Structural Expectations in Chinese Relative Clause...

Structural Expectations in Chinese Relative Clause Comprehension

Zhong Chen, Kyle Grove, John HaleDepartment of Linguistics

Cornell University

1

Saturday, April 23, 2011

Outline

Introduction

Chinese RC

Modeling

Conclusion

2


Relative clause comprehension

Subject relative clause (SR)

The senatori [who ei attacked the reporter] admitted the error.

Object relative clause (OR)

The senatori [who the reporter attacked ei] admitted the error.

Comprehension difficulty: SR << OR

3


Two major accounts

Memory-based

(Gibson 1998, 2000; Grodner & Gibson 2005)

Experience-based

(Mitchell et al 1995; Hale 2001; Lewis & Vasishth 2005; Levy 2008)

4


Memory-based approach

Subject Relative clause (SR)

The senatori [who ei attacked the reporter] admitted the error.

Object Relative clause (OR)

The senatori [who the reporter attacked ei] admitted the error.

Memory-based: SR << OR

5


Experience-based approach

English Penn Treebank: 86% SR vs 13% OR (Hale 2001)

German NEGRA: 74% SR vs 26% OR (Skut et al. 1997)

Experience-based: SR << OR

6


Outline

• Introduction

• Chinese RC

• Modeling

• Conclusion

7


Hsiao & Gibson (2003)

SR: [ei yaoqing fuhao de] guanyuani da-le jizhe

invite tycoon DE official hit reporter

‘The official who invited the tycoon hit the reporter.’

OR: [fuhao yaoqing ei de] guanyuani da-le jizhe

tycoon invite DE official hit reporter

‘The official who the tycoon invited hit the reporter.’

Memory-based: SR >> OR

Experience-based: SR << OR

8


Lin & Bever (2006)

• Subject-modifying SR (SR-S)

• [ei yaoqing fuhao de] guanyuani da-le jizhe

• Subject-modifying OR (OR-S)

• [fuhao yaoqing ei de] guanyuani da-le jizhe

9


Lin & Bever (2006)





• Object-modifying SR (SR-O)

• jizhe da-le [ei yaoqing fuhao de] guanyuani

9


Lin & Bever (2006)







• Object-modifying OR (OR-O)

• jizhe da-le [fuhao yaoqing ei de] guanyuani

9


Lin & Bever (2006, 2011)

CHINESE RELATIVE CLAUSES 15

V1(src)/N1(orc) N1(orc)/V1(src) DE N2(head)

600

700

800

900

1000

1100

Read

ing

Tim

e (m

s)

SR−SOR−SSR−OOR−O

Figure 3. Mean reading times in milliseconds for each condition in Experiment 2a.


600

700

800

900

1100

1300

1500

Read

ing

Tim

e (m

s)


Figure 4. Mean reading times in milliseconds for each condition in Experiment 2a.10


Lin and Bever (2006)







• Object-modifying OR (OR-O)

• jizhe da-le [fuhao yaoqing ei de] guanyuani

reporter hit tycoon invite11


Chen, Li, Kuo, Vasishth (submitted)


Analyses of reading times. Table 12 of the Appendix lists the word-by-word meanRTs of the region that includes two words before and after “DE”. Figure 2 plots RTsacross four RT types that refer to conditions in (10). A direct reading time comparison atthe head noun is presented in Figure 4. Table 4 reports the results of statistical analyses.

Over the five words taken as a whole, subject extractions (a and c) were read88.0 ms faster than object extractions (b and d) (p<0.05). At the head noun, we see amarginal SR advantage in subject-modifying conditions (a vs b) (t=0.79, n.s.). If RCsmodified the matrix object, this SR preference was intensified (c vs d) (t=2.32, p<0.05).Adding a spillover predictor in the model will not change the SR advantage in eithercontrast. At the next word of the head noun, the SR preference persisted.

