Transcript slides1

Evolució dels sistemes de
diàleg
• Millorar el procés de desenvolupament del
sistema
• Millorar la funcionalitat
– Utilizació en aplicacions més complexes
– Expansió de la cobertura lingüística
– Millora del controlador de diàleg
• Utilització del model del diàleg
• Utilització del model de tasques del sistema
– Integració amb altres modes: multimodalitat
Evolució
Millorar el procés de
desenvolupament del sistema
• Transportables a dominis diferents
• Sistemes i eines per desenvolupar mòduls
comunicatius
– INKA: Interfícies per construir Sistemes Experts
• Utilitza un Llenguate Structurat d’Interfícies
– NL-MENU: Interfícies per consultar bases de dades
– NAT: Interfícies per diferents llenguatges i aplicacions
Evolució
Utilizació en aplicacions més
complexes
Interfícies en LN per sistemes basats en el
coneixement
• El coneixement conceptual
implicat és més complexe
• Es necessiten noves
functionalitats
– Preguntes sobre l’aplicació
• El coneixement lingüístic
necessari és més gran
Incorporació
de la
representació
del domini
Evolució
Expansió de la cobertura
linguística
Eficiència Cobertura Reusabilitat
Basats en templetes
orientats a la tasca
Bona
Pobre
Recursos generals adaptables
a diferents aplicacions
Bona
Rica
Recursos generals
Pobre
Rica
Difícil
Fàcill
Fàcil
Integració amb altres modes:
multimodalitat
• La integració de speech permet una comunicació més
amistosa i noves aplicacions
–VOYAGER (MIT), Office Manager (CMU), MASK
(Multimodal Multimedia Automated Service Kiosk), ATIS (MIT,
CMU), Railtel, Sundial, Verbmobil
• La integració amb menus, gràfics i gest millora la
communicació en moltes aplicacions
–MMI2 (Multimodal Interface for Man Machine Interaction)
–MATIS (Multimodal Airline Travel Information System)
–COMET (Coordinated Multimedia Explanation Testbed),
ALFresco, CUBRICON
The functionality of GISE
GISE: Generador de Interfaces
para Sistemas Expertos
• It supports NL communication with
KBSs
• It automatically adapts
– General linguistic knowledge
• Represented in a Linguistic Ontology
– To application communication tasks
• Represented in a Conceptual Ontology
Aim of the study
GISE, a system for improving NL Interaction
with Knowledge Based Systems
• Reducing the run-time requirements for processing
user interventions
• Guiding the user about the system capabilities
• Reducing the cost of developing the grammar and
lexicon
• The GISE NLI uses:
- An application-restricted grammar and lexicon
- A menu-system
• GISE automatically adapts
- General linguistic knowledge to the application
knowledge represented in a Conceptual Ontology
GISE
The different types of knowledge
involved in the generation process
• Conceptual knowledge: Conceptual Ontology
– Application knowledge appearing in communication
– Communication tasks: general and specific
• Linguistic knowledge:
Linguistic Ontology
– Linguistic structures expressing the communication
tasks
• Control knowledge:
Control Rules
– Controlling the process of relating general linguistic
knowledge to application knowledge
GISE
Obtaining the applicationrestricted linguistic resources
Step 1. Providing the application
domain-specific knowledge
Step 2. Adapting the general communication
tasks to cover application knowledge
Step 3. Adapting general linguistic
knowledge to express the
application communication tasks
The functionality of GISE
Obtaining the applicationrestricted linguistic resources
Data Description
Conceptual Ontology
General knowledge
Application knowledge
Linguistic Ontology
General knowledge
Application lexicon
Control Description
Control rules
Dialogue
system
Application
grammar
Application
lexicon
The architecture of GISE
The Conceptual Ontology
• There are 3 basic entities represented in 3
separated taxonomies
– Concepts
– Attributes
• Describing the concepts
• They are classified according to a syntacicosemantic taxonomy
– Operations
• The communication tasks consist of the
expression of allowed operations over the CO
concepts
Conceptual Ontology
The syntactico-semantic
taxonomy of attributes
• Generalization of the relations between
– Application knowledge in the Conceptual
Ontology
– Linguistic knowledge in the Linguistic
Ontology
• Each class is related to the linguistic
structures expressing the consulting
and filling of the attributes in the class
Conceptual Ontology
The basic attribute taxonomy
• participants :
who_does
who_object
what_object
• being:
is
• possession:
has
• descriptions and relationships between
two or more objects :
of
• related processes: does
Conceptual Ontology
TOP
CONCEPT
ATTRIBUTE
TRANSPORT
lex: (transporte)
departure
arrival
departuretime
arrivaltime
price
TRAIN
BUS
OPERATION
Conceptual Ontology
ATTRIBUTE
OF
OF_QUANTITY
OF_TIME
ARRIVALTIME
lex: (llegar,...)
