Durational reduction in English (and Polish) words shows a contextual frequency effectAPAP 2019Kamil Kaźmierski, WA at AMU in PoznańJune 21st, 20191 / 35

Overall frequency effect

Higher lexical frequency → Shorter duration (e.g. Jurafsky et al. 2001)
- Perhaps caused by speakers: frequency of use causes articulatory reduction
- Perhaps moderated by listeners' expectations: highly frequent forms don't have to be pronounced carefully
Such reduction could result from online computation, performed on abstract phonological representations (cf. Levelt et al. 1999)

2 / 35

Contextual frequency effect on variation
 
    Language 
    Effect 
    Source 
  
    English 
    Prevocalic word-final /t/ glottalized more in words typically followed by consonants 
    Eddington & Channer (2010) 
  
    English 
    Word-final /t,d/ deletion more likely in words typically followed by consonants 
    Raymond et al. (2016) 
  
    English 
    Unstressed ING more likely to be /ɪn/ in words frequently occurring in /ɪn/ favoring contexts 
    Forrest (2017) 
  
    Spanish 
    Latin /fV-/ words frequently occurring after word-final non-high vowels likely to be  /ØV-/ in MSS 

    Brown & Raymond (2012) 
  
    New Mexican Spanish 
    Word-initial /s/ more likely to be reduced ([s] › [h] › [Ø]) in words often preceded by word-final non-high vowels 
    Raymond & Brown (2012) 
  
    English 
    Words typically predictable reduce in duration more 
    Seyfarth (2014) 
  
3 / 35

Language	Effect	Source
English	Prevocalic word-final /t/ glottalized more in words typically followed by consonants	Eddington & Channer (2010)
English	Word-final /t,d/ deletion more likely in words typically followed by consonants	Raymond et al. (2016)
English	Unstressed ING more likely to be /ɪn/ in words frequently occurring in /ɪn/ favoring contexts	Forrest (2017)
Spanish	Latin /fV-/ words frequently occurring after word-final non-high vowels likely to be /ØV-/ in MSS	Brown & Raymond (2012)
New Mexican Spanish	Word-initial /s/ more likely to be reduced ([s] › [h] › [Ø]) in words often preceded by word-final non-high vowels	Raymond & Brown (2012)
English	Words typically predictable reduce in duration more	Seyfarth (2014)

A table formatted this way will look OK

html 'entities' taken from here: https://onlineutf8tools.com/convert-utf8-to-html-entities

Predictability and Informativity4 / 35

Predictability

How likely is the current word given the word the speaker said right before it?

Transitional/conditional probability (cf. Jurafsky et al. 2001): $$P(w_i|w_{i-1})=\frac{C(w_{i-1}w_i)}{C(w_{i-1})}$$
One problem: bigrams with 0 occurrences in the corpus
A solution: add 1 to each bigram count
A better solution: Modified Kneser-Ney smoothing (cf. Chen & Goodman 1998) (r-cmscu R package)

5 / 35

High vs. low predictability

nice home

$$ $$

Lower (< 0.001) predictability of home given nice

fortress-like home

$$ $$

Higher (0.412) predictability of home given fortress-like

6 / 35

Higher predictability → More reduction

nice home

$$ $$

Lower predictability of home given nice

→ Less reduction in home

fortress-like home

$$ $$

Higher predictability of home given fortress-like

→ More reduction in home

7 / 35

Higher predictability → More reduction

Right-context predictability: How likely is this word, given the word the speaker is about to say next?

home wrecker

Higher predictability of home given wrecker

→ More reduction in home

home course

Lower predictability home given course

→ Less reduction in home

8 / 35

Two different predictability "profiles"

9 / 35

Informativity :: How unpredictable from its context is this word on average?10 / 35

Calculating Informativity

$$P(W = w | C = c_i)$$

11 / 35

Take Kneser-Ney smoothed probability of a particular word token in a particular context

Calculating Informativity

$$logP(W = w | C = c_i)$$

12 / 35

Seyfarth used bans (base 10 logarithm) as is apparently usual, I used natural logarithm (R's default) in nsp, corrected it to bans for gpsc

Calculating Informativity

$$\sum_{i=1}^NlogP(W = w | C = c_i)$$

13 / 35

Calculating Informativity

$$\frac{1}{N}\sum_{i=1}^NlogP(W = w | C = c_i)$$

14 / 35

Calculating Informativity

$$-\frac{1}{N}\sum_{i=1}^NlogP(W = w | C = c_i)$$

Steven T. Piantadosi, Harry Tily, & Edward Gibson 2011. "Word lengths are optimized for efficient communicationTitle", Proceedings of the National Academy of Sciences 108(9), 3526-3529.

