class: center, middle, inverse, title-slide # Durational reduction in English (and Polish) words shows a contextual frequency effect ## APAP 2019 ### Kamil Kaźmierski, WA at AMU in Poznań ### June 21st, 2019 --- layout: true background-image: url(theme/WA_logo_xaringan_small.png) background-position: 2% 98% background-size: 5% --- class: middle, keep-h1-up, font150 # 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) --- # Contextual frequency effect on variation <table class="table table-striped" style="width: auto !important; margin-left: auto; margin-right: auto;"> <thead> <tr> <th style="text-align:left;"> Language </th> <th style="text-align:left;"> Effect </th> <th style="text-align:left;"> Source </th> </tr> </thead> <tbody> <tr> <td style="text-align:left;"> English </td> <td style="text-align:left;"> Prevocalic word-final .ipa[/t/] glottalized more in words typically followed by consonants </td> <td style="text-align:left;"> Eddington & Channer (2010) </td> </tr> <tr> <td style="text-align:left;"> English </td> <td style="text-align:left;"> Word-final .ipa[/t,d/] deletion more likely in words typically followed by consonants </td> <td style="text-align:left;"> Raymond et al. (2016) </td> </tr> <tr> <td style="text-align:left;"> English </td> <td style="text-align:left;"> Unstressed ING more likely to be .ipa[/ɪn/] in words frequently occurring in .ipa[/ɪn/] favoring contexts </td> <td style="text-align:left;"> Forrest (2017) </td> </tr> <tr> <td style="text-align:left;"> Spanish </td> <td style="text-align:left;"> Latin .ipa[/fV-/] words frequently occurring after word-final non-high vowels likely to be <hv-> .ipa[/ØV-/] in MSS </hv-> </td> <td style="text-align:left;"> Brown & Raymond (2012) </td> </tr> <tr> <td style="text-align:left;"> New Mexican Spanish </td> <td style="text-align:left;"> Word-initial .ipa[/s/] more likely to be reduced (.ipa[[s]] › .ipa[[h]] › .ipa[[Ø]]) in words often preceded by word-final non-high vowels </td> <td style="text-align:left;"> Raymond & Brown (2012) </td> </tr> <tr> <td style="text-align:left;"> English </td> <td style="text-align:left;"> Words typically predictable reduce in duration more </td> <td style="text-align:left;"> Seyfarth (2014) </td> </tr> </tbody> </table> ??? A table formatted this way will look OK html 'entities' taken from here: https://onlineutf8tools.com/convert-utf8-to-html-entities --- layout: false class: center, middle, inverse # Predictability and Informativity --- layout: true background-image: url(theme/WA_logo_xaringan_small.png) background-position: 2% 98% background-size: 5% --- # Predictability .center[ ## 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*](https://github.com/jasonkdavis/r-cmscu) R package) --- class: center, middle, font200, keep-h1-up # High vs. low predictability .pull-left[ ## nice home $$ $$ **Lower** (< 0.001) predictability of *home* given *nice* ] .pull-right[ ## fortress-like home $$ $$ **Higher** (0.412) predictability of *home* given *fortress-like* ] --- class: center, middle, font200, keep-h1-up # Higher predictability → More reduction .pull-left[ ##nice home $$ $$ **Lower** predictability of *home* given *nice* → **Less** reduction in *home* ] .pull-right[ ## fortress-like home $$ $$ **Higher** predictability of *home* given *fortress-like* → **More** reduction in *home* ] --- class: center, middle, font200, keep-h1-up # Higher predictability → More reduction **Right**-context predictability: How likely is this word, given the word the speaker is **about to say** next? .pull-left[ ## home wrecker **Higher** predictability of *home* given *wrecker* → **More** reduction in *home* ] .pull-right[ ## home course **Lower** predictability *home* given *course* → **Less** reduction in *home* ] --- # Two different predictability "profiles" <img src="index_files/figure-html/unnamed-chunk-1-1.png" width="90%" height="90%" style="display: block; margin: auto;" /> --- layout: false class: center, middle, inverse # Informativity :: How unpredictable from its context is this word on average? --- layout: true background-image: url(theme/WA_logo_xaringan_small.png) background-position: 2% 98% background-size: 5% --- class: middle, keep-h1-up # Calculating Informativity .font200[ `$$P(W = w | C = c_i)$$` ] ??? Take Kneser-Ney smoothed probability of a particular word token in a particular context --- class: middle, keep-h1-up # Calculating Informativity .font200[ `$$logP(W = w | C = c_i)$$` ] ??? 