Text Generation for Instances in the Component Library

1
Text Generation for Instances in the Component
Library

• Ken Barker, Jarred McGinnis
• Monday afternoon

2
Outline

• Motivation
• Why UT is doing text generation
• Interface
• What goes in and what comes out
• Customization
• How to adapt for new classes
• Future Work
• What to do next

3
Motivation

• the point of SHAKEN is to make formal KR
  palatable to normal people, mostly through
  graphical I/O (CMAPs)
• there are many places where even a CMAP is too
  clunky or inappropriate
• English text is still the most efficient way to
  describe some bit of knowledge
• interpreting English is still too hard
• generating English from KM is a little easier

4
Still Not Convinced?

• This?
• (Manny has (instance-of (Person)) (trait
  ((aTrait-Value with (value (lazy))))) (size ((a
  Size-Value with (value ((pair big Person)))))))
• (_Cage44 has (instance-of (Cage))(is-possessed-by
  ((a Hamster with(is-possessed-by (Manny)))))
  (smell ((a Smell-Value with (value (stinky)))))
  (capacity ((a Capacity-Value with (value ((pair
  capacious Cage)))))) (is-inside ((a Room)))
  (is-near ((a Door))))

• or This?
• Mannys hamsters capacious stinky cage is near
  the door and inside the room.

5
Shouldnt This be part of SHAKEN?

• no.
• the lexicon and grammar rules are specific to
  components
• each component must be responsible for its own
  text generation
• any class can define its own text generation
  words, phrases and rules

6
Generic vs. Component-Specific
generic The entity move-intos the
tangible-entity with respect to the
tangible-entity along the spatial-entity.
component-specific The agent moves the object
into the container through the portal.
The entity let-falls the tangible-entity.
The agent lets the object fall.
The tangible-entity is received to the
tangible-entity from the tangible-entity.
The recipient receives the object from the
donor.
7
Outline

• Motivation
• Why UT is doing text generation
• Interface
• What goes in and what comes out
• Customization
• How to adapt for new classes
• Future Work
• What to do next

8
Interface

• every instance of Entity and Event has a slot
  called text-gen
• the text-gen of an Event is a sentence
• a single clause with the Event as the main verb
  and select Event?Entity slots as arguments
• can be redefined by any class
• the text-gen of an Entity is a noun phrase
• the Entity is the head noun, some property values
  are premodifiers, some Entity?Entity slots are
  postmodifiers, determiner may include
  is-possessed-by
• can be redefined
• the text-sentence of an Entity is its text-gen
  reworded as a sentence
• can be redefined

9
The text-gen of Entities

• for named instances just the name!
• KM (the text-gen of Manny)("manny")
• for anonymous instances
• KM (the text-gen of _Cage44)("manny's hamster's
  capacious stinky cage near the door and inside
  the room")

10
The text-sentence of Entities

• for named instances and anonymous instances
• is
• KM (the text-sentence of _Cage44)("manny's
  hamster's capacious stinky cage is near the door
  and inside the room.")
• is
• KM (the text-sentence of _Room74)("The room is
  inside the house.")
• is
• KM (the text-sentence of Manny)("Manny is big
  and lazy.")
• is a
• KM (the text-sentence of _Hamster43)("manny's
  hamster is a hamster.")

11
Heads

• for named instances just the name!
• for anonymous instances
• ( or )
• for very general Entities (E, T-E, I-E, S-E, P)
  is the Role played by the Entity
• otherwise, is one of the instance-of
• KM (make-phrase (the text-head of (a Boy called
  "Sue")))("boy (called Sue)")
• KM (make-phrase (the text-head of (a Entity
  with (plays ((a Container))))))("container")

12
Postmods

• is-part-of, material, content, spatial rels
  (is-above, etc.)
• any class can define its own postmod slots
• for each of these slots that has a value
• the of Self
• KM (the text-slot of is-part-of)(text-is-part-of
  )
• KM (the text-is-part-of of (a Knob))("part of")
• KM (the is-part-of of (a Knob))(_Door189)
• KM (make-phrase (the text-postmods of (a Knob)))
  ("part of the door")

13
Premods

• the properties (color, length, sex, etc.)
• for scalars
• only when the reference class
  is the same as the instance-of of the Entity
• for categoricals
• KM (make-phrase (the text-premods of
  (Manny)))("big lazy")

14
Determiners

• for named instances
• for anonymous instances
• if the Entity is-possessed-by P
• s
• otherwise the or a
• KM (make-phrase (the text-def-determiner of
  (_Hamster43)))("manny's")
• KM (make-phrase (the text-def-determiner of
  (_Cage44)))("manny's hamster's")
• KM (make-phrase (the text-indef-determiner of (a
  Door)))("a")

15
The text-gen of Events

• for active Events
• KM (the text-gen of _Move-Into49)("The agent
  moves the object into the container through the
  portal.")
• for passive Events
• KM (the text-gen of (a Move-Into))("The object
  gets moved into the container through the
  portal.")