unit: h/m
OF_COST
DEPARTURETIME
lex: (hora_salida, salir,..)
unit: h/m
PRICE
lex: (precio,..)
unit: Euro
Conceptual Ontology
TOP
CONCEPT
ATTRIBUTE
OF_TIME
OF_COST
OPERATION
MINIMUM_ATTRIBUTE
_VALUE_O
concept
attribute
TRAIN
lex: (tren)
departure
arrival
departuretime
arrivaltime
price
Which <concept_name>
<attribute_verb> first?
Which is the cheapest
<concept_name> ?
Which train departures first?
Which train arrives first?
Which is the cheapest train?
Conceptual Ontology
Operations
• Operations are represented as CO objects
– The attributes describing these objects
represent their parameters and their
preconditions (the conditions that must
hold for an operation to be executed)
• They are classified as Simple or complex
Constructive
Creating a conceptual instance, filling attributes
Consultative
Consulting the value of an instance attribute
The architecture of GISE
The Linguistic Knowledge
• It is organized following the basic principles of the
Nigel grammar
A large systemic functional grammar of English
It is based on Hallidays’s work
It has been used with GUM to generate NL
• It covers the Spanish communication with KBSs
• It is represented as an ontology
The grammar and lexicon generated
• Their size is not large -> Simple parsing
– They cover only the domain communication
tasks
– They incorporate dynamic categories
• They incorporate information from the
Conceptual Ontology -> Simple semantic
interpretation
– In the lexical entries
– In the features augmenting the categories
– In the preconditions associated with the rules
Linguistic Ontology
• Linguistic knowledge is organized in two
dimensions:
– Rank: The scale of the grammatical structures
represented
• Clause
• Group
• Word
– Metafunction: The type of meaning
• Interpersonal: The type of interaction
• Ideational: The propositional meaning and content
• Textual: The information organization
The architecture of GISE
The control rules
• They control the process of adapting
the general linguistic knowledge to
applications
• They establish general relations
between:
Concepts and operations in the CO
CO and LO objects
• Their form is: conditions ----> actions
• They are implemented in PRE (Production
Rules Environment)
The control rules
Adapting the general communication
tasks to cover application knowledge
for each CONCEPT in ONTOLOGY do
generate_CO_operations_ instance_modifying_concept (CONCEPT)
generate_CO_operations_ instance_consulting_concept (CONCEPT)
endfor
The control rules
Adapting general linguistic
knowledge to express the
application communication tasks
for each OPERATION_INSTANCE in ONTOLOGY do
generate_CLAUSE_instances (OPERATION_INSTANCE)
for each ARGUMENT in OPERATION_INSTANCE do
generate_GROUP/WORD_instances (OPERATION_INSTANCE , ARGUMENT)
endfor
endfor
The control rules
The basic set of rules
• It controls the generation of grammars
and lexicons for each application
• It contains 48 rules organized in 8 rulesets
• It covers different types of interfaces
Interfaces supporting descriptions
Interfaces supporting consults
Interfaces supporting consults and descriptions
• It can be enlarged easily
The control rules
A rule of the ruleset creating_instance
(rule cio
ruleset creating_instance
priority 1
control forever
(object ^con ?con ^pcc ?pcc)
--->
(?crinno := (create-name ‘criwno ?con)
(?concrinno := (create-object ?crinno ‘crinno))
(?oparg := (add-slots ?crinno ‘((con ?con)(pcc ?pcc))))
...)
