FrameNet and Ontologies

1
FrameNet and Ontologies

• Birte Lönneker-Rodman
• International Computer Science Institute
• Berkeley, CA

2
Mapping FrameNet to SUMO

• Align FrameNet Semantic Types with SUMO classes
• expressed in SUO-KIF (SUMO language)
• Also developed semi-automatic, domain-specific
  approach for linking FrameNet frame elements
  (FEs) to SUMO classes
• based on typical fillers of FEs in a particular
  domain
• Provide restricted, ontology-based types on the
  fillers of FEs

Scheffczyk et al. 2006
Jan Scheffczyk, Adam Pease, Michael Ellsworth.
Linking FrameNet to the Suggested Upper Merged
Ontology. In FOIS 2006, Baltimore, Maryland
(USA), 2006.
3
All FrameNet STs mapped

• First state that Semantic Types are a kind of
  SUMO Attributes
• Then, for each FN Semantic Type, find the closest
  SUMO class
• Difficulties are mentioned in comments

Included in FrameNet Distribution 1.3
ST_bindings.kif
4
Example ST-SUMO Mapping

• (subclass SEM_Physical_entity FN_SemanticType)
• (equivalentclass SEM_Physical_entity Object)
• (subclass SEM_Physical_entity Object)
• (partition SEM_Physical_entity SEM_Material
  SEM_Physical_object)
• (documentation SEM_Physical_entity "FN
  everything that is physical (FN distinguishes
  between countable physicals and uncountable
  physicals) SEM_Physical_entity equivalent to SUMO
  Object, i.e. not Process")

5
Frame Element-SUMO Mapping

• Not all Frame Elements are provided with Semantic
  Types
• FrameNet Semantic Types limited
• General Idea
• To use SUMO classes directly (or a combination
  thereof) to constrain Frame Element fillers,
• Find semantic head word of FE filler
• Use WordNet as interface to SUMO classes
• Select best SUMO class(es)

6
Experiments

• A tool was implemented by Scheffczyk et al. in
  2006
• I reused their method on the training data for
  the SemEval-2007 task on Frame Semantic
  Structure Extraction

Semeval task organizers Baker, Ellsworth, Erk
http//framenet.icsi.berkeley.edu/semeval/FSSE.htm
l
7
Full-text FrameNet annotation

• Each target word in a sentence is annotated
• A full semantic parse of the annotated sentence
  can be produced from FrameNet XML using SemEval
  task software

8
Example Semantic Parse
Semantic Head of Message Frame Element filler
(verify) evokes itself a frame, Verification.
Libya then invited the IAEA to verify the
elimination of nuclear weapon related activities
in-country.
9
Statement, FE Speaker, ANC texts

• Frames evoked
• NEperson 5
• Leadership 1
• NEorganization 1
• People 1
• Quantity 1

• SUMO classes mapped (all but NEs)
• SocialRole 3
• Position 3
• Position_at_ 2
• SubjectiveAssessmentAttribute 1
• Human 1
• FamilyGroup 1

Roles and Attributes are not subclasses of Human
in SUMO.
10
Statement, FE Speaker, NTI texts

• Frames evoked
• NEGPE 23
• NEperson 20
• Leadership 19
• NElocation 18
• NEorganization 12
• Aggregate 6
• Source_of_getting 3
• Organization 3
• Expertise 2
• Measure_duration 2
• Businesses 2
• Locale_by_use 1

• SUMO classes mapped
• Position 15
• SocialRole 9
• SubjectiveAssessmentAttribute 9
• Human 9
• origin 8
• Text 6
• part 6
• Group 6
• Proposition 5
• ElementalSubstance 5

e.g. authority, source
11
Intentionally_Act, FE Purpose, NTI

• Frames evoked
• Manufacturing 13
• Purpose 8
• Processing_materials 7
• Cause_to_make_progress 7
• Research 4
• Using 4
• Terrorism 4
• Project 4
• Cognitive_connection 3
• NEorganization 3
• Defend 3
• Being_at_risk 2
• Political_locales 2

• SUMO classes mapped
• Process 55
• SubjectiveAssessmentAttribute 46
• IntentionalProcess 32
• IntentionalPsychologicalProcess 27
• Communication 23
• Increasing 21
• ContentDevelopment 21
• Making 16
• Contest 14

e.g. nation, territory, town
12
Results and More Questions

• Frame Elements could be constrained by Semantic
  Types in terms of Frames or SUMO classes
• Find the best level or combination
• Constraint checking during annotation could
  prevent erroneous manual labeling
• But semantic construals, especially metonymies,
  would have to be taken into account

13
Semantic Construals

• Many metonymies are known as regular
• An organization can stand for the people
  employed.
• A building can stand for the organization it
  hosts.
• A text can stand for its author.
• How can we cover domain-specific and ad-hoc
  construals?
• She emptied the lid back into the bottle
• Should Emptying FE Source be restricted to ST
  Container? A lid is not usually itself a
  container.
• Remains a challenge for resource building and
  automatic semantic labeling

14
Conclusions

• FrameNet is a rich lexical database supported by
  a large body of annotations
• The Frame part of the database has been
  represented as OWL-DL and can be used for
  reasoning
• Different methods for restricting Frame Element
  fillers (arguments) have been investigated
• FrameNet Semantic Types best practice?