Introduction
Background
Complexity Metrics
Evaluation
Complexity Metrics in an Incremental, Right-corner Parser Stephen Wu, Asaf Bachrach, Carlos Cardenas, William Schuler University of Minnesota, INSERM-CEA, MIT, University of Minnesota and The Ohio State University
July 14, 2010 | ACL 2010
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Cognitively plausible Manner Memory limits Ex:
(Miller ’56, Cowan ’01) the nurse
Center-embedding
(Miller & Chomsky ’63, Gibson ’98)
Harder to understand Incremental Processing
(Tanenhaus et al. ’95)
Predictions: lexical, syntactic Context-dependent Right-corner HHMM parser PCFG Right-corner grammar (bounded memory) Chart parser HHMM (incremental processing)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Cognitively plausible Manner Memory limits Ex:
(Miller ’56, Cowan ’01)
the intern the nurse supervised
Center-embedding
(Miller & Chomsky ’63, Gibson ’98)
Harder to understand Incremental Processing
(Tanenhaus et al. ’95)
Predictions: lexical, syntactic Context-dependent Right-corner HHMM parser PCFG Right-corner grammar (bounded memory) Chart parser HHMM (incremental processing)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Cognitively plausible Manner Memory limits
(Miller ’56, Cowan ’01)
Ex: The drug the intern the nurse supervised administered
Center-embedding
(Miller & Chomsky ’63, Gibson ’98)
Harder to understand Incremental Processing
(Tanenhaus et al. ’95)
Predictions: lexical, syntactic Context-dependent Right-corner HHMM parser PCFG Right-corner grammar (bounded memory) Chart parser HHMM (incremental processing)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Cognitively plausible Manner Memory limits
(Miller ’56, Cowan ’01)
Ex: The drug the intern the nurse supervised administered cured the patient.
Center-embedding
(Miller & Chomsky ’63, Gibson ’98)
Harder to understand Incremental Processing
(Tanenhaus et al. ’95)
Predictions: lexical, syntactic Context-dependent Right-corner HHMM parser PCFG Right-corner grammar (bounded memory) Chart parser HHMM (incremental processing)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Cognitively plausible Manner Memory limits
(Miller ’56, Cowan ’01)
Ex: The drug [the intern [the nurse supervised] administered] cured the patient.
Center-embedding
(Miller & Chomsky ’63, Gibson ’98)
Harder to understand Incremental Processing
(Tanenhaus et al. ’95)
Predictions: lexical, syntactic Context-dependent Right-corner HHMM parser PCFG Right-corner grammar (bounded memory) Chart parser HHMM (incremental processing)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Cognitively plausible Manner Memory limits
(Miller ’56, Cowan ’01)
Ex: The drug [the intern [the nurse supervised] administered] cured the patient.
Center-embedding
(Miller & Chomsky ’63, Gibson ’98)
Harder to understand Incremental Processing
(Tanenhaus et al. ’95)
Predictions: lexical, syntactic Context-dependent Right-corner HHMM parser PCFG Right-corner grammar (bounded memory) Chart parser HHMM (incremental processing)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Cognitively plausible Manner Memory limits
(Miller ’56, Cowan ’01)
Ex: The drug [the intern [the nurse supervised] administered] cured the patient.
Center-embedding
(Miller & Chomsky ’63, Gibson ’98)
Harder to understand Incremental Processing
(Tanenhaus et al. ’95)
Ex: The left striker kicked the
Predictions: lexical, syntactic Context-dependent Right-corner HHMM parser PCFG Right-corner grammar (bounded memory) Chart parser HHMM (incremental processing)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Cognitively plausible Manner Memory limits
(Miller ’56, Cowan ’01)
Ex: The drug [the intern [the nurse supervised] administered] cured the patient.
Center-embedding
(Miller & Chomsky ’63, Gibson ’98)
Harder to understand Incremental Processing
(Tanenhaus et al. ’95)
Ex: The left striker kicked the penalty.
