Difference between revisions of "Berg-Kirkpatrick et al, ACL 2010: Painless Unsupervised Learning with Features"

From Cohen Courses
Jump to navigationJump to search
 
(21 intermediate revisions by the same user not shown)
Line 1: Line 1:
−
__TOC__
 
 
== Citation ==
 
== Citation ==
  
−
T. Berg-Kirkpatrick, A. Bouchard-Côté, J. DeNero, and D. Klein. '''Painless Unsupervised Learning with Features''', ''Human Language Technologies 2010'', pp. 582-590, Los Angeles, June 2010.
+
T. Berg-Kirkpatrick, A. Bouchard-Côté, J. DeNero and D. Klein. '''Painless Unsupervised Learning with Features''', ''Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL'', pp. 582-590, Los Angeles, June 2010.
  
 
== Online Version ==
 
== Online Version ==
Line 10: Line 9:
 
== Summary ==
 
== Summary ==
  
−
This [[Category::paper]] generalizes conventional HMMs to featurized HMMs, by replacing the multinomial conditional probability distributions (CPDs) with miniature log-linear models. Two algorithms for unsupervised training of featurized HMMs are proposed.
+
This [[Category::paper]] generalizes conventional [[UsesMethod::HMM|HMMs]] to [[UsesMethod::Featurized_HMM|featurized HMMs]], by replacing the multinomial conditional probability distributions (CPDs) with miniature log-linear models. Two algorithms for unsupervised training of featurized HMMs are proposed.
  
 
Featurized HMMs are applied to four unsupervised learning tasks:
 
Featurized HMMs are applied to four unsupervised learning tasks:
−
* POS induction (unsupervised version of POS tagging);
+
* [[AddressesProblem::Part_of_Speech_Induction|POS induction]] (unsupervised version of [[Part_of_Speech_Tagging|POS tagging]]);
−
* Grammar induction;
+
* [[AddressesProblem::Grammar_Induction|Grammar induction]];
−
* Word alignment;
+
* [[AddressesProblem::Word_Alignment|Word alignment]];
−
* Word segmentation.
+
* [[AddressesProblem::Word_Segmentation|Word segmentation]].
 
For all these four tasks, featurized HMMs are shown to outperform their unfeaturized counterparts by a substantial margin.
 
For all these four tasks, featurized HMMs are shown to outperform their unfeaturized counterparts by a substantial margin.
  
−
== Featurized HMMs ==
+
== Method ==
  
−
=== Definition ===
+
This paper proposes the concept of [[UsesMethod::Featurized_HMM|featurized HMMs]] and two algorithms for their unsupervised training. For a detailed elaboration, see the page [[UsesMethod::Featurized_HMM|Featurized HMMs]].
  
−
=== The Estimation Problem ===
+
The paper also comes up with featurized versions of other HMM-like models, e.g. [[UsesMethod::Dependency Model with Valence]] for [[AddressesProblem::Grammar_Induction|Grammar induction]].
  
−
=== The Decoding Problem ===
+
== Experiments ==
  
−
=== The Training Problem ===
+
=== POS Induction ===
  
−
== Experiments ==
+
POS induction is the unsupervised version of [[Part_of_Speech_Tagging|POS tagging]]. The output is clusters of words that the system believes to belong to the same part-of-speech. In order to evaluate the performance of a POS inductor, it is necessary to map the clusters to the actual POS tags. The best accuracy achieved by all the mapping is called the "many-1 accuracy".
  
−
=== POS Induction ===
+
{| border=1 cellpadding=3
 +
! align=right | Dataset
 +
| [[UsesDataset::Penn Treebank English WSJ]]
 +
|-
 +
! align=right | Criterion
 +
| Many-1 accuracy (the larger, the better)
 +
|-
 +
! align=right | Baseline
 +
| 63.1 ± 1.3 (HMM) (10 runs, mean ± standard deviation)
 +
|-
 +
! align=right | Performance of proposed systems
 +
| 68.1 ± 1.7 (Featurized HMM, Algorithm 1) <br> 75.5 ± 1.1 (Featurized HMM, Algorithm 2)
 +
|-
 +
! align=right | Performance of contrastive systems
 +
| 59.6 ± 6.9 (Featurized MRF) <nowiki>[</nowiki>[[RelatedPaper::Haghighi and Klein, ACL 2006: Prototype-Driven Learning for Sequence Models|Haghighi and Klein, ACL 2006]]<nowiki>]</nowiki>
 +
|}
  
 
=== Grammar Induction ===
 
=== Grammar Induction ===
 +
 +
{| border=1 cellpadding=3
 +
! align=right | Dataset
 +
| English: [[UsesDataset::Penn Treebank English WSJ]] <br> Chinese: [[UsesDataset::Penn Treebank Chinese]]
 +
|-
 +
! align=right | Criterion
 +
| Accuracy (the larger the better, not clear how it is defined)
 +
|-
 +
! align=right | Baseline
 +
| English 47.8, Chinese 42.5 (DMV)
 +
|-
 +
! align=right | Performance of proposed systems
 +
| English 48.3, Chinese 49.9 (Featurized DMV, Algorithm 1) <br> English 63.0, Chinese 53.6 (Featurized DMV, Algorithm 2)
 +
|-
 +
! align=right | Performance of contrastive systems
 +
| English 61.3, Chinese 51.9 <nowiki>[</nowiki>[[RelatedPaper::Cohen and Smith, ACL 2009: Shared Logistic Normal Distributions for Soft Parameter Tying in Unsupervised Grammer Induction|Cohen and Smith, ACL 2009]]<nowiki>]</nowiki>
 +
|}
  