15 / 35

Informativity vs. Frequency

16 / 35

Predictability vs. Informativity

Word frequently occurs in low-predictability contexts → high informativity

17 / 35

Predictability vs. Informativity

Word frequently occurs in low-predictability contexts → high informativity

Word frequently occurs in high-predictability contexts → low informativity

17 / 35

Target study: Seyfarth (2014)

Durational reduction in Buckeye (Pitt et al. 2007) and Switchboard-1 Release 2 (Calhoun et al. 2009; Godfrey & Holliman 1997)
Predictability and informativity estimateted from Fisher Part 2 corpus (Cieri et al. 2005)
Findings:
- Higher left-context and right-context predictability → More reduction
- Higher right-context (both corpora) or left-context (Switchboard) informativity → Less reduction
Implications:
- Predictability and reduction: could be an online effect
- Informativity and reduction: suggests storage of reduced forms

18 / 35

Research questionsRQ1: Do the results of Seyfarth (2014) replicate on another English dataset?19 / 35

Research questionsRQ1: Do the results of Seyfarth (2014) replicate on another English dataset?RQ2: Do these effects generalize to Polish?19 / 35

Method20 / 35

Source of English data

Nationwide Speech Project Corpus

(Clopper & Pisoni 2006)

21 / 35

Model architecture (N = 7,158)Response: Word duration

22 / 35

Model architecture (N = 7,158)Response: Word duration

Predictors of theoretical interest: Predictability given previous, Informativity given previous, Predictability given following, Informativity given following

22 / 35

Model architecture (N = 7,158)Response: Word duration

Predictors of theoretical interest: Predictability given previous, Informativity given previous, Predictability given following, Informativity given following

"Control" predictors:Part of speech, Orthographic length, No. of syllables, Dialect, Average speaking rate, Rate deviation

22 / 35

Model architecture (N = 7,158)Response: Word duration

Predictors of theoretical interest: Predictability given previous, Informativity given previous, Predictability given following, Informativity given following

"Control" predictors:Part of speech, Orthographic length, No. of syllables, Dialect, Average speaking rate, Rate deviation

Random terms:(1|Word), (1 + Informativity given following + Informativity given previous | Speaker)

22 / 35

Results23 / 35

Predictability given following

β = -0.029, p < 0.001

24 / 35

Informativity given following

β = 0.025, p < 0.001

25 / 35

Source of Polish data (N = 10,588)

Greater Poland Spoken Corpus

wa.amu.edu.pl/korpuswlkp

(Kaźmierski, Kul & Zydorowicz in press)

26 / 35

Predictability given previous

β = -0.007, p < 0.001

27 / 35

Predictability given following

β = -0.016, p < 0.001

28 / 35

Informativity given following

β = 0.02, p < 0.001

29 / 35

Summary of resultsRQ1: Do the results of Seyfarth (2014) replicate on another English dataset?✓ YesDifferences: predictability given previous; informativity given previous (Switchboard)30 / 35

Summary of resultsRQ2: Do these effects generalize to Polish?✓ Yeslocal predictability given both previous and following word, as well as informativity given following word are significant predictors of word duration in Polish words31 / 35

Conclusions→ The effect of local predictability is stronger for right-hand context than left-hand context→ On top of local predictability, both English and Polish show the effect of informativity→ The latter effect suggests phonological storage of reduced forms32 / 35

Thank you!