Predictions: lexical, syntactic Context-dependent Right-corner HHMM parser PCFG Right-corner grammar (bounded memory) Chart parser HHMM (incremental processing)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Cognitively plausible Manner Memory limits
(Miller ’56, Cowan ’01)
Ex: The drug [the intern [the nurse supervised] administered] cured the patient.
Center-embedding
(Miller & Chomsky ’63, Gibson ’98)
Harder to understand Incremental Processing
(Tanenhaus et al. ’95)
Ex: The left striker kicked the penalty.
Predictions: lexical, syntactic Context-dependent Right-corner HHMM parser PCFG Right-corner grammar (bounded memory) Chart parser HHMM (incremental processing)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Cognitively plausible Manner Memory limits
(Miller ’56, Cowan ’01)
Ex: The drug [the intern [the nurse supervised] administered] cured the patient.
Center-embedding
(Miller & Chomsky ’63, Gibson ’98)
Harder to understand Incremental Processing
(Tanenhaus et al. ’95)
Ex: The left striker kicked the bucket.
Predictions: lexical, syntactic Context-dependent Right-corner HHMM parser PCFG Right-corner grammar (bounded memory) Chart parser HHMM (incremental processing)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Cognitively plausible Manner Memory limits
(Miller ’56, Cowan ’01)
Ex: The drug [the intern [the nurse supervised] administered] cured the patient.
Center-embedding
(Miller & Chomsky ’63, Gibson ’98)
Harder to understand Incremental Processing
(Tanenhaus et al. ’95)
Ex: The left striker kicked the penalty.
Predictions: lexical, syntactic Context-dependent Right-corner HHMM parser PCFG Right-corner grammar (bounded memory) Chart parser HHMM (incremental processing)
Introduction
Background
Complexity Metrics
Evaluation
Does this match human sentence processing? Corpus studies (Schuler et al. ’08) HHMM depth limit no memory 1 memory element 2 memory elements 3 memory elements 4 memory elements 5 memory elements TOTAL
sentences 127 3,496 25,909 38,902 39,816 39,832 39,832
coverage 0.32% 8.78% 65.05% 97.67% 99.96% 100.00% 100.00%
Parsing Acuracy (Schuler et al. ’10) with punctuation: (≤ 40 wds) KM’03: unmodified, devset KM’03: par+sib, devset CKY: binarized, devset HHMM: par+sib, devset CKY: binarized, sect 23 HHMM: par+sib, sect 23
LR − − 80.3 84.1 78.8 83.4
LP − − 79.9 83.5 79.4 83.7
F 72.6 77.4 80.1 83.8 79.1 83.5
sentence failure 0 0 0.8 0.5 0.1 0.1
Predictions of reading difficulty? (this paper)
error reduction 17.5% 18.6% 21.1%
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Does this match human sentence processing? Corpus studies (Schuler et al. ’08) HHMM depth limit no memory 1 memory element 2 memory elements 3 memory elements 4 memory elements 5 memory elements TOTAL
sentences 127 3,496 25,909 38,902 39,816 39,832 39,832
coverage 0.32% 8.78% 65.05% 97.67% 99.96% 100.00% 100.00%
Parsing Acuracy (Schuler et al. ’10) with punctuation: (≤ 40 wds) KM’03: unmodified, devset KM’03: par+sib, devset CKY: binarized, devset HHMM: par+sib, devset CKY: binarized, sect 23 HHMM: par+sib, sect 23
LR − − 80.3 84.1 78.8 83.4
LP − − 79.9 83.5 79.4 83.7
F 72.6 77.4 80.1 83.8 79.1 83.5
sentence failure 0 0 0.8 0.5 0.1 0.1
Predictions of reading difficulty? (this paper)