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 --- class: middle, keep-h1-up # Calculating Informativity .font200[ `$$\sum_{i=1}^NlogP(W = w | C = c_i)$$` ] --- class: middle, keep-h1-up # Calculating Informativity .font200[ `$$\frac{1}{N}\sum_{i=1}^NlogP(W = w | C = c_i)$$` ] --- class: middle, keep-h1-up # Calculating Informativity .font200[ `$$-\frac{1}{N}\sum_{i=1}^NlogP(W = w | C = c_i)$$` ] .footnote[ 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. ] --- # Informativity vs. Frequency <img src="index_files/figure-html/freq_vs_inf-1.png" width="90%" height="90%" style="display: block; margin: auto;" /> --- class: font200, center, middle, keep-h1-up # Predictability vs. Informativity Word frequently occurs in **low-predictability** contexts → **high** informativity -- Word frequently occurs in **high-predictability** contexts → **low** informativity --- class: font140 # 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 --- class: middle, keep-h1-up # Research questions ## **RQ1**: Do the results of Seyfarth (2014) replicate on another English dataset? -- ## **RQ2**: Do these effects generalize to Polish? --- layout: false class: center, middle, inverse # Method --- layout: true background-image: url(theme/WA_logo_xaringan_small.png) background-position: 2% 98% background-size: 5% --- class: center, middle, font180, keep-h1-up # Source of English data Nationwide Speech Project Corpus <img src="media/map_nspc_transparent.png" width="70%" height="90%" style="display: block; margin: auto;" /> [(Clopper & Pisoni 2006)](https://www.ncbi.nlm.nih.gov/pubmed/21423815) --- class: font160, middle, keep-h1-up # 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)` --- layout: false class: center, middle, inverse # Results --- layout: true background-image: url(theme/WA_logo_xaringan_small.png) background-position: 2% 98% background-size: 5% --- # Predictability given following <img src="index_files/figure-html/effect_nspc_pred-1.png" width="90%" height="90%" style="display: block; margin: auto;" /> .footnote[ `β = -0.029, p < 0.001` ] --- # Informativity given following <img src="index_files/figure-html/effect_nspc_inf-1.png" width="90%" height="90%" style="display: block; margin: auto;" /> .footnote[ `β = 0.025, p < 0.001` ] --- class: center, middle, font180, keep-h1-up # Source of Polish data (*N* = 10,588) .pull-left[ .middle[ Greater Poland Spoken Corpus [**wa.amu.edu.pl/korpuswlkp**](http://wa.amu.edu.pl/korpuswlkp) (Kaźmierski, Kul & Zydorowicz in press) ] ] .pull-right[ .middle[ <img src="media/map_combined_xaringan_background.png" width="90%" height="90%" style="display: block; margin: auto;" /> ] ] --- # Predictability given previous <img src="index_files/figure-html/effect_gpsc_pred_prev-1.png" width="90%" height="90%" style="display: block; margin: auto;" /> .footnote[ `β = -0.007, p < 0.001` ] --- # Predictability given following <img src="index_files/figure-html/effect_gpsc_pred_foll-1.png" width="90%" height="90%" style="display: block; margin: auto;" /> .footnote[ `β = -0.016, p < 0.001` ] --- # Informativity given following <img src="index_files/figure-html/effect_gpsc_inf_foll-1.png" width="90%" height="90%" style="display: block; margin: auto;" /> .footnote[ `β = 0.02, p < 0.001` ] --- class: middle, keep-h1-up # Summary of results ## **RQ1**: Do the results of Seyfarth (2014) replicate on another English dataset? ## ✓ Yes ### Differences: predictability given previous; informativity given previous (Switchboard) --- class: middle, keep-h1-up # 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 --- class: middle, keep-h1-up # 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 --- class: center, middle, keep-h1-up # Thank you! ## Right-context informativity influences word durations in English and in Polish <img src="index_files/figure-html/unnamed-chunk-3-1.png" width="90%" height="90%" style="display: block; margin: auto;" /> [kamil.kazmierski@wa.amu.edu.pl](!mailto:kamil.kazmierski@wa.amu.edu.pl) --- # Model summary: English
--- # Model summary: Polish