V1(src)/N1(orc) N1(orc)/V1(src) DE N2(head) N2+1

450

550

650

750

Read

ing

Tim

e (m

s)



Table 4: Summary of statistical analyses of Experiment 2a

Experiment Region Contrast/Predictor Coefficient Std. Error t value

2a head nouna vs b 0.012 0.0153 0.79

c vs d 0.037 0.0159 2.35

2a head nouna vs b 0.006 0.0148 0.41

c vs d 0.033 0.0153 2.13spillover 0.350 0.0308 11.39

Experiment 2b: Method

Participants. Experiment 2b was also conducted in Dalian, China. However, theparticipants were 60 college students who did not take Experiment 2a. Each participant

12


Outline

• Introduction

• Chinese RC

• Modeling

• Conclusion

13


Surprisal (Hale, 2001)

14



• is a model of sentence processing difficulty

14




• is built on a language model, such as a PCFG

14





• quantifies the “unlikelihood” (surprise) of integrating a word in the sentence

14





• quantifies the “unlikelihood” (surprise) of integrating a word in the sentence

• is used to predict word-by-word reading times

14


Surprisal: an example

A six-word string:

0 The 1 horse 2 raced 3 past 4 the 5 barn 6

15



A six-word string:

0 The 1 horse 2 raced 3 past 4 the 5 barn 6S

NP

DT

the

NN

horse

VP

VBD

raced

PP

IN

past

NP

DT

the

NN

barn

S

NP

NP

DT

the

NN

horse

VP

VBN

raced

PP

IN

past

NP

DT

the

NN

barn

VP

S

NP

NP

DT

the

NN

horse

VP

VBN

raced

PP

IN

past

NP

DT

the

NN

barn

VP

VBD

fell

3

15



A six-word string:

0 The 1 horse 2 raced 3 past 4 the 5 barn 6

main-clause reading >> reduced RC reading

S

NP

DT

the

NN

horse

VP

VBD

raced

PP

IN

past

NP

DT

the

NN

barn

S

NP

NP

DT

the

NN

horse

VP

VBN

raced

PP

IN

past

NP

DT

the

NN

barn

VP

S

NP

NP

DT

the

NN

horse

VP

VBN

raced

PP

IN

past

NP

DT

the

NN

barn

VP

VBD

fell

3

15



S

NP

DT

the

NN

horse

VP

VBD

raced

PP

IN

past

NP

DT

the

NN

barn

S

NP

NP

DT

the

NN

horse

VP

VBN

raced

PP

IN

past

NP

DT

the

NN

barn

VP

S

NP

DT

the

NN

horse

VP

VBD

raced

PP

IN

past

NP

DT

the

NN

barn

S

NP

NP

DT

the

NN

horse

VP

VBN

raced

PP

IN

past

NP

DT

the

NN

barn

VP

VBD

fell

3

The next word leads to a structural reanalysis:

0 The 1 horse 2 raced 3 past 4 the 5 barn 6 fell 7

16



S

NP

DT

the

NN

horse

VP

VBD

raced

PP

IN

past

NP

DT

the

NN

barn

S

NP

NP

DT

the

NN

horse

VP

VBN

raced

PP

IN

past

NP

DT

the

NN

barn

VP

S

NP

DT

the

NN

horse

VP

VBD

raced

PP

IN

past

NP

DT

the

NN

barn

S

NP

NP

DT

the

NN

horse

VP

VBN

raced

PP

IN

past

NP

DT

the

NN

barn

VP

VBD

fell

3

Xmain-clause reading



16



S

NP

DT

the

NN

horse

VP

VBD

raced

PP

IN

past

NP

DT

the

NN

barn

S

NP

NP

DT

the

NN

horse

VP

VBN

raced

PP

IN

past

NP

DT

the

NN

barn

VP

S

NP

DT

the

NN

horse

VP

VBD

raced

PP

IN

past

NP

DT

the

NN

barn

S

NP

NP

DT

the

NN

horse

VP

VBN

raced

PP

IN

past

NP

DT

the

NN

barn

VP

VBD

fell

3

Xmain-clause reading

√reduced-relative reading



16


Surprisal: the calculation


S

NP

DT

the

NN

horse

VP

VBD

raced

PP

IN

past

NP

DT

the

NN

barn

S

NP

NP

DT

the

NN

horse

VP

VBN

raced

PP

IN

past

NP

DT

the

NN

barn

VP

VBD

fell

surprisal = log2(αn−1

αn) surprisal = log2(

α(0,6)

