a joint model of implicit arguments for nominal predicates matthew gerber and joyce y. chai...
TRANSCRIPT
A Joint Model of Implicit Arguments for Nominal Predicates
Matthew Gerber and Joyce Y. ChaiDepartment of Computer Science
Michigan State UniversityEast Lansing, Michigan, USA
{gerberm2,jchai}@cse.msu.edu
Language &InteractionResearch
Robert BartComputer Science and Engineering
University of WashingtonSeattle, Washington, USA
2
Implicit Arguments
Georgia-Pacific and Nekoosa produce market pulp,
containerboard and white paper. The goods could be
manufactured closer to customers, saving shipping costs.
• What can traditional SRL systems tell us?
3
Implicit Arguments
• What can traditional SRL systems tell us?– Who is the producer?– What is produced?
Georgia-Pacific and Nekoosa produce market pulp,
containerboard and white paper. The goods could be
manufactured closer to customers, saving shipping costs.
4
Implicit Arguments
• What can traditional SRL systems tell us?– Who is the producer?– What is produced?– What is manufactured?
• But that’s not the whole story…– Who is the manufacturer?
Georgia-Pacific and Nekoosa produce market pulp,
containerboard and white paper. The goods could be
manufactured closer to customers, saving shipping costs.
5
Implicit Arguments
• What can traditional SRL systems tell us?– Who is the producer?– What is produced?– What is manufactured?
• But that’s not the whole story…– Who is the manufacturer?– Who ships?
Georgia-Pacific and Nekoosa produce market pulp,
containerboard and white paper. The goods could be
manufactured closer to customers, saving shipping costs.
6
Implicit Arguments
• What can traditional SRL systems tell us?– Who is the producer?– What is produced?– What is manufactured?
• But that’s not the whole story…– Who is the manufacturer?– Who ships what?
Georgia-Pacific and Nekoosa produce market pulp,
containerboard and white paper. The goods could be
manufactured closer to customers, saving shipping costs.
7
Implicit Arguments
• What can traditional SRL systems tell us?– Who is the producer?– What is produced?– What is manufactured?
• But that’s not the whole story…– Who is the manufacturer?– Who ships what to whom? Implicit arguments
Georgia-Pacific and Nekoosa produce market pulp,
containerboard and white paper. The goods could be
manufactured closer to customers, saving shipping costs.
8
Model Formulation (Gerber and Chai, 2010)
• Candidate selection– PropBank/NomBank arguments– Two-sentence candidate window
• Coreference chaining–
• Binary classification function–
c2
c3
c1
)... ,',iarg,shipping|Pr( 31 c
},{' 233 ccc
Georgia-Pacific and Nekoosa produce market pulp,
containerboard and white paper. The goods could be
manufactured closer to customers, saving shipping costs.
Assume independent arguments
9
Are Arguments Independent?
The president is struggling to manage the country’s economy.
If he cannot get it under control, loss of the next election
might result.
10
Are Arguments Independent?
• What entity might lose?– Economies lose jobs, value, etc.– Presidents lose votes, allegiance, etc.
• Implicit arguments are not independent• A joint model would be more natural
The president is struggling to manage the country’s economy.
If he cannot get it under control, loss of the next election
might result.
11
Related Work
• Joint verbal SRL (Toutanova et al. (2008))– Re-rank full argument structures– Joint label sequence
• [arg0, Predicate, arg1]• [arg0, Predicate, arg0]
• Joint selectional preferences (Ritter et al. (2010))– [Arg0 economy] [Predicate lost] [Arg1 jobs]
– [Arg0 economy] [Predicate lost] [Arg1 election]
– Relies on TextRunner extraction system
12
TextRunner
• Open Information Extraction (OIE) database• Query
– Arg0: ?– Predicate: lose– Arg1: election
• Answer– [Arg0 The president] [Predicate lost] [Arg1 the election].
• Use TextRunner to identify joint implicit arguments
13
Joint Implicit Argument Model
• Model joint occurrence of iarg0 and iarg1
• Consider all possible candidate assignments
The president is struggling to manage the country’s economy.
If he doesn’t succeed by the next election, a loss might result.
14
Joint Implicit Argument Model
• Using TextRunner queries
• Query 1: <president, lose, ?>1. <Kenyan president, lose, election>
2. <president, lose, ally>
3. …
• Query 2: <?, lose, election>1. <Republican party, lose, election>
2. <president, lose, election>
3. …
• Match rank• Match similarity• Local model scores
15
Evaluation Setting
• Data created by Gerber and Chai (2010)– 1,200 annotations of 10 predicates
• Only test instances that take iarg0 and iarg1
• Ten-fold cross-validation• Baseline: independent classification model
16
• Methodology (Ruppenhofer et al., 2010)
– Ground-truth implicit arguments:– Predicted implicit argument:– Prediction score:
– P: total prediction score / prediction count– R: total prediction score / true implicit positions
Evaluation Setting
p},,{ 1 nttT
),Dice(max),Score( iTt
tpTpi
Georgia-Pacific and Nekoosa produce market pulp,
containerboard and white paper. The goods could be
manufactured closer to customers, saving shipping costs.
17
Evaluation Results
• Overall results– Baseline F1: 72.2%
– Joint F1: 73.1%
• Per-predicatePredicate Baseline F1 (%) Joint F1 (%)
bid 66.7 78.2
investment 53.3 62.5
18
Example Improvement
Big investors can decide to ride out market storms without
selling stock. They often do that because stocks have proved
to be the best-performing investment, attracting $1 trillion.
• What was invested?• Who invested?
– Baseline (independent) model is incorrect– Joint model is correct
[iarg1 money]
19
Example Improvement
Big investors can decide to ride out market storms without
selling stock. They often do that because stocks have proved
to be the best-performing investment, attracting $1 trillion.
• Query 1: <investor, invest, ?>– Answers: money, amount, million
• Query 2: <?, invest, money>– Answers: government, business, investor
[iarg1 money]
20
Summary
• Implicit arguments– Frequent– Nearby– Can be automatically recovered
• Semantic arguments are not independent– OIE can help identify argument dependencies– Joint model can recover from simple errors
21
Future Work
• Extension to other predicates– Only 10 are currently considered
• Extension to other argument positions– iarg2 and iarg3 are also common
• Computational complexity– Exhaustive search is intractable– Heuristic search– Gibbs sampling for joint inference