Easy-First Coreference Resolution Veselin Stoyanov and Jason Eisner Johns Hopkins University.

Easy-First Coreference Resolution

Veselin Stoyanov and Jason EisnerJohns Hopkins University

Coreference Resolution

• Given a peace of text, identify all noun phrases that refer to the same real-world entity



Former Boxing Champion Mike Tyson revealed that he had caught Hollywood heartthrob Brad Pitt with his ex-wife Robin Givens.In an interview Tyson said that he was amidst the process of divorce from Givens in the late 1980s when he found her sleeping with the Hollywood star.








• Typical pairwise approaches• But, inherently global task



• Inherently global task



• Inherently global task


Global Approaches to Coreference Resolution

• A difficult search problem– Exponentially many possible clusterings

• Recent approaches have been successful:– [Culotta et al., 2007] – train a global coherence

measure; use greedy inference at test time– [Poon and Domingos, 2008] – use an unsupervised

approach based on MLNs to enforce global properties

A Different Approach

• A Multi-Pass Sieve [Raghunathan et al., 2010]– Build coreference clusters incrementally– Applying sieves (hand-specified rules) of

decreasing precision and increasing recall– Winner of the CoNLL-2011 shared task

A Multi-Pass Sieve for Coreference Resolution [Raghunathan et al. 2010]



• Sieves:1. Exact match



• Sieves:1. Exact match2. Precise Constructs



• Sieves:1. Exact match2. Precise Constructs



• Sieves:1. Exact match2. Precise Constructs3. Strict Head match



• Sieves:1. Exact match2. Precise Constructs3. Strict Head match



• Sieves:1. Exact match2. Precise Constructs3. Strict Head match………….

7. Pronoun MatchingFormer Boxing Champion Mike Tyson revealed that he had caught Hollywood heartthrob Brad Pitt with his ex-wife Robin Givens.In an interview Tyson said that he was amidst the process of divorce from Givens in the late 1980s when he found her sleeping with the Hollywood star.


• Sieves:1. Exact match2. Precise Constructs3. Strict Head match………….

7. Pronoun MatchingFormer Boxing Champion Mike Tyson revealed that he had caught Hollywood heartthrob Brad Pitt with his ex-wife Robin Givens.In an interview Tyson said that he was amidst the process of divorce from Givens in the late 1980s when he found her sleeping with the Hollywood star.


• Use the same idea:– Greedily build clusters– Using more information for the less confident

decisions• But learn automatically what constitutes an

“easy” (confident) decision– Leads to empirical improvement over the Multi-

Pass Sieve approach and a competitive baseline


• Incrementally and greedily build clusters– Using machine learning

• Learn policy for the test time inference procedure

Easy-First Coreference: Talk Overview

• Test-Time Inference• Training Algorithm• Experimental results

TEST-TIME INFERENCE

Test-Time Inference

• Begin with each NP in a separate cluster


Test-Time Inference


Cluster 1SemType: uknwnGender: uknwnNumber: singleNames: NIL

Cluster 2SemType: personGender: maleNumber: singleNames: Mike Tyson

Cluster 3SemType: personGender: maleNumber: singleNames: NIL

Cluster 3SemType: personGender: maleNumber: singleNames: NIL

• Begin with each NP in a separate cluster

• Compute a feature vector for each pair of clusters

Test-Time Inference





Test-Time Inference




SemTypeSame = False


Test-Time Inference




SemTypeSame = FalseSemTypeComp = True


Test-Time Inference




SemTypeSame = FalseSemTypeComp = TrueNumberSame = True


Test-Time Inference




SemTypeSame = FalseSemTypeComp = TrueNumberSame = TrueRoleAppositive = True


Test-Time Inference




SemTypeSame = FalseSemTypeComp = TrueNumberSame = TrueRoleAppositive = TrueSentDistance = 0


Test-Time Inference




SemTypeSame = FalseSemTypeComp = TrueNumberSame = TrueRoleAppositive = TrueSentDistance = 0

Globalfeatures

Localfeatures


Test-Time Inference




f(joinij)

• Compute a weight for each join action

Test-Time Inference




w.f(joinij)

• Compute a weight for each join action

Test-Time Inference




w.f(joinij)weightij=

• Perform the action with the highest weight

Test-Time Inference




w.f(joinij)aij=argmaxij


Test-Time Inference




w.f(joinij)a12=argmaxij


Test-Time Inference



w.f(joinij)a12=argmaxij


Test-Time Inference



Test-Time Inference



Test-Time Inference



Test-Time Inference



Test-Time Inference



Test-Time Inference



Test-Time Inference



Test-Time Inference



Test-Time Inference



Test-Time Inference



Test-Time Inference


• Until the action with the highest weight is the special action HALT

Test-Time Inference


Time complexity

• Let n = the number of mentions in a document• At most n clusters at any time• s = number of joins performed by the

algorithm• s<=n at most n joins before HALT (usually

much less)

Time complexity

• Each step requires:– Popping the top action -- O(logn) (priority queue)– Performing the join – O(1)– Deleting O(n) old actions – O(nlogn)– Computing scores for O(n) new actions -- O(n)– Inserting those -- O(nlogn) or O(n)

• Total runtime = O(n2logn)+O(snlogn)

TRAINING ALGORITHM

Learning

• Learn weight vector w that leads the Easy-First algorithm to good clusterings

• At each step we are given a state s and a set of possible actions {joinij}∪{HALT}

• We want to learn what is “the best” action in each state

• But the current state s depends on the action taken so far

Reinforcement Learning

• A classic reinforcement learning setting• Act in the world and learn “good” actions in

each state• We tried several reinforcement learning

algorithms– Policy gradient [Sutton et al., 2000]– Q-learning [Watkins and Dayan, 1992]– No success: distant supervision; too many possible

states

Perceptron-Style Training

• A version of the early update structured perceptron [Collins and Roark, 2004]– Inspired by the success of easy-first parsing