α(0,7))

5

17


Probabilistic Grammar

0.9261 S → NPSBJ VP

0.0739 S → NPRC VP 0.5 V → yaoqing0.5935 NPRC → CPSR NP 0.5 V → dale0.4065 NPRC → CPOR NP 0.3333 N → fuhao0.8615 VP → V NPOBJ 0.3333 N → guanyuan0.1385 VP → V NPRC 0.3334 N → jizhe1.0 CPSR → VP DEC 1.0 DEC → de1.0 CPOR → S/NP DEC 1.0 NP/NP → �1.0 S/NP → NPSBJ VP/NP 1.0 pro → �1.0 VP/NP → V NP/NP0.4215 NPSBJ → NP

0.5785 NPSBJ → pro

0.9928 NPOBJ → NP

0.0072 NPOBJ → pro

1.0 NP → N

Table 2: A richer PCFG for Chinese Relative Clauses with pro-drop rules

Type Count

noun subject 10870

pro subject 14921

RC subject 2057

noun object 14041

pro object 109

RC object 2273

SRC 895

ORC 613

Table 3: Attestation count from Chinese Treebank 6.0

4.2 Parser

An statistical prefix parsing system (Grove, 2010) was used to obtain surprisals for the prefix

in each of the four examples. It employs a bottom-up chart parsing strategy in the style of

Shieber et al. (1995). The parser constructs a probabilistic model at each prefix calculating

the inside probability of each non-terminals. For example, . . .

If more than one parse is available. The total probability should be the summation of

the probability of all parses.

5 Results

Starting from the simpler grammar.

Subject-modifying Relative Clauses as in Figure 3(a)

5

18

0.9261 S → NPSBJ VP

0.0739 S → NPRC VP 0.5 V → yaoqing0.5935 NPRC → CPSR NP 0.5 V → dale0.4065 NPRC → CPOR NP 0.3333 N → fuhao0.8615 VP → V NPOBJ 0.3333 N → guanyuan0.1385 VP → V NPRC 0.3334 N → jizhe1.0 CPSR → VP DEC 1.0 DEC → de1.0 CPOR → S/NP DEC 1.0 NP/NP → �1.0 S/NP → NPSBJ VP/NP 1.0 pro → �1.0 VP/NP → V NP/NP0.4215 NPSBJ → NP

0.5785 NPSBJ → pro

0.9928 NPOBJ → NP

0.0072 NPOBJ → pro

1.0 NP → N

Table 2: A richer PCFG for Chinese Relative Clauses with pro-drop rules

Type Count

noun subject 10870

pro subject 14921

RC subject 2057

noun object 14041

pro object 109

RC object 2273

SRC 895

ORC 613

Table 3: Attestation count from Chinese Treebank 6.0

4.2 Parser

An statistical prefix parsing system (Grove, 2010) was used to obtain surprisals for the prefix

in each of the four examples. It employs a bottom-up chart parsing strategy in the style of

Shieber et al. (1995). The parser constructs a probabilistic model at each prefix calculating

the inside probability of each non-terminals. For example, . . .

If more than one parse is available. The total probability should be the summation of

the probability of all parses.

5 Results

Starting from the simpler grammar.

Subject-modifying Relative Clauses as in Figure 3(a)

5


Parser

• bottom-up chart-parsing algorithm (Goodman, 1999)

• the incremental parser considers multiple parses at each position.

19


Results: Subj-modifying RCTotal:

SR 8.89OR 10.47

02

46

8Surprisal

V1(SR)N1(OR) N1(SR)V1(OR) DE N2(head)

1.74

2.90

1.81

1.0

3.76

4.99

1.58 1.58

SROR

20


Results: Subj-modifying RC

1.23

Total:SR 8.89

OR 10.47

02

46

8Surprisal


1.74

2.90

1.81

1.0

3.76

4.99

1.58 1.58

SROR

20


Calculate surprisals

• In SR-S, before “de”: V N “invite tycoon ...”

