PROCESSING TONE SEQUENCES IN STANDARD MANDARIN

Isabelle LinM.A. thesis2015-2016

TONE, SEGMENTS, AND LEXICAL INFORMATION

v4 tones in Standard Mandarin

vThe classic example: ma

mother linen horse scold

1 妈 2 麻 3 马 4 骂

vFor successful retrieval of the correct lexical item, both segmental and tonal information must be processed

vHow informative is tone?

2

TONE, SEGMENTS, AND LEXICAL INFORMATION

vBack to the classic example: ma

mother linen horse scold

1 妈 2 麻 3 马 4 骂

vBut also:

symbol agate

3码 3 玛

vAnd then:

password jetty

mi4 ma3 密码 码头 ma3 tou2 …

vAccessing segments + tone on a monosyllable does not provide access to an definitive lexical item

3

TONE IS LESS INFORMATIVE THAN SEGMENTS

vHomophone decision: Cantonese and Mandarin speakers are slower to respond to tone mismatches than segmental mismatches (Taft & Chen, 1992) (qu3-qu4 slower than qi4 vs. qu4)

vLexical decision: Cantonese speakers often accept tone mismatch non-words as words more often than vowel mismatches (Cutler & Chen, 1997)(real word /bok8-si6/, tone mismatch /bok8-si2/ vowel mismatch /bok8-sy6/)

ØSegmental content alone provides reliable information: speakers can make a confident guess at the corresponding lexical entry

4

TONE AND SEGMENT INFORMATION ARE SEPARABLEvGanong effect on tone categorization: synthetic monosyllabic token ambiguous between two tones: participants give tone responses that would make it a real word. (Fox and Unkefer, 1985)

ØLexical entries are stored with tone information and participants use this tone information from the lexicon to supplement insufficient tone information in the signal.

5

TONE INFORMATION IS AVAILABLE LATER

vTone identification in Mandarin depends on a pitch contour, tone information becomes available towards the middle or late portions of the syllable (Howie,1976, Whalen and Xu, 1992)

vDetection of semantic anomalies, ERP: Mandarin speakers show earlier ERP components for anomalies induced by segmental information (Brown-Schmidt, Canseco-Gonzalez, 2004)

vCOHORT model of lexical retrieval (Marslen-Wilson, 1984): an auditory input triggers the activation of all possible word candidates (lexical access), followed by competition among candidates before a unique item is retrieved (lexical selection)

va difference in processing timelines might cause tonal information to be active at a later stage during lexical retrieval.

6

PRIMING EFFECTS OF SEGMENTS AND TONE

vMatching segmental primes are facilitatory (Lee, 2007, Sereno and Li, 2015)

vAuditory priming: Mandarin speakers show facilitation for identity primes at 250 msISI, but not with a segmental-only (lou2-lou3) or tone-only match (lou2-mang2). At 50 ms, facilitation is found for the segmentally matched pair (lou2-lou3) (Lee, 2007)

vLexical decision and shadowing: auditory primes mismatching in segmental content but matching in tone with the target (mo3-ba3) caused a response delay (Poss, Hung and Will, 2008)

vAll previous studies on monosyllabic contrasts

vPriority to segmental information: in terms of meaning carried and timeline. How efficient is a ‘segments first’ strategy for identifying words?

7

FUNCTIONAL LOAD AND INFORMATION LOSS

Surendran and Niyogi (2003)

vBased on Shannon’s entropy:

vThe functional load of a given contrast is the relative difference (in percentage) in entropy between L and L’ (L without the contrast):

vThis estimates the increased homophony induced by neutralization of the contrast

8

INFORMATIVENESS OF SEGMENTS AND TONE

v Measure of functional load (Oh, Pellegrino, Coupé & Marsico, 2013). Estimates the increased homophony induced by neutralization of the contrast

v Tones in Mandarin are about as informative as vowels, both much less informative than consonants

9adapted from Oh, Pellegrino, Coupé & Marsico, 2013

EXPERIMENT 1: INFORMATIVENESS OF TONE

vMost legal monosyllabic segments+tone combinations can form independent words/morphemes, but every such combination can correspond to a number of homophones

vMost content words are instead disyllabic compounds, the other syllable of the compound disambiguates the target morpheme