[Goldberg and Elhadad, 2010]• The algorithm:

– Keep acting in the world until the next action will result in an error (reduction in score)

– Update weights away from the top action and towards the highest scoring positive action

Training algorithm

• Perform N iterations:– For each document d in the training set do:

• Initialize C so that each mention in its own cluster• Initialize A = {joinij}∪{HALT}

• repeat until amax != HALT:– for k steps do:

» amax =A.peekMax()

» if amax.isPositive() break

» apos =A.peekMaxPositive()

» w+=(feat(apos,C)-feat(amax ,C))» Recompute scores in A using new w

update weight

Training algorithm

• Perform N iterations:– For each document d in the training set do:

• Initialize C so that each mention in its own cluster• Initialize A = {joinij}∩{HALT}

• repeat until amax != HALT:– Update weights– amax =A.popMax()

– if amax is a join

» perform amax

EXPERIMENTAL EVALUATION

Experimental Setup

• Five datasets:– MUC6, MUC7, ACE03, ACE04, ACE05

• Two evaluation metrics: – MUC score– B^3 score

• Implemented in the Reconcile research platform [Stoyanov et al. 2010]– Code available upon request

• Two settings:– Automatic mentions– Gold (manual) mentions

Baselines

• A pairwise baseline [Stoyanov et al. 2009]• Reimplementation of the pairwise baseline using

different features• Easy-First coreference resolution

– Easy-Firstpercep – weights from the pairwise baseline

– Easy-Firststruct – weights from our training algorithm

• Only for the gold mentions data– Entity-centered approach [Haghighi and Klein, 2009]– Multi-pass sieve [Raghunathan et al., 2010]

Results: Automatic Mentions

MUC6 MUC7 ACE03 ACE04 ACE05 B^3 MUC B^3 MUCB^3 MUC B^3 MUC B^3 MUC Reconcile 70.9 68.5 65.9 62.8 79.4 67.9 76.5 62.0 73.7 67.4


MUC6 MUC7 ACE03 ACE04 ACE05 B^3 MUC B^3 MUCB^3 MUC B^3 MUC B^3 MUC Reconcile 70.9 68.5 65.9 62.8 79.4 67.9 76.5 62.0 73.7 67.4Pairwise baseline 70.7 67.9 62.2 57.9 81.7 73.5 79.4 68.0 75.7 70.2






MUC6 MUC7 ACE03 ACE04 ACE05 B^3 MUC B^3 MUCB^3 MUC B^3 MUC B^3 MUC Reconcile 70.9 68.5 65.9 62.8 79.4 67.9 76.5 62.0 73.7 67.4Pairwise baseline 70.7 67.9 62.2 57.9 81.7 73.5 79.4 68.0 75.7 70.2Easy-First percep 71.5 68.4 64.4 58.1 82.8 74.1 80.3 68.4 76.0 70.0


MUC6 MUC7 ACE03 ACE04 ACE05 B^3 MUC B^3 MUCB^3 MUC B^3 MUC B^3 MUC Reconcile 70.9 68.5 65.9 62.8 79.4 67.9 76.5 62.0 73.7 67.4Pairwise baseline 70.7 67.9 62.2 57.9 81.7 73.5 79.4 68.0 75.7 70.2Easy-First percep 71.5 68.4 64.4 58.1 82.8 74.1 80.3 68.4 76.0 70.0Easy-First struct 72.1 68.2 64.6 57.7 83.1 74.5 80.9 68.8 75.9 70.9





Results: Gold Mentions

MUC6 ACE04

B^3 MUC B^3 MUC

Haghighi and Klein 75.0 81.9 76.9 76.5

Raghunathan et al. 73.2 77.7 78.9 78.1


MUC6 ACE04

B^3 MUC B^3 MUC



Stoyanov et al. 76.1 88.2 77.8 76.2


MUC6 ACE04

B^3 MUC B^3 MUC



Stoyanov et al. 76.1 88.2 77.8 76.2


MUC6 ACE04

B^3 MUC B^3 MUC



Stoyanov et al. 76.1 88.2 77.8 76.2


MUC6 ACE04

B^3 MUC B^3 MUC



Stoyanov et al. 76.1 88.2 77.8 76.2

Pairwise baseline 72.7 88.2 75.9 78.9


MUC6 ACE04

B^3 MUC B^3 MUC



Stoyanov et al. 76.1 88.2 77.8 76.2

Pairwise baseline 72.7 88.2 75.9 78.9


MUC6 ACE04

B^3 MUC B^3 MUC



Stoyanov et al. 76.1 88.2 77.8 76.2

Pairwise baseline 72.7 88.2 75.9 78.9Easy-First percep 73.3 88.2 78.4 79.8


MUC6 ACE04

B^3 MUC B^3 MUC



Stoyanov et al. 76.1 88.2 77.8 76.2

Pairwise baseline 72.7 88.2 75.9 78.9Easy-First percep 73.3 88.2 78.4 79.8Easy-First struct 77.5 88.2 81.8 80.1

Conclusions

• A new approach to coreference using greedy incremental clustering– More accurate clustering by utilizing global

information as it becomes available for less confident decisions

• Learning policy for the test-time inference procedure leads to improvements

THANK YOU!

Easy-First Coreference Resolution Veselin Stoyanov and Jason Eisner Johns Hopkins University.

Documents

interview tyson

hollywood star

boxing champion mike

hollywood heartthrob

exwife robin givens

coreference resolution

process of divorce

shared task slide