort food is is a Gruyere cheese toastie."
_         pizza    "My favorite pizza is pepperoni because I like pizza with meat toppings."

Data - JSON

Classic JSON data used in web calls
Represented internally as facts
Functions to create, query, etc.

Functions

Classic functions with arguments start with ^

$$tmp = ^pick(~animals)

Call by reference
Variety of predefined functions
Can define user functions for patterns or output
Most useful are:

^gambit(~topic), ^respond(~topic)
^refine(), ^reuse(label)

Functions vs Rules vs Topics

Executes script vs Matches patterns, Functions Rules vs Rules Script (via If)
Can be called - Functions (with args) Rules (0 args)

So how are they different? They barely are.

Rules are separately labelled
Topics can be invoked by association (keyword)
Rules stop when output is generated
Functions & Rules execute to completion

Std Engine Control Flow

A control topic manages engine execution.
Below default engine behavior can be altered.
Input volley is processed sentence by sentence
Rule execution is top to bottom in a topic
Rules can invoke topics or other rules
New rules tried until user output is generated.
Rules erase after generating user output
Rules decline to repeat output close together

Agenda - What is the NL pipeline

Section 0 – Introduction
Section 1 Fundamentals of Scripting
Section 2 How does the engine work
- What is the NL pipeline
- How does the engine work

Natural Language (NL) Pipeline

Process characters into words/sentences
Apply standard transformations
- adjust case ( i->I )British → American
- expand contractions texting → American
- spell correction dialog act conversion
- remove end punctuation interjection splitting
- merge dates, numbers, proper nouns
- substitutions (FB -> Facebook, 1am → 1 am)
- separation (5'11” → 5 foot 11 inch, Bob's → Bob 's)
- remove noise (_anonymous strangerv -> stranger)

NL Pipeline

Part-of-Speech tagging and parsing
Dual stream of original and canonical words
Mark words with what concepts and where

All stored in dictionary for hash lookup.

Can use :prepare to see results or :trace

Now ready to execute script.

Agenda - How does the engine work

Section 0 – Introduction
Section 1 Fundamentals of Scripting
Section 2 How does the engine work
- What is the NL pipeline
- How does the engine work

CS Design Goals

Low memory – runnable locally in cellphone
High speed – runnable on weak processors
Concise – rules take few characters
Precise – can represent specific meanings
General – beyond chat, general NL tool
Commercial grade
- scaleable
- script testing and analytic support
- Well documented

Engine sequencing

Startup load dictionary, layer 0, layer 1
Read user input (username + botname + input)
Read in user topic file
Run NL pipeline
Run Script
Write changed user topic file back
Send output to user

Dictionary

Every word, phrase, fact value, concept name, function name, etc stored in hashtable
Lower & uppercase words have same hash
Uppercase words use next bucket up
Enables case-based lookups of same word
New words from user go into reserved bucket
Are ripped back at end of volley
Every word has lists of facts using it as S V O
Enables query by field value
Facts also have lists of facts using it as S V O

Marking

Dictionary entry for every word, phrase, concept
Words have timestamp + loc. in sentence.
Words have lists of facts they participate in.
Marking takes original/canonical word
Fills in location in sentence
Propagates from facts in which word is subject
using MEMBER & IS to object of fact and marks that
Propagate restricted by POS or Wordnet meaning
Pattern test can then look up word for status

Memory Management

No dynamic malloc/free
Mark/release per volley, no interim release
Memory pre-allocated to pools of:
- Dictionary entries
- Facts
- String space
- Buffers (buffer pool array, not mark-release)
Facts & Words inter-locked, unpeel to release

Script Control

Bot definition assigns topics to variables. Only $cs_controlmain is required.

$cs_controlpre – once per volley at start
$cs_prepass - once per sentence
$cs_controlmain – once per sentence
$cs_controlpost – once per volley at end

Simple Control Topic

The Harry control flow processes sentences until a single output is generated. Does this:

See if rejoinder matches
Respond in current topic
Respond in other topics based on keyword.
Respond from keywordless topic.
Quibble sometimes.
Gambit from current topic.
Gambit from other topics based on keyword.
Gambit from any topic.

Error Handling

All functions return 2 things:

Output - is either assigned or dumped to user
- ^myfunction(_0) - result seen in user output
- $var = ^myfunction(_0) - result stored on variable
Error code - tells how to proceed
- NoProblem
- FailRule FailTopic FailInput FailSentence
- EndRule EndTopic EndSentence EndInput
- RetryRule RetryTopic RetrySentence RetryInput

Out-of-band data

ChatScript manages 2 channels of data

Input and output with the user
Input and output with the application
Application data (OOB) in [``] first.

[category: vet specialty: cat] Hi, I'm Rosa.
[callback: 600] My name is Pearl.

Application responsible for handling [``] in output
Script responsible for handling [``] in input
Engine aware of [``], does not apply NL pipeline

`^`Function Support

API routines for:
Manipulating topics
Word manipulation
Marking and parsing
Manipulating facts
Input control
Manipulating JSON
Manipulating numbers
Debugging
Output generation and access Postgres DB use
Controlling flow

ChatScript Server

One thread devoted to CS main engine
Other threads accept and read input
Then queue for engine service
Then handle output and file writing
Can run server on each core, tied to same port.
Server is CPU bound.

`:` Command Support

Testing and display- :trace, :testpattern
Proving your code works- :verify, :regress
Analytics on logs- :trim :abstract
Document reading- :document
Word information- :up, :down, :word, :prepare
Control- :build, :restart, :reset
Topic data- :topicstats, :topicinfo
and lots more

Extensive Documentation

Manuals: Guide to Documentation
Basic + advanced User manuals
System functions (API guide)
JSON, facts
System variables and Engine-defined concepts
Finalizing a bot
Debugging
Papers on history and theory of ChatScript/chat

Files

ChatScript-Training.md

Latest commit

History