error reduction 17.5% 18.6% 21.1%
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Does this match human sentence processing? Corpus studies (Schuler et al. ’08) HHMM depth limit no memory 1 memory element 2 memory elements 3 memory elements 4 memory elements 5 memory elements TOTAL
sentences 127 3,496 25,909 38,902 39,816 39,832 39,832
coverage 0.32% 8.78% 65.05% 97.67% 99.96% 100.00% 100.00%
Parsing Acuracy (Schuler et al. ’10) with punctuation: (≤ 40 wds) KM’03: unmodified, devset KM’03: par+sib, devset CKY: binarized, devset HHMM: par+sib, devset CKY: binarized, sect 23 HHMM: par+sib, sect 23
LR − − 80.3 84.1 78.8 83.4
LP − − 79.9 83.5 79.4 83.7
F 72.6 77.4 80.1 83.8 79.1 83.5
sentence failure 0 0 0.8 0.5 0.1 0.1
Predictions of reading difficulty? (this paper)
error reduction 17.5% 18.6% 21.1%
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Does this match human sentence processing? Corpus studies (Schuler et al. ’08) HHMM depth limit no memory 1 memory element 2 memory elements 3 memory elements 4 memory elements 5 memory elements TOTAL
sentences 127 3,496 25,909 38,902 39,816 39,832 39,832
coverage 0.32% 8.78% 65.05% 97.67% 99.96% 100.00% 100.00%
Parsing Acuracy (Schuler et al. ’10) with punctuation: (≤ 40 wds) KM’03: unmodified, devset KM’03: par+sib, devset CKY: binarized, devset HHMM: par+sib, devset CKY: binarized, sect 23 HHMM: par+sib, sect 23
LR − − 80.3 84.1 78.8 83.4
LP − − 79.9 83.5 79.4 83.7
F 72.6 77.4 80.1 83.8 79.1 83.5
sentence failure 0 0 0.8 0.5 0.1 0.1
Predictions of reading difficulty? (this paper)
error reduction 17.5% 18.6% 21.1%
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Does this match human sentence processing? Corpus studies (Schuler et al. ’08) HHMM depth limit no memory 1 memory element 2 memory elements 3 memory elements 4 memory elements 5 memory elements TOTAL
sentences 127 3,496 25,909 38,902 39,816 39,832 39,832
coverage 0.32% 8.78% 65.05% 97.67% 99.96% 100.00% 100.00%
Parsing Acuracy (Schuler et al. ’10) with punctuation: (≤ 40 wds) KM’03: unmodified, devset KM’03: par+sib, devset CKY: binarized, devset HHMM: par+sib, devset CKY: binarized, sect 23 HHMM: par+sib, sect 23
LR − − 80.3 84.1 78.8 83.4
LP − − 79.9 83.5 79.4 83.7
F 72.6 77.4 80.1 83.8 79.1 83.5
sentence failure 0 0 0.8 0.5 0.1 0.1
Predictions of reading difficulty? (this paper)
error reduction 17.5% 18.6% 21.1%
Conclusion
Introduction
Background
Complexity Metrics
Reading Times
Linear mixed-effect regression
Evaluation
Conclusion
(Bachrach et al)
Introduction
Background
Complexity Metrics
Reading Times
Linear mixed-effect regression
Evaluation
Conclusion
(Bachrach et al)
Introduction
Background
Complexity Metrics
Reading Times
Linear mixed-effect regression
Evaluation
Conclusion
(Bachrach et al)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Right-corner transform S
S
NP
S/NN DT
NN
VBD
NN
NP S/NP
the engineers
NN
S/NN
VP
VBD
PRT
DT
pulled
off
an
NN NN
NN
engineering trick
VBD
NP
VBD/PRT
NP/NN
NN
VBD
DT
engineers
pulled
the
Incremental by nature (Active/Awaited) Flatter structure Trunks and memory
DT
S/VP
an PRT off
trick
engineering
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Right-corner transform S
S
NP
S/NN DT
NN
VBD
NN
NP S/NP
the engineers
NN
S/NN