ØStudies using monosyllabic stimuli might underestimate the contribution of tone information to lexical access.

vTo assess whether this is the case, we examined corpus data, using %of corpus as proxy for % of lexicon

10

MANDARIN MONOSYLLABLES

vMonosyllables: 11.15% of the SUBTLEX-CH corpus (Cai and Brysbaert, 2010). The corpus contains 46.8 million characters and 33.5 million words, from movie and television subtitles

v For the same pinyin sequence (segmental content), there are on average12 homophones

v 0.64% of the corpus can be defined using segments only (5.7% of monosyllables)

v 0.68% of the corpus can be defined using segments + tone (5.9% of monosyllables)

11

12

MANDARIN DISYLLABLES

vDisyllables: 65.56 % of the SUBTLEX-CH corpus

vFor the same pinyin sequence (segmental content), there are on average 1 word (1.53)

v51.26% of the corpus can be defined using segments only (78% of disyllables)

v61.94% of the corpus can be defined using segments + tone (94% of disyllables)

13

14

INFORMATIVENESS OF SEGMENTS AND TONE

v Fisher’s exact test: the number of words with homophones after segments + tone have been specified significantly differed by word length (p < 0.001)

vTone contributes more in defining unique disyllables than monosyllables

v Looking at monosyllables would underestimate the informativeness of tone

v There is a subset of segmental sequences where only tone disambiguates two segmentally identical homophones. (10.94% of corpus). We will use them to investigate the role of tone.

15

DISTRIBUTION OF TONE SEQUENCES

vOne possibility is that the effect of tone on Mandarin lexical access is modulated by how likely it is to encounter a certain sequence of tones in the language.

vSpeakers are sensitive to word frequencies, and segment phonotactics, maybe also to tone sequence frequencies?

vFor disyllabic words, we need to examine frequency for sequences of two tones rather than individual tones. There are at least two possible ways in which one can measure frequency of tone sequences in disyllables, which we do in Experiment 2

16

EXPERIMENT 2: TWO MEASURES OF TONE SEQUENCE FREQUENCYvHow likely is it for a disyllabic word to bear a given TaTb sequence?

Øcomputed using SUBTLEX-CH (46.8 million syllables)ØDitone frequency

vHow likely is it for a given TaTb sequence to occur in running speech?

Øcomputed using a corpus of semi-spontaneous speech (22571 syllables)ØTone bigram frequency

17

A CORPUS OF SEMI-SPONTANEOUS SPEECH

v 9 native speakers of Standard Mandarin from Beijing, China (5 female, 4 male, ages 19-53)

v Semi-spontaneous dialogue with a friend, optional conversation prompts provided

v1 hour per pair, separate track for each speaker (double room studio)

vHand-segmented at the syllable level, coded for pinyin, tone, utterance-level translation

vAutomatic tagging for PoS, gloss (from CC-CEDICT)

18

19

ALTERNATIVE MEASURES OF TONE SEQUENCE FREQUENCYvDitone frequency: How likely is it for a disyllabic word to bear a given TaTbsequence?

vTone bigram frequency: How likely is it for a given TaTb sequence to occur in running speech?

vComparing the two measures:

vTwo-sample Kolmogorov-Smirnov test for equality of distribution functions: the distributions as defined by ditone and bigram measures were significantly different (p < .05)

20

EXPERIMENT 3: WORD RECOGNITION

vThere is a subset of disyllabic segmental sequences where only tone disambiguates two segmentally identical homophones. (10.94% of corpus)

vIf the preferred strategy for processing disyllabic words is:vrely on segmental content for lexical access, activating all segmental homophonesvuse tone information to complete lexical selection among candidates

v Then, on such items, listeners’ recognition of disyllabic minimal pairs could be biased by tone primes matching one or the other candidate

vTone frequency was indexed using ditone vs. bigram frequencies to determine the appropriate measure

21

HYPOTHESIS

v Overall, we expected participants to be more inclined to choose the more frequent word alternative.

vHowever, if tone sequences are decisive in lexical selection, we expected that the tone sequence on the prime should also influence the identification of the ambiguous target as one or the other word candidate.