S

NPSBJ

pro

VP

V

yaoqing

NPOBJ

NP

N

fuhao

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

NP

VP

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

de

NP

VP

S

NPSBJ

NP

N

fuhao

VP

V

yaoqing

NPOBJ

S

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V

yaoqing

NP/NP

�

DEC

NP

VP

1

(1) pro-drop main clause (2) SR0

24

68

Surprisal


1.74

2.90

1.81

1.0

3.76

4.99

1.58 1.58

SROR

21



• In SR-S, after “de”: V N de “invite tycoon de ...”

S

NPSBJ

pro

VP

V

yaoqing

NPOBJ

NP

N

fuhao

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

NP

VP

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

de

NP

VP

S

NPSBJ

NP

N

fuhao

VP

V

yaoqing

NPOBJ

S

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V

yaoqing

NP/NP

�

DEC

NP

VP

1

DE

X(1) pro-drop main clause (2) SR

02

46

8Surprisal


1.74

2.90

1.81

1.0

3.76

4.99

1.58 1.58

SROR

22



• In OR-S, before “de”: N V “tycoon invite ...”

S

NPSBJ

pro

VP

V

yaoqing

NPOBJ

NP

N

fuhao

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

de

NP

VP

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

NP

VP

S

NPSBJ

NP

N

fuhao

VP

V

yaoqing

NPOBJ

S

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V

yaoqing

NP/NP

�

DEC

NP

VP

S

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V

yaoqing

NP/NP

�

DEC

de

NP

VP

1

S

NPSBJ

pro

VP

V

yaoqing

NPOBJ

NP

N

fuhao

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

de

NP

VP

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

NP

VP

S

NPSBJ

NP

N

fuhao

VP

V

yaoqing

NPOBJ

S

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V

yaoqing

NP/NP

�

DEC

NP

VP

S

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V

yaoqing

NP/NP

�

DEC

de

NP

VP

1

(1) main clause (2) OR0

24

68

Surprisal


1.74

2.90

1.81

1.0

3.76

4.99

1.58 1.58

SROR

23



• In OR-S, after “de”: N V “tycoon invite de ...”

S

NPSBJ

pro

VP

V

yaoqing

NPOBJ

NP

N

fuhao

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

de

NP

VP

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

NP

VP

S

NPSBJ

NP

N

fuhao

VP

V

yaoqing

NPOBJ

S

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V

yaoqing

NP/NP

�

DEC

NP

VP

S

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V

yaoqing

NP/NP

�

DEC

de

NP

VP

1

S

NPSBJ

pro

VP

V

yaoqing

NPOBJ

NP

N

fuhao

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

de

NP

VP

S

NPRC

CPSR

VP

V

yaoqing

NPOBJ

NP

N

fuhao

DEC

NP

VP

S

NPSBJ

NP

N

fuhao

VP

V

yaoqing

NPOBJ

S

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V

yaoqing

NP/NP

�

DEC

NP

VP

S

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V

yaoqing

NP/NP

�

DEC

de

NP

VP

1

(1) main clause (2) OR

XDE

02

46

8Surprisal


1.74

2.90

1.81

1.0

3.76

4.99

1.58 1.58

SROR

24


Results: Obj-modifying RCTotal:

SR 8.04OR 9.61

02

46

8Surprisal


4.17

1.822.25

6.21

0.04 0

1.58 1.58

SROR

25


Results: Obj-modifying RC

3.96

Total:SR 8.04OR 9.61

02

46

8Surprisal


4.17

1.822.25

6.21

0.04 0

1.58 1.58

SROR

25



SR reading has already been recognized.

S

NPSBJ

N

jizhe

VP

V

dale

NPRC

CPSR

VP

V

yaoqing

NPOBJ

DEC

NP

2

S

NPSBJ

N

jizhe

VP

V

dale

NPRC

CPSR

VP

V

yaoqing

NPOBJ

DEC

NP

2

fuhao0

24

68

Surprisal


4.17

1.822.25

6.21

0.04 0

1.58 1.58

SROR

26

• In SR-O, N V V “reporter hit invite ...” → N V V N



• In OR-O, N V N “reporter hit tycoon ...”

