Transcript Frequency effects in the processing of Chinese inflection

Frequency
effects in the
processing
of Chinese
inflection
1
James Myers
Yu-chi Huang
National Chung Cheng University
Taiwan
Assistants:
Wen-Ling Wang, Pei-Ying Chou
Funded by National Science Council (Taiwan)
NSC-90-2411-H-194-023
First author email: [email protected]
2
Overview
• Frequency affects the processing of English
regular inflection, but only in speeded tasks.
• Inflection in Chinese is even more
transparent than in English (though there
are semantic selectional restrictions).
• Frequency effects are thus unexpected in
Chinese inflection, yet a series of
experiments suggests that they do exist.
3
English inflection and rules
• Advocates of real-time inflectional rule
application point to frequency effects as
evidence for their position (Pinker 1999).
• In particular, off-line judgment tasks don’t
show frequency effects (e.g. Ullman 1999),
presumably because in such tasks rule
application applies extra-lexically.
4
English inflection in real time
• However, even real-time rule advocates
admit that speeded tasks do show frequency
effects (e.g. Sereno & Jongman 1997).
• Pinker (1999) claims this is consistent with
his position, since retrieval of memory
traces (inevitably formed during earlier rule
applications) is more efficient than actual
rule application when speed is crucial.
5
Sereno & Jongman (1997)
[Experiments 2A-B]
• The relative frequencies of base vs. plural
were varied for regular English nouns:
– high-base/low-plural: river-rivers
– low-base/high-plural: window-windows
• Higher frequencies sped up responses for
both bases and plurals:
– river RT < window RT
– rivers RT > windows RT
6
Sereno & Jongman (1997)
[Experiments 3A-B]
• Base frequencies were matched and only
plural frequencies were varied:
– equal-base/low-plural: desk-desks
– equal-base/high-plural: tree-trees
• Higher-frequency plurals were responded
to more quickly:
– desk RT = tree RT
– desks RT > tress RT
7
Relevance to Chinese
Chinese is notoriously “isolating”, so finding
frequency effects with Chinese inflection
would be particularly dramatic evidence that
even highly regular morphological processes
involve lexical retrieval, at least in speeded
tasks.
8
Chinese inflection?
• Chinese does have grammatical morphemes
that serve inflection-like functions, such as:
– plural marker men (們), e.g. ren-men (“people”)
– perfective aspect marker le (了), e.g. sidiao-le
(“died”)
– experiential aspect marker guo (過), e.g. qu-guo
(“went previously”)
– durative aspect marker zhe (著), e.g. kan-zhe
(“watching”)
9
It’s not derivation
• There is agreement between the plural
marker and classifiers (Li 1999):
– xuesheng(-men) “students”
– san-ge xuesheng(*-men) “three students”
• Similarly, interactions between the verb
aspect markers and other elements in the
sentence have been well documented
(Huang 1982, Ma 1985, Chiu 1993)
10
Inflection or cliticization?
• Like clitics, the above morphemes:
– have no lexical exceptions
– show no allomorphy
– are highly productive (see e.g. Packard 1993)
• Yet they also have selectional restrictions,
an affix diagnostic (Zwicky & Pullum 1983)
11
Selectional restrictions
• The plural marker men can only appear with
human nouns
• Aspect markers only appear with particular
verb classes, e.g. zhe with the following:
– verbs of activity: pao-zhe “running”
– verbs of state: ai-zhe “loving”
– *sidiao-zhe “dying” and *qu-zhe “going” are
bad since these verbs imply change of state
12
Isolating Chinese inflection
• Whether these morphemes are affixes or
clitics, the orthography and phonology of
Chinese make them trivial to parse out:
– 記載著 jizaizhe
– 批判著 pipanzhe


記載 + 著
批判 + 著
• Thus, given the logic of real-time rule
application, the frequency of verb + aspect
sequences shouldn’t affect reading times.
13
Experiment overview
• Following Sereno & Jongman (1997), we
presented readers with base forms and
“suffixed forms” with varying frequencies
• Following Huang (2001) we chose the
durative marker zhe (著) :
– aspect markers are more clitic-like than plural
and thus more challenging to frequency claims
– unlike le (了) and guo (過), zhe (著) has no
easily confusable homonyms
14
Material selection
• Only verb-verb compounds were used since
zhe is exclusively suffixing with these
• Frequencies were measured using a Taiwan
Mandarin corpus (Chen et al. 1996)
• To supplement the low zhe form counts in
the corpus, we also counted websites using
chinese.yahoo.com (highly correlated
with corpus counts: r > 0.8)
15
Frequency manipulations
Frequency relations were similar to Exp. 2 of
Sereno & Jongman (1997) (highlighted
frequencies were the ones we measured):
HFZ (have high-frequency zhe forms)
Total frequency for string 記載 ("record")
Freq. of 記載著
Freq. of 記載 alone Other
LFZ (have low-frequency zhe forms)
Total frequency for string 批判 ("criticize")
Other
Freq. of 批判著
Freq. of 批判 alone
16
Experiments 1a-b methods
• 15 HFZ and 15 LFZ base forms matched for
total frequency (also character frequency)
• 15 HFZ and 15 LFZ zhe forms differing in
frequency but matching in acceptability
(pretested off-line with 4-point scale)
• Separate visual lexical decision tasks for
base (Exp 1a) and zhe forms (Exp. 1b)
• 20 participants each
• Errors and RTs > 2SD from mean rejected
17
Experiments 1a-b results
730
RT (msec)
• The difference
in base forms
just missed
significance
(p = 0.06)
• The difference
in zhe forms just
reached
significance
(p = 0.03)
710
HFZ
LFZ
690
670
650
630
base (Exp. 1a)
zhe (Exp. 1b)
18
Semantics-frequency confound?