VP
VBD
PRT
DT
pulled
off
an
NN NN
NN
engineering trick
VBD
NP
VBD/PRT
NP/NN
NN
VBD
DT
engineers
pulled
the
Incremental by nature (Active/Awaited) Flatter structure Trunks and memory
DT
S/VP
an PRT off
trick
engineering
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Right-corner transform S
S
NP
S/NN DT
NN
VBD
NN
NP S/NP
the engineers
NN
S/NN
VP
VBD
PRT
DT
pulled
off
an
NN NN
NN
engineering trick
VBD
NP
VBD/PRT
NP/NN
NN
VBD
DT
engineers
pulled
the
Incremental by nature (Active/Awaited) Flatter structure Trunks and memory
DT
S/VP
an PRT off
trick
engineering
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Right-corner transform S
S
NP
S/NN
VP S/NN
DT
NN
VBD
NP S/NP
the engineers
VBD
PRT
DT
pulled
off
an
NN NN
NN
engineering trick
VBD
NP
VBD/PRT
NP/NN
NN
VBD
DT
engineers
pulled
the
Incremental by nature (Active/Awaited) Flatter structure Trunks and memory
DT
S/VP
NN NN
an PRT off
trick
engineering
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Right-corner transform S
S
NP
S/NN
VP S/NN
DT
NN
VBD
NP S/NP
the engineers
VBD
PRT
DT
pulled
off
an
NN NN
NN
engineering trick
VBD
NP
VBD/PRT
NP/NN
NN
VBD
DT
engineers
pulled
the
Incremental by nature (Active/Awaited) Flatter structure Trunks and memory
DT
S/VP
NN NN
an PRT off
trick
engineering
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Right-corner transform S
S
NP
S/NN DT
NN
VBD
NN
NP S/NP
the engineers
NN
S/NN
VP
VBD
PRT
DT
pulled
off
an
NN NN
NN
engineering trick
VBD
NP
VBD/PRT
NP/NN
NN
VBD
DT
engineers
pulled
the
Incremental by nature (Active/Awaited) Flatter structure Trunks and memory
DT
S/VP
an PRT off
trick
engineering
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Right-corner transform S
S
NP
S/NN DT
NN
VBD
NN
NP S/NP
the engineers
NN
S/NN
VP
VBD
PRT
DT
pulled
off
an
NN NN
NN
engineering trick
VBD
NP
VBD/PRT
NP/NN
NN
VBD
DT
engineers
pulled
the
Incremental by nature (Active/Awaited) Flatter structure Trunks and memory
DT
S/VP
an PRT off
trick
engineering
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Parsing with a Hierarchic Hidden Markov Model t=1
t=4
t=5
t=6
S/
···
···
P
N
N
VP
N
S/
S/
nn
VB
···
D rt /p
d vb
g in er ne gi en
an
f of
d lle pu
s er ne gi en
e th
word
VP
P/
d=2
t=3
S/
N
dt
d=1
t=2
···
1 center-embedding depth/memory element per trunk 1 word per time Many hypotheses (partial trees)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Parsing with a Hierarchic Hidden Markov Model t=1
t=4
t=5
t=6
···
S/ VP D
VB rt /p
d vb
f of
d lle pu
word
VP
s er ne gi en
nn
P/
e th
d=2
t=3
S/
N
dt
d=1
t=2
1 center-embedding depth/memory element per trunk 1 word per time Many hypotheses (partial trees)
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
The Trellis in HHMM Inference
◦
◦
◦
◦
◦
◦
◦
◦
◦
◦
.. . Incomplete trees in parallel Ranked by P(q1..t o1..t ) Beam Bt Viterbi (use backpointers)
◦
◦
◦
◦
d =1
···
◦
◦
◦
◦
◦
d =2
◦
d =1
d =2
.. .
◦
◦
◦
◦
◦
◦
◦
◦
◦
d =2
◦
.. .
d=3
◦
◦
◦
d=2
d=1
d=2
◦
◦
d=1
◦
◦
d =1
◦
d=1
◦
d =2
◦
◦
◦
◦
d =1
d=2
d=2
···
◦
d=1
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
The Trellis in HHMM Inference
◦
◦
◦
◦
◦
◦
◦
◦
◦
◦
.. . Incomplete trees in parallel Ranked by P(q1..t o1..t ) Beam Bt Viterbi (use backpointers)
◦
◦
◦
◦
d =1
···
◦
◦
◦
◦
◦
d =2
◦
d =1
d =2
.. .