16
The of Events

• and defined
  for every Event instance, with both singular and
  plural forms
• KM (the text-gen of (a Break))("The object gets
  broken.")
• KM (the text-gen of (a Break with (agent ((a
  Entity)))))("The agent breaks the object.")
• KM (the text-gen of (a Break with (agent ((a
  Living-Entity) (a Substance)))))("The
  living-entity and the substance break the
  object.")
• KM (the text-gen of (a Break with (object ((a
  Object) (a Bond)))))("The object and the bond
  get broken.")

17
The of Events

• of the agent or instrument (or some
  other role)
• for very general Entities, use the case name
  instead
• KM (the text-gen of (a Move with (agent
  (_Hamster43))))("Manny's hamster moves the
  object.")
• KM (the text-gen of (a Move with (instrument
  ((a Tangible-Entity with (plays ((a
  Vehicle))))))))("The vehicle moves the object.")
• KM (the text-gen of (a Receive with (recipient
  (Manny))))("Manny receives the object.")

18
The of Events

• of the object (or some other role)
• for very general Entities, use the case name
  instead
• if there is no object, use something
• KM (the text-gen of (a Move with (object
  (_Hamster43))))("Manny's hamster moves.")
• KM (the text-gen of (a Move with (instrument
  ((a Tangible-Entity with (plays ((a
  Vehicle))))))))("The vehicle moves the object.")
• KM (the text-gen of (a Produce))("The result
  gets produced from the raw-material.")

19
The of Events

• for some list of slots
• for very general Entities, use the case name
  instead
• KM (the text-gen of (a Store with (base ((a
  Tangible-Entity)))))("The object gets stored in
  the base.")
• KM (the text-gen of (a Withdraw with (base ((a
  Tangible-Entity)))))("The object gets withdrawn
  from the base.")
• KM (the text-gen of (a Enter with (base ((a
  Tangible-Entity)))))
• ("The object enters the container through the
  portal.")
• KM (the text-gen of _Move138)("The wheel moves
  manny's hamster from the cageto the door.")

20
text-particle

• one more thing
• some verbs have particles that are part of the
  verbseq, but may be separated from the main verb
• look up a number ? look it up
• the text-particle slot can appear anywhere among
  the arguments
• KM (the text-gen of (a Knock-Down))("Something
  gets knocked down.")
• KM (the text-gen of (a Knock-Down with (agent
  ((a Entity))) (object ((a Object)))))("The agent
  knocks the object down.")

21
The text-gen of States

• mostly covered by the text-gen of Events
• the is by default keeps
• the is by default is
• the state itself is handled by text-particle
• KM (the text-gen of (a Be-Accessible))("The
  object is accessible.")
• KM (the text-gen of (a Be-Accessible
  with (agent ((a Entity)))))("The agent keeps
  the object accessible.")

22
Outline

• Motivation
• Why UT is doing text generation
• Interface
• What goes in and what comes out
• Customization
• How to adapt for new classes
• Future Work
• What to do next

23
Customizing Text Generation

• the text generation system involves 33 special
  slots that specify the words used in text
  generation and the rules for combining them
• all slots are inherit-with-overrides so any
  subclass can redefine any part of text generation
• all slots are dont-cache-values so when instances
  change their text-gen changes

24
Roll Your Own

• usually the inherited text generation machinery
  is good enough for new classes
• KM (Sneak-In has (superclasses
  (Enter)))(Sneak-In)
• KM (the text-gen of (a Sneak-In with (base ((a
  Theater)))))("The object enters the theater
  through the portal.")

25
Roll Your Own (cont.)

• customizing text generation for a new Event class
  usually involves changes to only a few lexicon
  slots
• KM (every Sneak-In has (text-active-verbseq-sg
  ("sneaks"))
• (text-active-verbseq-pl ("sneak"))
• (text-base ("into")))(Sneak-In)
• KM (the text-gen of (a Sneak-In with (base ((a
  Theater)))))("The object sneaks into the theater
  through the portal.")

26
Roll Your Own (cont.)

• sometimes customization demands more
• KM (Bequeath has (superclasses (Action)))
• KM (every Bequeath has (donor ((a
  Entity))) (recipient ((a Entity))))
• KM (the text-gen of (a Bequeath))("Something
  has bequeath done to it.")
• KM (every Bequeath has (text-active-verbseq-sg
  ("bequeaths")) (text-active-verbseq-pl
  ("bequeath")) (text-passive-verbseq-sg ("is
  bequeathed")) (text-passive-verbseq-pl ("are
  bequeathed")) (text-subj ((the donor of
  Self))) (text-argslots ((seq recipient
  instrument))) (text-recipient
  ("to")) (text-instrument ("in")))
• KM (the text-gen of (a Bequeath with (object
  ((a Money))) (instrument ((a Will)))))("The
  donor bequeaths the money to the object in the
  will.")

27
Outline

• Motivation
• Why UT is doing text generation
• Interface
• What goes in and what comes out
• Customization
• How to adapt for new classes
• Future Work
• What to do next

28
Future Work

• stop working on text gen, dammit!