22

EXPERIMENT 3: STIMULI

vDisyllabic lexicalized compounds

vSame segmental structure, different words distinguished by tone sequence

vReduce possibilities: only 2 lexicalized alternatives

example： juli 42 距离 distance 34 举例 to give an example

(This, for instance, was not usable：

jishi 23 即使 even if 22 及时 on time 24 集市 market 24 急事 urgent matter)

vAlternatives are orthographically, semantically unrelated

v30 pairs

23

STIMULI

vRecorded from native speaker:

alternative a 42 距离 distance

alternative b 34 举例 to give an example

same syllables with flat pitch

v Praat script: average pitch contour on the two alternatives, normalized for durationtransfer the averaged contour onto the flat pitch item

resulting test item

v In half the items, the most frequent alternative bore the most frequent ditonesequence

24

STIMULI

vPrime with tones matching alternative a or alternative b, or with unrelated tones, word frequency is average of that of the two alternatives

42 至于 as to 34 午夜 midnight 15称呼 to address

Test item

vFillers: 60 pairs of unrelated words

- overall word frequency and tone sequence frequency distribution matched to that of targets + primes

- pitch contours of the two words of a pair are averaged- resulting contour added on one member of the pair

25

FREQUENCY CONDITIONS

26

vCondition A: the more frequent word bears the more frequent tone sequence. This word should be maximally salient during lexical access. No way to disambiguate between the effects of tone or word frequency.

vConditions B, C and D: the more frequent word does not bear the more frequent tone sequence.

vIf word frequency favours one candidate while tone sequence frequency favours the other (condition D), participants’ preference for the more frequent word should not be as strong, and therefore such items are most likely to be sensitive to tone priming.

PROCEDUREv47 speakers of Standard Mandarin (36 female, 11 male, mean age = 22, SD = 6, range 18-50)

vcan handwrite simplified characters

vInstructions: identify the word pronounced strangely

vConstrain responses without indicating overtly that there is ambiguity between 2 alternatives

vHandwritten responses:

preserve ‘one word’ impression

detect ad hoc compounding

vRecording: in case handwriting is unclear

v30 test trials + 60 filler trials, 2 groups counterbalanced for prime condition

v~30 - 40 minutes27

PROCEDURE

28

DATAvTest items: look at image to determine response

vExcluded ad hoc compounds/other words

vResponses coded for:match primehigher word frequency alternative (categorical)higher tone sequence frequency alternative (categorical)

v5 participants excluded from analysis (1 technical issue, 1 did not follow instructions, 3 gave >2SD of ‘other’ responses)

v6 word pairs were excluded because (1 for >2SD ‘other’ responses, 5 only one alternative given)

29

RESULTS – FREQUENCY MATCH / MISMATCH

30

RESULTS – FREQUENCY MATCH CONDITION (A)

v Items where the more frequent word candidate also bore the more frequent tone pattern (according to both ditone and bigram frequency; condition A)

vWe expected the least amount of priming in this condition.

vmixed-effects logistic regression, prime type as fixed effect, and participant (SUBJ) and ITEM as crossed random effects (random slopes and intercepts).

vParticipants were overall likely to answer with the more frequent word alternative, priming conditions did not significantly affect this preference (68% of more frequent word responses across all priming conditions, above chance at 50%).

vWe cannot distinguish the role of word and tone frequency for these items.

31

RESULTS – FREQUENCY MISMATCH CONDITION (D)vCondition D: the more frequent word alternative bears the less frequent tone pattern (according to both ditone and bigram frequency).

vIf tone sequence frequency makes a word more or less salient during processing, this mismatch should reduce the saliency of the more frequent word. Participants might be more sensitive to priming with tone patterns.

vParticipants were still likely to answer with the more frequent word candidate in the unrelated prime condition (59% of more frequent word responses vs 68% in the frequency match condition - A)

vWhen presented with a prime matching the tone pattern of the more frequent word, participants were less likely to answer with that word (p < .005). Though nonsignificant, the complementary trend was present.

vThis suggests that tone sequence frequency plays a separate role from word frequency and interferes with word frequency information during the processing of disyllables. This interference makes items where the frequencies of words and tone sequences mismatch sensitive to priming effects, but is not strong enough to influence the likelihood of choosing the more frequent word candidate when word and tone sequence frequencies match.

