Why the monkey ate the banana

1
Why the monkey ate the banana?

• Christer Johansson1, Anders Nøklestad2, Chris
  Biemann3
• 1 Universitetet i Bergen
• 2 Universitetet i Oslo
• 3 Universität Leipzig
• Workshop on Anaphora Resolution, Mjølfjell 2005

2
The monkey ate the banana because

• it was hungry.
• it was ripe.
• it was tea time

3
Ontologies / semantic information

• Need to describe relations between words
• If we had an ontology we would know
• A) that monkeys can be hungry
• B) That bananas cannot be hungry
• C) that bananas can be ripe
• D) that monkeys are not likely to be ripe
• E) that tea time is a situation when you eat
• And so on

4
Benefits

• Well defined
• Formal model
• Supports (automatic) reasoning
• Provides justification for decisions
• EU buzzword, which facilitates funding

5
Problems

• Ontologies are hard to get
• Never complete
• Hard to expand
• Hard to merge
• Hard to prepare for the unexpected
• Not suited for the domain you have
• Usually demands human efforts
• Lots of sweat, coffee and money

6
Our approach

• Accept that we cannot get the perfect ontology.
• We can do without a well defined formal model
• We may not need to justify the automatic
  decisions for some applications
• We can do something useful, fast and with high
  coverage.

7
Using free resources

• If just enough text is analyzed, it might be
  possible
• to identify hungry as a typical property of
  monkeys
• to find ripe as a property of bananas
• by seeing those adjectives modifying the
  respective nouns quite often.

8
Statistical Co-occurrences

• Co-occurrence occurrence of two words A and B
  within a well-defined unit of information or in
  some specific position.
• Significant co-occurrences reflect relations
  between words. The mechanism extracts typical
  associations between words.
• To distinguish co-occurrences from frequency
  counts, a significance measure is used, here
  (log-likelihood)
• sig(A,B) x k log x log k!
• k number of units containing A and B
• n total number of units
• x (ab)/n

9
Finding properties

• Restrict co-occurrence statistics to grammatical
  functions.
• Tags are provided by a Constraint-Grammar Tagger.
• A Norwegian Corpus of 20 Million words was used
  to extract about 4.2 million units of different
  grammatical relations
• Example
• The quick fox eats the pizza. The pizza is
  cold.
• adjective-noun modifiers (AN) quick fox
• verb-subject relations (VS) eat fox
• verb-object relations (VO) eat pizza
• subject-predicate relations (SP) pizza cold
• The co-occurrence statistics cut the noise and
  leave the typical cases

10
What can be done to bananas?

• Co-occurrences only leave specific relations, but
  lose coverage

11
Norwegian girls and boys
12
How to integrate semantic knowledge into
anaphora resolution
13
Memory Based Learning

• Vector space model
• simplest nearest neighbor
• look at a majority of k nearest.

Red is nearest
Majority of 5 nearest are blue
14
Feature Vectors

• Pairs of pronoun and potential antecedent are
  represented as (boolean) feature vectors.
• Does the pair match on surface form?
• Do they have the same
• syntactic function?
• gender?
• number?
• Is the anaphor a distinctly human pronoun (han
  he or hun she) and the potential antecedent
  a proper name (John or Mary)?
• If so, do they also have the same gender?

15
cont.

• Is the antecedent a Subject?
• Is the anaphor reflexive (seg), and the
  antecedent its closest subject?
• Is the antecedent in the current or previous
  sentence?
• Is the antecedent in the current,previous, or
  penultimate sentence?
• Anaphor and antecedent lemmas concatenated

16
Results

• Overall F62
• non-identical antecedent and anaphoraF 40
• (Some errors may have other anaphorical
  relations, without having been.)

17
Automatically detected syntactic/semantic
relations
18
Are the candidate and the anaphor described in
the same way?

• The monkey ate the banana because it was
  hungry.
• Is there a potential antecedent of it that is
  associated with hungry?
• We find out this using corpora and significance
  measures.
• --------------------------------------------------
  ----------------------
• The quick fox ate the slug. It was not satisfied.
• If we find (satisfied ? fox) the chances
  increase that the antecedent is fox.
• If we find (satisfied ? slug) the chances
  increase for slug.

19
What does it do?

• verb-subject relations (VS)
• The dog ate the pill, and it swallowed it whole.
• do dogs swallow?
• do pills swallow?

20
What is done to it?

• verb-object relations (VO)
• The dog ate the pill, and it swallowed it whole.
• Are pills swallowed?
• Are dogs swallowed?

21
Even safe features fail

• He bought Harry a hamburger and they sat down on
  plastic seats to eat them. (J.K. Rowling Harry
  Potter and the Philosophers Stone)
• Judging from grammatical features only, Harry and
  Hagrid would eat plastic seats
• We need world knowledge, and we may get some
  from large corpora

22
Training

• Positive example
• anaphor antecedent
• Negative examples
• anaphor any markable that is closer to the
  anaphor than the actual antecedent.

23
Testing

• Starting from the pronoun,
• search backwards in the text
• until we find a markable that is classified as
  an antecedent for the pronoun.
• For each classification decision If the
  mechanism finds more negative examples in the
  database within the nearest neighbors then the
  answer is no, otherwise yes.

24
Feature Percolation

• If the markable is part of an already established
  coreference chain
• then features are allowed to match on earlier
  markables of that chain.
• We say that feature values percolate towards the
  most recent referent.

25
Results
26
Coverage

• At least one of the semantic features is present
  in about 8 of the positive cases.
• The corpus we used for constructing the features
  were about 20 million running words
• Which means about 4 million (non-unique)
  instances
• Which leaves us about 135 000 (unique)
  statistical collocations.

27
Coverage

• Increase by using much larger corpora
• Work in progress 400 million words from
  Norwegian news wire.

28
When is it useful?

• 1/4 of the cases with semantic features selects
  for a positive instance of co-occurrence.
• Background 1/10 is a positive instance
• Gives something, but not much.
• Hope positive interaction with other features.

29
Discussion
30
Discussion

• It is a general problem that there are so many
  negative examples compared to the positive ones.
• What can be done to reduced the number of
  negative examples?

31
Conclusion
32
Conclusion

• Very simple features account for most of the
  correctly made decisions.
• Hard to come up with features that improve the
  results.
• The syntactic/semantic features improve results,
  although not drastically.

33
Questions?

• THANKS!

34

• War is over
• if you want it