Right-context informativity influences word durations in English and in Polish

kamil.kazmierski@wa.amu.edu.pl

33 / 35

Model summary: English

Show entries

Search:

	Term	Estimate	Std. Error	t value	Pr(>\|t\|)
1	(Intercept)	0.256	0.015	17.131	0
2	local_pred_given_prev_sd	-0.001	0.001	-0.952	0.341
3	inf_given_prev	0.002	0.005	0.444	0.657
4	local_pred_given_foll_sd	-0.029	0.001	-20.253	0
5	inf_given_foll	0.025	0.005	5.004	0
6	posAdjective	-0.013	0.008	-1.745	0.081
7	posAdverb	-0.036	0.008	-4.683	0
8	posVerb	-0.031	0.006	-5.486	0

Showing 1 to 8 of 17 entries

Previous1 2 3Next

34 / 35

Model summary: Polish

Show entries

Search:

	Term	Estimate	Std. Error	t value	Pr(>\|t\|)
1	(Intercept)	0.251	0.012	20.612	0
2	pred_giv_prev_sd	-0.007	0.001	-5.299	0
3	inf_giv_prev	-0.007	0.004	-1.881	0.06
4	pred_giv_foll_sd	-0.016	0.001	-12.281	0
5	inf_giv_foll	0.02	0.004	5.507	0
6	posadj	-0.001	0.006	-0.112	0.911
7	posadv	-0.001	0.007	-0.132	0.895
8	posverb	-0.02	0.005	-4.349	0

Showing 1 to 8 of 12 entries

Previous1 2Next

35 / 35

Overall frequency effect

Higher lexical frequency → Shorter duration (e.g. Jurafsky et al. 2001)

Perhaps caused by speakers: frequency of use causes articulatory reduction
Perhaps moderated by listeners' expectations: highly frequent forms don't have to be pronounced carefully

Such reduction could result from online computation, performed on abstract phonological representations (cf. Levelt et al. 1999)

↑, ←, Pg Up, k	Go to previous slide
↓, →, Pg Dn, Space, j	Go to next slide
Home	Go to first slide
End	Go to last slide
Number + Return	Go to specific slide
b / m / f	Toggle blackout / mirrored / fullscreen mode
c	Clone slideshow
p	Toggle presenter mode
t	Restart the presentation timer
?, h	Toggle this help

Durational reduction in English (and Polish) words shows a contextual frequency effect

APAP 2019

Kamil Kaźmierski, WA at AMU in Poznań

June 21st, 2019

Overall frequency effect

Contextual frequency effect on variation

Predictability and Informativity

Predictability

How likely is the current word given the word the speaker said right before it?

High vs. low predictability

nice home

fortress-like home

Higher predictability → More reduction

nice home

fortress-like home

Higher predictability → More reduction

home wrecker

home course

Two different predictability "profiles"

Informativity :: How unpredictable from its context is this word on average?

Calculating Informativity

Calculating Informativity

Calculating Informativity

Calculating Informativity

Calculating Informativity

Informativity vs. Frequency

Predictability vs. Informativity

Predictability vs. Informativity

Target study: Seyfarth (2014)

Research questions

RQ1: Do the results of Seyfarth (2014) replicate on another English dataset?

Research questions

RQ1: Do the results of Seyfarth (2014) replicate on another English dataset?

RQ2: Do these effects generalize to Polish?

Method

Source of English data

Model architecture (N = 7,158)

Model architecture (N = 7,158)

Model architecture (N = 7,158)

Model architecture (N = 7,158)

Results

Predictability given following

Informativity given following

Source of Polish data (N = 10,588)

Predictability given previous

Predictability given following

Informativity given following

Summary of results

RQ1: Do the results of Seyfarth (2014) replicate on another English dataset?

✓ Yes

Differences: predictability given previous; informativity given previous (Switchboard)

Summary of results

RQ2: Do these effects generalize to Polish?

✓ Yes

local predictability given both previous and following word, as well as informativity given following word are significant predictors of word duration in Polish words

Conclusions

→ The effect of local predictability is stronger for right-hand context than left-hand context

→ On top of local predictability, both English and Polish show the effect of informativity

→ The latter effect suggests phonological storage of reduced forms

Thank you!

Right-context informativity influences word durations in English and in Polish

Model summary: English

Model summary: Polish

Overall frequency effect

Help