S

NPSBJ

N

jizhe

VP

V

dale

NPRC

CPSR

VP

V

yaoqing

NPOBJ

DEC

NP

S

NPSBJ

N

jizhe

VP

V

dale

NPOBJ

NP

N

fuhao

S

NPSBJ

N

jizhe

VP

V

dale

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V NP/NP

DEC

NP

2

(1) main clause (2) OR0

24

68

Surprisal


4.17

1.822.25

6.21

0.04 0

1.58 1.58

SROR

27



• In OR-O, N V N V “reporter hit tycoon invite ...”

S

NPSBJ

N

jizhe

VP

V

dale

NPRC

CPSR

VP

V

yaoqing

NPOBJ

DEC

NP

S

NPSBJ

N

jizhe

VP

V

dale

NPOBJ

NP

N

fuhao

S

NPSBJ

N

jizhe

VP

V

dale

NPRC

CPOR

S/NP

NPSBJ

NP

N

fuhao

VP/NP

V NP/NP

DEC

NP

2

(1) main clause

Xyaoqing

(2) OR0

24

68

Surprisal


4.17

1.822.25

6.21

0.04 0

1.58 1.58

SROR

28


Outline

• Introduction

• Chinese RC

• Modeling

• Conclusion

29


Conclusion

30


Conclusion

Surprisal

• uses structural frequencies as a reflection of readers’ linguistic experience

• models the resolution of incremental ambiguity

30


Conclusion

Surprisal

• uses structural frequencies as a reflection of readers’ linguistic experience

• models the resolution of incremental ambiguity

Results are consistent with recent empirical data

• argue against the memory-based account

• support the experience-based account

30


Acknowledgment

• Chien-Jer Charles Lin (Indiana)

• Shravan Vasishth (Potsdam)

• Jiwon Yun (Cornell)

• This work is supported by

• Cornell Cognitive Science Program

• Cornell East Asian Program

• a NSF CAREER Award (0741666) to JTH

31


谢谢！ Thank you!

32


ReferencesGibson, E. (1998). Linguistic Complexity: Locality of Syntactic Dependencies. Cognition, 68: 1–76.Gibson, E. (2000). The dependency locality theory: A distance-based theory of linguistic complexity.

In Miyashita, Y., Marantz, A., & O'Neil, W. (Eds.), Image, language, brain (pp. 95-126), Cambridge, MA: MIT Press.

Goodman, J. (1999). Semiring parsing. Computational Linguistics, 25, 573–605.

Grodner, D. & Gibson, E. (2005). Consequences of the serial nature of linguistic input. Cognitive

Science, 29,261-291.Hale, J. T. (2001). A probabilistic Earley parser as a psycholinguistic model. In Proceedings of the Second

Meeting of the North American Chapter of the Association for Computational Linguistics. Pittsburgh, PA.

Hsiao, F. & Gibson, E. (2003). Processing relative clauses in Chinese. Cognition, 90, 3-27.

Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106(3):1126-1177. Lewis, R. & Vasishth, S. (2005). An activation-based model of sentence processing as skilled memory

retrieval. Cognitive Science, 29:1-45

Lin, C.-J. C., & Bever, T. G. (2006). Subject preference in the processing of relative clauses in chinese. In D. Baumer, D. Montero, & M. Scanlon (Eds.), Proceedings of the 25th WCCFL (p. 254-260). Cascadilla Proceedings Project.

Mitchell, D. C., Cuetos, F., Corley, M. M. B., & Brysbaert, M. (1995). Exposure-based models of human parsing: Evidence for the use of coarse-grained (nonlexical) statistical records. Journal of Psycholinguistic Research, 24, 469-488.

33


34


Lin and Bever (2011)



600

700

800

900

1000

1100

Read

ing

Tim

e (m

s)




600

700

800

900

1100

1300

1500

Read

ing

Tim

e (m

s)



No strong effect

35


Structural Expectations in Chinese Relative Clause...

Documents

Transcript of Structural Expectations in Chinese Relative Clause...