◦
◦
◦
◦
◦
◦
◦
◦
◦
d =2
◦
.. .
d=3
◦
◦
◦
d=2
d=1
d=2
◦
◦
d=1
◦
◦
d =1
◦
d=1
◦
d =2
◦
◦
◦
◦
d =1
d=2
d=2
···
◦
d=1
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
The Trellis in HHMM Inference
◦
◦
◦
◦
◦
◦
◦
◦
◦
◦
.. . Incomplete trees in parallel Ranked by P(q1..t o1..t ) Beam Bt Viterbi (use backpointers)
◦
◦
◦
◦
d =1
···
◦
◦
◦
◦
◦
d =2
◦
d =1
d =2
.. .
◦
◦
◦
◦
◦
◦
◦
◦
◦
d =2
◦
.. .
d=3
◦
◦
◦
d=2
d=1
d=2
◦
◦
d=1
◦
◦
d =1
◦
d=1
◦
d =2
◦
◦
◦
◦
d =1
d=2
d=2
···
◦
d=1
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
The Trellis in HHMM Inference
◦
◦
◦
◦
◦
◦
◦
◦
◦
◦
.. . Incomplete trees in parallel Ranked by P(q1..t o1..t ) Beam Bt Viterbi (use backpointers)
◦
◦
◦
◦
d =1
···
◦
◦
◦
◦
◦
d =2
◦
d =1
d =2
.. .
◦
◦
◦
◦
◦
◦
◦
◦
◦
d =2
◦
.. .
d=3
◦
◦
◦
d=2
d=1
d=2
◦
◦
d=1
◦
◦
d =1
◦
d=1
◦
d =2
◦
◦
◦
◦
d =1
d=2
d=2
···
◦
d=1
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
The Trellis in HHMM Inference
◦
◦
◦
◦
◦
◦
◦
◦
◦
◦
.. . Incomplete trees in parallel Ranked by P(q1..t o1..t ) Beam Bt Viterbi (use backpointers)
◦
◦
◦
◦
d =1
···
◦
◦
◦
◦
◦
d =2
◦
d =1
d =2
.. .
◦
◦
◦
◦
◦
◦
◦
◦
◦
d =2
◦
.. .
d=3
◦
◦
◦
d=2
d=1
d=2
◦
◦
d=1
◦
◦
d =1
◦
d=1
◦
d =2
◦
◦
◦
◦
d =1
d=2
d=2
···
◦
d=1
Introduction
Background
Surprisal
Complexity Metrics
Evaluation
Conclusion
(Hale ’01; Levy ’07; Demberg & Keller ’08; Boston et al ’08; Roark et al ’09)
How much each word narrows possible interpretations
Introduction
Background
Surprisal
Complexity Metrics
Evaluation
(Hale ’01; Levy ’07; Demberg & Keller ’08; Boston et al ’08; Roark et al ’09)
How much each word narrows possible interpretations Pre(o1..t ) =
X
P(q1..t o1..t )
q1..t
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦
.. .
.. .
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦
···
Conclusion
···
.. .
Introduction
Background
Surprisal
Complexity Metrics
Evaluation
(Hale ’01; Levy ’07; Demberg & Keller ’08; Boston et al ’08; Roark et al ’09)
How much each word narrows possible interpretations Pre(o1..t ) =
X q1..t
|{z}
P(q1..t o1..t ) | {z } trellis prob.
from qt
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦
.. .
.. .
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦
···
Conclusion
···
.. .
Introduction
Background
Surprisal
Complexity Metrics
Evaluation
(Hale ’01; Levy ’07; Demberg & Keller ’08; Boston et al ’08; Roark et al ’09)
How much each word narrows possible interpretations Pre(o1..t ) =
X
P(q1..t o1..t )
q1..t
Surprisal(t) = log2 Pre(o1..t-1 ) − log2 Pre(o1..t )
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦
.. .
.. .
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦
···
Conclusion
···
.. .