32

RESULTS – COMPARING TONE FREQUENCY MEASURES (B,C)

vWe investigated whether a mismatch in ditone or bigram frequency alone (Condition B or C) shows inhibitory priming of the sort observed in Condition D

vIf frequency mismatches in ditone alone influence lexical processing, then we should see inhibitory priming in condition B, but not C

vIf frequency mismatches in bigram frequency alone influence lexical processing, then we should see inhibitory priming in condition C, but not B.

33

RESULTS – COMPARING TONE FREQUENCY MEASURES (B,C)

34

RESULTS – COMPARING TONE FREQUENCY MEASURES (B,C)vToo few items to conduct statistical comparison, but pattern suggests that frequency mismatch for bigrams (C) is most likely to account for the inhibitory priming (D)

vIf this effect generalizes to more items, this would point to tone bigram frequency as the appropriate measure of tone sequence frequency. Specifically, it would show that Mandarin listeners’ lexical selection is influenced by the likelihood of encountering a given sequence of tones in running speech, regardless of word boundaries.

35

CONCLUSIONS

v Tone is more informative on disyllables than monosyllables

vTo assess frequency of tone sequences, ditone-based and bigram-based measures of tone sequence frequency do provide significantly different measures of tone sequence frequency

vTone frequency information interferes with word frequency information during lexical selection

vAntipriming for frequent tone sequences? If a prime matches the frequent alternative in tone, participants were less likely to answer with the more frequent word

vDid not seem to affect the less frequent word: less competitive alternative to start with?

vSpeaker adaptation issue: same voice on clear (prime) and ambiguous (target) tokens. Contrasting a clear tone sequence to an unclear one, more confidence if frequent tone sequence?

36

37

THANK YOU!

ACKNOWLEGMENTS

Thanks to Kie Zuraw, Megha Sundara and Patricia Keating for advice on design and corpus data collection

RAs: Dong Dai, Jin Weifeng, Wang Junyi, Xiong Yihan, Zhang Haixin

Olivier Wang for help optimizing experiment script

38

REFERENCES

Brown-Schmidt, S., and Canseco-Gonzalez, E. (2004). Who do you love, your mother or your horse? An event-related brain potential analysis of tone processing in Mandarin Chinese. Journal of psycholinguistic research, 33(2), 103-135.

Cai Q, Brysbaert M (2010) SUBTLEX-CH: Chinese Word and Character Frequencies Based on Film Subtitles. PLoS ONE 5(6)

Cutler, A., and Chen, H.-C. (1997). Lexical tone in Cantonese spoken-word processing. Perception and Psychophysics, 59, 165-179

Fox, R. A., and Unkefer, J. (1985). THE EFFECT OF LEXICAL STATUS ON THE PERCEPTION OF TONE. Journal of Chinese Linguistics, 69-90

Howie, J. M. (1976). Acoustical studies of Mandarin vowels and tones (Vol. 6). Cambridge University Press

Lee, C. Y. (2007). Does horse activate mother? Processing lexical tone in form priming. Language and Speech, 50(1), 101-123.

Marslen-Wilson, W. D. (1984). Parallel processes in lexical access. Unpublished manuscript, University of Cambridge

Oh, Y. M., Pellegrino, F., Coupé, C., and Marsico, E. (2013). Cross-language comparison of functional load for vowels, consonants, and tones. Interspeech (August 2013, pp. 3032-3036).

Poss, N., Hung, T. H., and Will, U. (2008). The effects of tonal information on lexical activation in Mandarin. Proceedings of the 20th North American Conference on Chinese Linguistics (NACCL-20) (Vol.1, pp. 205-211).

Sereno, J. A., and Lee, H. (2015). The contribution of segmental and tonal information in Mandarin spoken word processing. Language and Speech, 58(2), 131-151.

Surendran, D., and Niyogi, P. (2003). Measuring the functional load of phonological contrasts. Unpublished manuscript, Chicago, IL.

Taft, M., and Chen, H. C. (1992). Judging homophony in Chinese: The influence of tones. Advances in psychology, 90, 151-172.

Whalen, D. H., and Xu, Y. (1992). Information for Mandarin tones in the amplitude contour and in brief segments. Phonetica, 49(1), 25-47.

39