The Current State of FrameNet

1
The Current State of FrameNet

• CLFNG
• June 26, 2006
• Fillmore

2
The Starting Challenge

• A standard dictionary often does not give us
• the background to the meanings of words
• for specialist vocabulary nobody doubts the need
  of such background
• frame semantics proposes that ALL (content)
  words, including the ones that everybody knows,
  are best understood in terms of their background
  frames
• a description of a words combinatory affordances
  (valences)
• this requires a description of the elements of
  each frame the things that can be relevantly
  talked about within phrases structurally related
  to the word being described and a
  classification of relevant syntactic forms and
  functions
• examples illustrating its common uses
• and these should be determined from examples from
  a large corpus

3
The Starting Solution

• Describe a situation type (frame) that holds for
  one or more words in the language.
• Revenge
• Find the frame elements for this frame.
• Avenger, Offender, Injury,Injured_Party,
  Punishment
• Choose words that belong to the frame.
• avenge, retaliate
• Create valence descriptions as sets of triples
  frame elements, grammatical function, phrase
  type for each word.
• one possibility for avenge
• Av,Subj,NP, Inj,Obj,NP, Pun, Dep,
  PPingby

4
Do this by

• Having trained humans annotate good example
  sentences by assigning FE labels to the syntactic
  dependents of words being described.
• Having automatic processes assign syntactic form
  and function properties to the labeled phrases.
• Having automatic processes derive the valences
  from the annotations.
• Noticing whether the number of syntactic phrases
  found with the head word matches the number of
  frame elements known to belong to the frame, and
  figure out what to make of discrepancies when you
  find them.

5
The Product

• Frame Descriptions
• Lexical Entries specifying
• frame membership
• valences
• access to examples
• Annotated Sentences
• Map of Frame-to-Frame Relations

6
Immediately Emerging Issues

• For our lexicographic purposes we wanted to
  select only simple clear examples.
• First, thats not possible.
• Second, we need to be able to annotate full
  texts.
• There are cases where the major frame in a
  construction was evoked by a syntactic dependent,
  not a syntactic head
• support verbs or prepositions with dependent
  frame-bearing nouns
• say a prayer, swear an oath, make a promise
• transparent nouns selected by their complements
  (unitizers, etc.)
• stick of gum, kernel of rice,
• There are cases of mismatches between the number
  of syntactic constituents we find in a sentence
  and the number of frame elements that we might
  expect to see represented. For example
• null representation of some FEs (for which we
  have stories)
• multiple or discontinuous representations of some
  FEs

7
More Recently Emerging Issues

• Not all FE are of equal status. We proposed
  distinguishing
• Core,
• Peripheral and
• Extrathematic FEs,
• if only because we wanted to say SOMETHING about
  all phrases that were in syntactically
  appropriate positions relative to each target LU.
• We bombed their village yesterday in retaliation.
• The set of frames is not a flat list. There are
  frame-to-frame relations, of various kinds,
  importantly
• inheritance
• composition
• perspective

8
Issues that dont go away

• Public Relations
• Some of our lexicography and NLP friends object
  that since we work one frame at a time,rather
  than one word at a time, itll take forever
  before we can provide the research community with
  any serious kind of systematic polysemy analysis
  hence were useless for WSD.
• Speed
• Well always be slow, but a current grant
  providing collaboration with Adam Killgarriff and
  the use of his WordSketch tool for FrameNet
  vanguarding shows promise. (Vanguarding is what
  you do before you turn material over to the
  annotators selecting correct senses of
  polysemous words, sorting and selecting samples
  that can together show the variety of syntactic
  patterns found with a given word, choosing the
  most relevant collocations, etc.. Collin will
  talk about this.)

9
Continuing Ambitions

1. A speeding up of the whole process (the
   WordSketch connection)
2. Collaborative alignment with WordNet and SUMO
3. Full sample FN analysis of some portions of the
   ANC
4. Some open-source way of enhancing the growth of
   the FrameNet lexicon
5. Success in funding 1-4.

10
Current FN Visitors

• Carlos Subirats
• Jan Scheffsczyk
• Thomas Schmidt
• Kyoko Hirose
• Fran Valverde