Introduction
Background
Complexity Metrics
Evaluation
Embedding Difference Explicit representation for center-embedding “Memory cost” (Gibson ’98, ’00)
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Embedding Difference Explicit representation for center-embedding “Memory cost” (Gibson ’98, ’00) µEMB (o1..t ) =
X
d(qt ) · P(o1..t q1..t )/
P(o1..t q′1..t )
q′t ∈Bt
qt ∈Bt
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦
.. .
.. .
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦
···
X
···
.. .
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Embedding Difference Explicit representation for center-embedding “Memory cost” (Gibson ’98, ’00) frontier depth
X X z }| { P(o1..t q′1..t ) µEMB (o1..t ) = d(qt ) · P(o1..t q1..t ) / | {z } ′ qt ∈Bt
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦
.. .
.. .
qt ∈Bt
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦
···
trellis prob.
···
.. .
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Embedding Difference Explicit representation for center-embedding “Memory cost” (Gibson ’98, ’00) µEMB (o1..t ) =
X
d(qt ) · P(o1..t q1..t )/
P(o1..t q′1..t )
q′t ∈Bt
qt ∈Bt
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦
.. .
.. .
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦
···
X
···
.. .
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Embedding Difference Explicit representation for center-embedding “Memory cost” (Gibson ’98, ’00) µEMB (o1..t ) =
X
d(qt ) · P(o1..t q1..t )/
X
P(o1..t q′1..t )
q′t ∈Bt
qt ∈Bt
EmbDiff(o1..t ) = µEMB (o1..t ) − µEMB (o1..t−1 )
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦
.. .
.. .
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦
···
···
.. .
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Entropy Reduction How much info is in a probability distribution
Conclusion
(Hale ’03, Hale ’06, etc)
Introduction
Background
Complexity Metrics
Evaluation
Entropy Reduction
(Hale ’03, Hale ’06, etc)
How much info is in a probability distribution Ht =
Conclusion
X q1..t
P(q1..t o1..t ) log2 P(q1..t o1..t )
Introduction
Background
Complexity Metrics
Evaluation
Entropy Reduction
(Hale ’03, Hale ’06, etc)
How much info is in a probability distribution Ht =
X
P(q1..t o1..t ) log2 P(q1..t o1..t )
q1..t
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦
.. .
.. .
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦
···
Conclusion
···
.. .
Introduction
Background
Complexity Metrics
Evaluation
Entropy Reduction
(Hale ’03, Hale ’06, etc)
How much info is in a probability distribution Ht =
X
P(q1..t o1..t ) log2 P(q1..t o1..t )
q1..t
ER(ot ) = max(0, Ht−1 − Ht )
◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦
◦ ◦ ◦ ◦ ◦ ◦ ◦
.. .
.. .
◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
◦ ◦ ◦
···
Conclusion
···
.. .
Introduction
Background
Complexity Metrics
Evaluation
Linear Mixed-Effects Regression =
+
+
···
Fixed effects e.g., surprisal, bigram Random effects e.g., by word, by subj Slope (Coeff) Confidence (t) HHMM surprisal & entropy rdn. — significant predictors Embedding difference — independent contribution
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Linear Mixed-Effects Regression =
+
Fixed effects e.g., surprisal, bigram Random effects e.g., by word, by subj Slope (Coeff) Confidence (t)
+
···
F ULL DATA Coefficient Std. Err. t-value (Intcpt) order rlength unigrm bigrm embdiff etrpyrd srprsl
-9.340·10−3 -3.746·10−5 -2.002·10−2 -8.090·10−2 -2.074·10+0 9.390·10−3 2.753·10−2 3.950·10−3