• Acceptability scores for zhe forms (4=least
acceptable) are correlated with frequency:
– according to corpus: r = -0.24, p = 0.1
– according to web: r = -0.42, p = 0.006
• A new off-line task with more sensitive 7point scale (7=most acceptable) revealed
difference in acceptability for HFZ and LFZ
zhe forms in Exps 1a-b (6.0 vs.4.9, p<0.001)
19
Experiments 2a-b methods
• Same as Exps. 1a-b, except:
– New sets of 12 HFZ, 12 LFZ base/zhe forms
– Acceptability more carefully controlled
(pretested with 7-point scale)
• Control of acceptability affected frequency:
– Exps. 2a-b had lower base form frequencies
– Exps. 1a-b: LFZ zhe forms about 20 times less
frequent than HFZ zhe forms
– Exps. 2a-b: LFZ zhe forms only about 7 times
less frequent than HFZ zhe forms
20
Experiments 2a-b results
730
RT (msec)
• Neither
difference was
anywhere near
significance
(p = 0.7)
710
HFZ
LFZ
690
670
650
630
base (Exp. 2a)
zhe (Exp. 2b)
21
Conclusions
• Unlike English, Mandarin inflection shows
significant frequency-acceptability
correlations in off-line judgment tasks
• Like English, on-line response times are
affected by the frequency of inflected
forms, but due to the above confound, it is
unclear if this is a “pure” frequency effect
• Like English, the frequency of uninflected
forms also affects response times
22
Appendix: Exp. 1 materials
• Exp 1a: base forms
HFZ-base
孕育
守護
思索
洋溢
流傳
盼望
記載
追逐
陳列
散發
期盼
凝聚
擁抱
隱藏
籠罩
LFZ-base
反抗
抗拒
批判
防守
呼叫
爭奪
爭論
穿梭
負荷
搭乘
運送
逼近
違背
銜接
讚美
• Exp 1b: zhe forms
HFZ-zhe
孕育著
守護著
思索著
洋溢著
流傳著
盼望著
記載著
追逐著
陳列著
散發著
期盼著
凝聚著
擁抱著
隱藏著
籠罩著
LFZ-zhe
反抗著
抗拒著
批判著
防守著
呼叫著
爭奪著
爭論著
穿梭著
負荷著
搭乘著
運送著
逼近著
違背著
銜接著
讚美著
23
Appendix: Exp. 2 materials
• Exp 2a: base forms
HFZ-base
包圍
打量
伴隨
夾雜
抱持
玩弄
閃動
搖晃
跟隨
摸索
環繞
蘊藏
LFZ-base
交談
交織
依靠
珍藏
追隨
塗抹
感嘆
遍佈
漂浮
滿載
編織
襯托
• Exp 2b: zhe forms
HFZ-zhe
包圍著
打量著
伴隨著
夾雜著
抱持著
玩弄著
閃動著
搖晃著
跟隨著
摸索著
環繞著
蘊藏著
LFZ-zhe
交談著
交織著
依靠著
珍藏著
追隨著
塗抹著
感嘆著
遍布著
漂浮著
滿載著
編織著
襯托著
24
Selected references
Chen, K.-J., Huang, C.-R., Chang, L.-P., & Hsu, H.-L. (1996). SINICA
CORPUS: design methodology for balanced corpora. Language,
Information and Computation, 11, 167-176.
Huang, Y-C. (2001). Frequency effects on the processing of an aspect
marker in Mandarin. Unpublished master's thesis, National Chung
Cheng University, Chiayi, Taiwan.
Sereno, J. A., & Jongman, A. (1997). Processing of English inflectional
morphology. Memory and Cognition, 25(4), 425-437.
Ullman, M. T. (1999). Acceptability ratings of regular and irregular pasttense forms: Evidence for a dual-system model of language from word
frequency and phonological neighborhood effects. Language and
Cognitive Processes, 14 (1), 47-67.
25