Shallow Semantics with Shallow Syntax

1
Shallow Semantics with Shallow Syntax

• Nimar S. Arora

2
Shallow Semantics

• Semantics is about constructing a logical form of
  human sentences to extract information, answer
  questions, etc. Ex.
• I ate popcorn ? eat (I, popcorn, t) tltnow
• Shallow semantics is about identifying the key
  players in a sentence and how they interact
• A prerequisite for creating the logical form, but
  only a first step for bigger tasks

3
(No Transcript)
4
The Berkeley FrameNet Project

• Collin F. Baker, Charles J. Fillmore, John B.
  Lowe, 1998
• Frames are an attempt to generalize semantic
  patterns into a hierarchy
• Consist of a target (the predicate) and frame
  elements (arguments)
• However, the roles are attached to the Frame and
  not the target

5
Frame Example

• Duplication Frame
• The teacher was demonstrating on the blackboard ,
  Creator the children Target copying Goal
  onto slates .
• Body_movement
• Agent Scott Target threw Agent, Body_part
  himself Manner with enthusiasm Goal into this
  exacting assignment .

6
The Task

• Pre-segmented constituents
• Given Frame and the target
• Label each constituent with a role
• Ignore null-instantiated elements
• 130K training sentences 13K test
• Parts-of-speech tags are available for training

7
Baseline

• Most Frequent Role 43
• Assigns all frame elements to the most frequent
  role for the frame
• i.e. argmax role P (role frame)
• Most Frequent Role Before or After 62
• Assigns all frame elements before target to one
  role and all those elements after the target to
  another

8
Phrase Type / First Word

• Recall ..Goal onto slates.. Goal into this
  exacting assignment
• The Goal role seems to begin with a preposition
• Similarly Agents, Creators, Inventors seem to
  begin with articles
• Such similarity exists across frames

9
Phrase Type / First Word

• Argmaxrole P(role frame)P(first word role)
• 76
• Duplication The teacher was demonstrating on the
  blackboard Creator ? Original the children
  Target copying Goal onto slates .
• Children cant be copied!
• Somehow need to incorporate P(childrenOriginal)
  vs. P(childrenCreator)

10
Head Word

• Last word doesnt help
• ..Goal to another file specially reserved for
  the purpose.. ? Purpose
• P(purposePurpose) higher than P(purposeGoal)
• Want to compare P(fileGoal) to P(filePurpose)
• First noun in the constituent is the head word or
  None (single word constituents)
• Different from Collins (1999) where prepositions
  are the head of preposition phrases

11
Head Word

• Used part-of-speech (POS) tag labels (BNC and
  PENN style) in training data to build taggers
• Argmaxrole P(role frame)P(first word role)
  P(head word role)
• Problem Too little data
• Duplication And Original ? Goal the layout of
  the walled garden long narrow beds , with
  plants grouped according to their botanic family
  , has been Target copied Place worldwide .
• P(layoutOriginal) 0 but P(layoutGoal) gt 0

12
Hierarchical Counting

• FrameNet provides a hierarchy relating roles in
  frames to super type roles in super type frames
• For example, ManufacturingManufacturer ?
  Intentionally_CreateCreator
• An occurrence of ManufacturingManufacturer can
  also be considered an occurrence of
  Intentionally_CreateCreator and so on up the
  super type hierarchy

13
Hierarchical Counting

• We can no longer condition on role because roles
  dont have a hierarchy
• Need to condition on both frame type and role

• 79
• Best is around 85
• using external resources for clustering nouns
• 80.4 otherwise (complicated interpolation)
• on FrameNet 1.2 which is half of FrameNet1.3

14
Future Approaches

• Deeper Syntactical information like parse trees
• Need more work on head word rules, example
  children s book the head word is book rather
  than children
• Use adjectives and other phrases to better
  discriminate head word sense.
• Example the highlighted block is mistaken to
  be a Place rather than Original
• Use relations other than inheritance between
  frames to enrich counts
• Hierarchical smoothing, i.e. interpolate
  P(headframe, role) with P(headparent frame,
  parent role)

15
Conclusion

• The first word and the head word of a constituent
  can give very good information about semantic
  role
• Head word can be computed by looking at
  part-of-speech tags without parsing
• The Frame rather than the target word is much
  more critical for semantic role labeling
• Hierarchical information about frames and roles
  can help in smoothing the probabilities and
  improve accuracy