The dialogue system
Dialogue sytem
Menu system
Parser
User
Grammar
Lexicon
Dialogue
Controller
Communication Manager
Conceptual
Ontology
Application
The grammar and lexicon
• They are obtained from the LO objects
• They are represented in the definiteclause grammar (DCG) formalism
because:
– Definite-clause grammars are more
expressive than conventional context-free
grammars
– They can be efficiently parsed
– They are automatically generated
The lexicon
A lexical entry representing the verb ser
String
es
Category
Interpretation
verbser(syn(num(s),tense(p))) (((l,X),(l,Y)),(X,Y))
syntactic
number singular
tense present
The lexicon
A lexical entry representing the concept
ARCHITECT
un_arquitecto
• String
• Category
indefngcon (syn(gen(m),num(s)), sem(con(architect)))
syntactic
gender masculine
number singular
• Semantic Interpretation
semantic
concept architect
architect
The lexicon
Dynamic entries
Representing instances of concepts
Category
pngi(sem(con(person)))
function
instance_of(person)
Representing values of attributes requested
to the user during communication
Category
defngattrof(sem(con(person), attr(name)))
function
name
Representing all possible values of an attribute
defngvalofcause(sem(con(requirementobuild),attr(reasonotbuilt)))
menu(reasonotbuilt)
The lexicon
Dynamic entries
• The number of lexical entries to be
considered is reduced
• They allow the introduction of new values
during communication
• They guide the user to introduce
specialized terms
The parser
• It is based on the Ross version of the
Left-corner algorithm
• It assures there is always a correct
choice to continue from a correct prefix
(prefix correctness)
• It can parse
– A word and predicts the set of all possible
next words
The Dialogue Controller (DC)
• The DC completes and disambiguates the
semantic interpretation of the user request
– The result is a complete specification of an
operation over the Conceptual Ontology
• The DC controls the execution of the
operation
• The DC passes the resulting information to
the interface
The Dialogue Controller
• The DC completes and disambiguates the
semantic interpretation of the user request
using:
– History of dialogue
• Concept and parameters of the previous operations
– The Conceptual Ontology
• The definition of the operation: mandatory
arguments, default values,...
• This process is simple when users build the
requests using the NL options shown in the
screen
– Mistakes and misunderstandings are avoided
Applications of GISE
SIREDOJ, an expert system in law
• Previously its communicative tasks
– were fully integrated with functional tasks
– were based on a set of menus
• Applying GISE improves the
communication:
– Complex concepts can be expressed in one
sentence
– User-initiative dialogues are allowed
– The size of the linguistic resources is not big:
26 grammar rules and 112 lexical entries
Conclusions
• Main contribution:
– Proposing an organization of the
knowledge involved in communication
that improves the obtaining of the
linguistic resources most appropriate
for each application
Conclusions
Proposing a reusable organization
• The Conceptual Ontology
– It provides a general framework for
representing application communication
tasks
– It includes a syntactic-semantic taxonomy
of attributes
• Capturing the relations between
application communication tasks and
their linguistic realization
Conclusions
Proposing a reusable organization
• The Linguistic Ontology
– It is an adaptation of NIGEL grammar for
communication with KBSs in Spanish
• The Control Rules
– They control the process of adapting
linguistic knowledge to each application
– A basic set of rules controls this process for
different types of applications
Conclusions
Improving the NL processing
Using grammars and lexicon restricted to the
application communication tasks
• Their size is not large: The parsing is simple
– Dynamic categories are used
– A menu-system is integrated in the NLI
• They incorporate information from the
Conceptual Ontology: The interpretation is
simple
• In the lexical entries
• In the features augmenting the categories
• In the preconditions associated with the rules
Conclusions
Improving communication and
user satisfaction
• Using an easy and clear language
– Guiding the user about application specific
information
• Using a menu-system to introduce NL
– The user is guided about the system
requirements
– The user can avoid typing sentences
• Tools helping the user are incorporated
into the interface
GIWEB
Interface
User
Grammar
Lexicon
Parser
Conceptual
Ontology
Dialogue
Controller
Wrapper1
Wrapper2
Wrappern
Internet
Source1
Source2
Sourcen