fast full parsing by linear-chain conditional random fields

Fast Full Parsing by Linear-Chain Conditional Random Fields

Yoshimasa Tsuruoka, Jun’ichi Tsujii, and Sophia Ananiadou

The University of Manchester

Outline• Motivation

• Parsing algorithm• Chunking with conditional random fields• Searching for the best parse

• Experiments• Penn Treebank

• Conclusions

Motivation• Parsers are useful in many NLP applications– Information extraction, Summarization, MT, etc.

• But parsing is often the most computationally expensive component in the NLP pipeline

• Fast parsing is useful when– The document collection is large

– e.g. MEDLINE corpus: 70 million sentences– Real-time processing is required

– e.g. web applications

Parsing algorithms

• History-based approaches– Bottom-up & left-to-right (Ratnaparkhi, 1997)– Shift-reduce (Sagae & Lavie 2006)

• Global modeling– Tree CRFs (Finkel et al., 2008; Petrov & Klein 2008)– Reranking (Collins 2000; Charniak & Johnson, 2005)– Forest (Huang, 2008)

Chunk parsing• Parsing Algorithm

1. Identify phrases in the sequence.2. Convert the recognized phrases into new non-

terminal symbols.3. Go back to 1.

• Previous work– Memory-based learning (Tjong Kim Sang, 2001)• F-score: 80.49

– Maximum entropy (Tsuruoka and Tsujii, 2005)• F-score: 85.9

Parsing a sentence

Estimated volume was a light 2.4 million ounces .

VBN NN VBD DT JJ CD CD NNS .

QP

NP

VP

NP

S

Estimated volume was a light 2.4 million ounces .

VBN NN VBD DT JJ CD CD NNS .

QPNP

1st iteration

volume was a light million ounces .

NP VBD DT JJ QP NNS .

NP

2nd iteration

volume was ounces .

NP VBD NP .

VP

3rd iteration

volume was .

NP VP .

S

4th iteration

was

S

5th iteration

Estimated volume was a light 2.4 million ounces .

VBN NN VBD DT JJ CD CD NNS .

QP

NP

VP

NP

S

Complete parse tree

Chunking with CRFs

• Conditional random fields (CRFs)

• Features are defined on states and state transitions

Feature functionFeature weight

F

i

n

tttiin yytf

ZyyP

1 111 ,,,exp

1)|...( xx

Estimated volume was a light 2.4 million ounces .

VBN NN VBD DT JJ CD CD NNS .

QPNP

Estimated volume was a light 2.4 million ounces .

VBN NN VBD DT JJ CD CD NNS .

Chunking with “IOB” tagging

B-NP I-NP O O O B-QP I-QP O O

NP QP

B : Beginning of a chunk I : Inside (continuation) of the chunkO : Outside of chunks

Features for base chunking

Estimated volume was a light 2.4 million ounces .

VBN NN VBD DT JJ CD CD NNS .

?

Features for non-base chunking

volume was a light million ounces .

NP VBD DT JJ QP NNS .

NP

VBN NN

Estimated volume

?

Finding the best parse

• Scoring the entire parse tree

• The best derivation can be found by depth-first search.

h

iiipscore

0

| xy

Depth first searchPOS tagging

Chunking (base)

Chunking

Chunking Chunking

Chunking

Chunking (base)

Chunking

Chunking

Finding the best parse

Extracting multiple hypotheses from CRF

• A* search– Uses a priority queue– Suitable when top n hypotheses are needed

• Branch-and-bound– Depth-first– Suitable when a probability threshold is given

CRF

BIOOOB

0.3

BIIOOB

0.2

BIOOOO

0.18

Experiments• Penn Treebank Corpus– Training: sections 2-21– Development: section 22– Evaluation: section 23

• Training– Three CRF models

• Part-of-speech tagger• Base chunker• Non-base chunker

– Took 2 days on AMD Opteron 2.2GHz

Training the CRF chunkers

• Maximum likelihood + L1 regularization

• L1 regularization helps avoid overfitting and produce compact modes– OWLQN algorithm (Andrew and Gao, 2007)

i

ij

jj CpL xy |log

Chunking performance

Symbol # Samples Recall Precison F-score

NP 317,597 94.79 94.16 94.47

VP 76,281 91.46 91.98 91.72

PP 66,979 92.84 92.61 92.72

S 33,739 91.48 90.64 91.06

ADVP 21,686 84.25 85.86 85.05

ADJP 14,422 77.27 78.46 77.86

: : : : :

All 579,253 92.63 92.62 92.63

Section 22, all sentences

Beam width and parsing performance

Beam Recall Precision F-score Time (sec)

1 86.72 87.83 87.27 16

2 88.50 88.85 88.67 41

3 88.69 89.08 88.88 61

4 88.72 89.13 88.92 92

5 88.73 89.14 88.93 119

10 88.68 89.19 88.93 179

Section 22, all sentences (1,700 sentences)

Comparison with other parsers

Recall Prec. F-score Time (min)

This work (deterministic) 86.3 87.5 86.9 0.5

This work (beam = 4) 88.2 88.7 88.4 1.7

Huang (2008) 91.7 Unk

Finkel et al. (2008) 87.8 88.2 88.0 >250

Petrov & Klein (2008) 88.3 3

Sagae & Lavie (2006) 87.8 88.1 87.9 17

Charniak & Johnson (2005) 90.6 91.3 91.0 Unk

Charniak (2000) 89.6 89.5 89.5 23

Collins (1999) 88.1 88.3 88.2 39

Section 23, all sentences (2,416 sentences)

Discussions

• Improving chunking accuracy– Semi-Markov CRFs (Sarawagi and Cohen, 2004)– Higher order CRFs

• Increasing the size of training data– Create a treebank by parsing a large number of

sentences with an accurate parser– Train the fast parser using the treebank

Conclusion

• Full parsing by cascaded chunking– Chunking with CRFs– Depth-first search

• Performance– F-score = 86.9 (12msec/sentence)– F-score = 88.4 (42msec/sentence)

• Available soon