5.347·10−2 7.808·10−6 1.635·10−2 3.690·10−1 8.132·10−1 3.268·10−3 6.792·10−3 3.452·10−4
-0.175 -4.797∗ -1.225 -0.219 -2.551∗ 2.873∗ 4.052∗ 11.442∗
HHMM surprisal & entropy rdn. — significant predictors Embedding difference — independent contribution
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Linear Mixed-Effects Regression =
+
Fixed effects e.g., surprisal, bigram Random effects e.g., by word, by subj Slope (Coeff) Confidence (t)
+
···
F ULL DATA Coefficient Std. Err. t-value (Intcpt) order rlength unigrm bigrm embdiff etrpyrd srprsl
-9.340·10−3 -3.746·10−5 -2.002·10−2 -8.090·10−2 -2.074·10+0 9.390·10−3 2.753·10−2 3.950·10−3
5.347·10−2 7.808·10−6 1.635·10−2 3.690·10−1 8.132·10−1 3.268·10−3 6.792·10−3 3.452·10−4
-0.175 -4.797∗ -1.225 -0.219 -2.551∗ 2.873∗ 4.052∗ 11.442∗
HHMM surprisal & entropy rdn. — significant predictors Embedding difference — independent contribution
Introduction
Background
Complexity Metrics
Evaluation
Conclusion
Linear Mixed-Effects Regression =
+
Fixed effects e.g., surprisal, bigram Random effects e.g., by word, by subj Slope (Coeff) Confidence (t)
+
···
F ULL DATA Coefficient Std. Err. t-value (Intcpt) order rlength unigrm bigrm embdiff etrpyrd srprsl
-9.340·10−3 -3.746·10−5 -2.002·10−2 -8.090·10−2 -2.074·10+0 9.390·10−3 2.753·10−2 3.950·10−3
5.347·10−2 7.808·10−6 1.635·10−2 3.690·10−1 8.132·10−1 3.268·10−3 6.792·10−3 3.452·10−4
-0.175 -4.797∗ -1.225 -0.219 -2.551∗ 2.873∗ 4.052∗ 11.442∗
HHMM surprisal & entropy rdn. — significant predictors Embedding difference — independent contribution
Introduction
Background
Complexity Metrics
Conclusion Simple complexity metrics in HHMM Significant predictors Surprisal Entropy reduction Embedding difference
∴ Modeling bounded memory in a parser: quantitative broad-coverage ... while not hurting the parse
Evaluation
Conclusion
Introduction
Background
Complexity Metrics
Conclusion Simple complexity metrics in HHMM Significant predictors Surprisal −→ standout Entropy reduction Embedding difference
∴ Modeling bounded memory in a parser: quantitative broad-coverage ... while not hurting the parse
Evaluation
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Conclusion Simple complexity metrics in HHMM Significant predictors Surprisal −→ standout Entropy reduction Embedding difference
∴ Modeling bounded memory in a parser: ↑ short-term memory use quantitative broad-coverage ... while not hurting the parse
↑ linguistic complexity
Conclusion
Introduction
Background
Complexity Metrics
Thank you! William Schuler Tim Miller Brian Roark Mark Holland
I have moved to Mayo Clinic.
[email protected]
Evaluation
Conclusion
Introduction
Background
Complexity Metrics
Evaluation
Results F ULL DATA Coefficient Std. Err. t-value (Intcpt) order rlength unigrm bigrm embdiff etrpyrd srprsl order rlength unigrm bigrm emdiff etrpyrd srprsl
± + + + +
O PEN t-value -0.237 -4.621∗ 0.554 -0.391 -3.248∗ 0.539 0.063 6.285∗
-9.340·10−3 5.347·10−2 -0.175 -3.746·10−5 7.808·10−6 -4.797∗ -2.002·10−2 1.635·10−2 -1.225 -8.090·10−2 3.690·10−1 -0.219 -2.074·10+0 8.132·10−1 -2.551∗ 9.390·10−3 3.268·10−3 2.873∗ 2.753·10−2 6.792·10−3 4.052∗ 3.950·10−3 3.452·10−4 11.442∗ (Intr) order rlngth ungrm bigrm emdiff entrpy .000 -.006 -.003 .049 .000 -.479 .001 .005 -.006 -.073 .000 .009 -.049 -.089 .095 .000 .003 .016 -.014 .020 -.010 .000 -.008 -.033 -.079 .107 .362 .171
C LOSED ± t-value - -0.794 - -4.232∗ - -0.865 - -0.644 - -2.645∗ + 3.082∗ + 4.857∗ + 9.286∗
Conclusion