 
=== Word Alignment ===
 
=== Word Alignment ===
 +
 +
{| border=1 cellpadding=3
 +
! align=right | Dataset
 +
| [[UsesDataset::NIST 2002 Chinese-English Development Set]]
 +
|-
 +
! align=right | Criterion
 +
| Alignment Error Rate (the smaller the better)
 +
|-
 +
! align=right | Baseline
 +
| 38.0 (Model 1) <nowiki>[</nowiki>[[RelatedPaper::Brown et al, CL 1994: The Mathematics of Statistical Machine Translation: Parameter Estimation|Brown et al, CL 1994]]<nowiki>]</nowiki> <br> 33.8 (HMM) <nowiki>[</nowiki>[[RelatedPaper::Ney and Vogel, CL 1996: HMM-Based Word Alignment in Statistical Translation|Ney and Vogel, CL 1996]]<nowiki>]</nowiki>
 +
|-
 +
! align=right | Performance of proposed systems
 +
| 35.6 (Featurized Model 1, Algorithm 1) <br> 30.0 (Featurized HMM, Algorithm 1)
 +
|}
  
 
=== Word Segmentation ===
 
=== Word Segmentation ===
 +
 +
{| border=1 cellpadding=3
 +
! align=right | Dataset
 +
| [[UsesDataset::Bernstein-Ratner Corpus]]
 +
|-
 +
! align=right | Criterion
 +
| Segment F1 score (the larger the better)
 +
|-
 +
! align=right | Baseline
 +
| 76.9 ± 0.1 (Unigram) (10 runs, mean ± standard deviation)
 +
|-
 +
! align=right | Performance of proposed systems
 +
| 84.5 ± 0.5 (Featurized Unigram, Algorithm 1) <br> 88.0 ± 0.1 (Featurized Unigram, Algorithm 2)
 +
|-
 +
! align=right | Performance of contrastive systems
 +
| 87 <nowiki>[</nowiki>[[RelatedPaper::Johnson and Goldwater, ACL 2009: Improving Non-Parametric Bayesian Inerence: Experiments on Unsupervised Word Segmentation with Adaptor Grammars|Johnson and Goldwater, ACL 2009]]<nowiki>]</nowiki>
 +
|}

Latest revision as of 14:37, 18 September 2011

Citation

T. Berg-Kirkpatrick, A. Bouchard-Côté, J. DeNero and D. Klein. Painless Unsupervised Learning with Features, Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pp. 582-590, Los Angeles, June 2010.

Online Version

PDF version

Summary

This paper generalizes conventional HMMs to featurized HMMs, by replacing the multinomial conditional probability distributions (CPDs) with miniature log-linear models. Two algorithms for unsupervised training of featurized HMMs are proposed.

Featurized HMMs are applied to four unsupervised learning tasks:

For all these four tasks, featurized HMMs are shown to outperform their unfeaturized counterparts by a substantial margin.

Method

This paper proposes the concept of featurized HMMs and two algorithms for their unsupervised training. For a detailed elaboration, see the page Featurized HMMs.

The paper also comes up with featurized versions of other HMM-like models, e.g. Dependency Model with Valence for Grammar induction.

Experiments

POS Induction

POS induction is the unsupervised version of POS tagging. The output is clusters of words that the system believes to belong to the same part-of-speech. In order to evaluate the performance of a POS inductor, it is necessary to map the clusters to the actual POS tags. The best accuracy achieved by all the mapping is called the "many-1 accuracy".

Dataset Penn Treebank English WSJ
Criterion Many-1 accuracy (the larger, the better)
Baseline 63.1 ± 1.3 (HMM) (10 runs, mean ± standard deviation)
Performance of proposed systems 68.1 ± 1.7 (Featurized HMM, Algorithm 1)
75.5 ± 1.1 (Featurized HMM, Algorithm 2)
Performance of contrastive systems 59.6 ± 6.9 (Featurized MRF) [Haghighi and Klein, ACL 2006]

Grammar Induction

Dataset English: Penn Treebank English WSJ
Chinese: Penn Treebank Chinese
Criterion Accuracy (the larger the better, not clear how it is defined)
Baseline English 47.8, Chinese 42.5 (DMV)
Performance of proposed systems English 48.3, Chinese 49.9 (Featurized DMV, Algorithm 1)
English 63.0, Chinese 53.6 (Featurized DMV, Algorithm 2)
Performance of contrastive systems English 61.3, Chinese 51.9 [Cohen and Smith, ACL 2009]

Word Alignment

Dataset NIST 2002 Chinese-English Development Set
Criterion Alignment Error Rate (the smaller the better)
Baseline 38.0 (Model 1) [Brown et al, CL 1994]
33.8 (HMM) [Ney and Vogel, CL 1996]
Performance of proposed systems 35.6 (Featurized Model 1, Algorithm 1)
30.0 (Featurized HMM, Algorithm 1)

Word Segmentation

Dataset Bernstein-Ratner Corpus
Criterion Segment F1 score (the larger the better)
Baseline 76.9 ± 0.1 (Unigram) (10 runs, mean ± standard deviation)
Performance of proposed systems 84.5 ± 0.5 (Featurized Unigram, Algorithm 1)
88.0 ± 0.1 (Featurized Unigram, Algorithm 2)
Performance of contrastive systems 87 [Johnson and Goldwater, ACL 2009]
Retrieved from "http://curtis.ml.cmu.edu/w/courses/index.php?title=Berg-Kirkpatrick_et_al,_ACL_2010:_Painless_Unsupervised_Learning_with_Features&oldid=6236"