An exploration of lexical selection in adults who stutter

1.1Avoiding words

Decades ago, Petrunik and Shearing (1983) stated that verbal avoidance represents the PWS attempts to conceal their stuttering. Avoidance behaviour takes various forms in conversational speech and may include word substitution or omission, circumlocution, and simply ‘letting others talk’, as a means of protection from the emotional harm of stuttering and to maintain the identity of a normally fluent speaker (Cream et al., 2003; Guitar, 2019; Lowe et al., 2017; 2021; Plexico et al., 2009). When a person avoids words, s/he either needs to choose a different word or change what is to be said (Lowe et al., 2017). As a result, behavioural reinforcement occurs, giving rise to habitual patterns of speech avoidance in PWS. Although there is wide variation in the form and frequency of avoidance strategies used by PWS (Crichton-Smith, 2002; Mancinelli, 2018; Plexico et al., 2009), data suggest that word substitution is one of the most commonly employed and earliest-appearing strategies (Crichton-Smith, 2002; Jackson et al., 2015), accounting for up to 82%of all coping responses (Vanryckeghem et al., 2004).

Successfully ‘problem-solving’ anticipated stuttering through word substitution and other similar strategies has been reported to result in distraction, miscommunication, restricted opportunities, and reduced capacity to communicate for social purposes (Amster & Klein, 2018; Cream et al., 2003; Jackson et al., 2015; Plexico et al., 2009). Indeed, linguistic data concerning the expression of politeness and opinion by PWS suggest a reduced openness to interpersonal engagement in these speakers, which the researchers interpret as evidence of systemic language change related to verbal avoidance (Lee et al., 2015). Post-treatment interviews with PWS indicate the persistence of avoidance behaviours in some cases, where a cumulative load of typical speech content planning, anticipatory word selection, and the application of newly learned stutter-control techniques then becomes apparent (Cream et al., 2003; Cream et al., 2004; Linklater, 2020).

Verbal avoidance is a significant feature in the clinical presentation of stuttering, with far-reaching implications for the communicative freedom and effectiveness of PWS. To date, the majority of research examining verbal avoidance in PWS is based on personal accounts during structured or semi-structured interviews (Cream et al., 2003, 2004; Crichton-Smith, 2002; Jackson et al., 2015; Plexico et al., 2009) and self-report checklists (Blomgren et al., 2005; Vanryckeghem et al., 2004). That is, the only essential evidence of an instance of spontaneous verbal avoidance resides in the speaker’s own awareness. The aim of the current study was to explore the use of a behavioural measure to examine verbal avoidance. Creating a measurable context for objectively documenting avoidance behaviour could provide additional insight regarding the experience of stuttering. Research related to confrontational picture naming performance in PWS may indicate one possible approach.

1.2Clues from confrontation naming

Lexical access is a highly automated process but not entirely passive, as the speaker exerts some degree of cognitive control over the process. According to Miyake et al. (2000) and others (Levelt et al., 1999; Shao et al., 2013; Shao et al., 2014), naming pictures involves executive control over lexical competitors. That is, word selection is achieved by suppressing the activation of alternative lexical choices. Presumably, the speed at which a picture is named (i.e., latency) is representative of the amount of competition among possible word choices (Folstein & Van Petten, 2008). A picture that could be named using a variety of word choices would result in a longer latency of response.

Interest in the relationship between lexical processing ability and stuttering has given rise to a cluster of studies examining the performance of PWS on tasks requiring the rapid naming of picture stimuli. These have attempted to explain disfluency in relation to models of lexical access, activation, and production, using both behavioural (e.g., Crowe & Kroll, 1991; Hennessey et al., 2008; Newman & Bernstein Ratner, 2007; Pellowski, 2011; Prins et al., 1997; Taylor et al., 1970) and electrophysiological data (e.g., Maxfield, 2020, 2021; Maxfield et al., 2010; Maxfield et al., 2015). Results from confrontation naming tasks in PWS have been equivocal, producing a complex picture of lexical access behaviour in PWS. For example, Maxfield et al. (2015) and Pellowski (2011) reported on groups of PWS and people with no stuttering (PWNS) who completed computerized naming tasks, where the target responses were noun labels for common objects. Both studies found a significantly longer naming latency for PWS participants, which was interpreted to reflect diminished ability to control lexical selection. However, similar studies reporting no significant naming latency differences between PWS and PWNS also exist (Hennessey et al., 2008; Maxfield, 2020, 2021; Newman & Bernstein Ratner, 2007, Prins et al., 1997). Prins et al. found no significant main effect of group status on the speech onset latency of 12 PWS and 12 PWNS, in response to line drawings depicting target nouns or verbs. A comparatively greater difference in response latency was nonetheless seen between groups for target verbs than nouns. Based on this observation and the tendency for disfluency to appear on verb phrases from the time of early childhood (Bloodstein et al., 2021), the authors suggested a particular role of slowed verb processing in precipitating stuttering. In the case of Maxfield (2020, 2021), the latency of naming words depicting common and uncommon objects/actions was measured in PWS and PWNS. Both groups were found to exhibit differential latencies for common (short) and uncommon (long) words. However, the magnitude of the effect was larger for PWS.

There is evidence that, when compared to PWNS on speech latency measures, PWS respond differently to changes in processing demands, such as those related to linguistic complexity and emotionality (Hennessey et al., 2014; Newman & Bernstein Ratner, 2007; Prins et al., 1997; Tsiamtsiouris & Cairns, 2009). While confrontation naming thus far has been used to examine lexical access and activation among PWS, the methodology also allows for the accuracy of naming responses. Errors in naming have typically included any verbal responses other than a pre-determined target word, as well as hesitations, stutters, and non-responses, and have not been shown to distinguish groups of PWS and PWNS (e.g., Hennessey et al., 2008; Maxfield et al., 2010, 2015; Pellowski, 2011). Newman and Bernstein Ratner (2007) employed a more elaborate method of accuracy analysis, with interesting results. The study involved a timed computerized picture-naming task containing noun and verb stimuli controlled for word frequency factors and matched for initial phoneme, which was completed by 25 PWS and 25 PWNS participants. Response errors resulting from word-finding difficulty or visual confusion of the pictured stimulus were tallied, but response variants representing dialectal differences, alternate word choices, or elaborations, were counted separately and classified as correct. This was intended to account for any occurrences of word avoidance. Statistical analysis revealed that the PWS group produced a significantly higher frequency of errors than the control group, but the two groups demonstrated a similar frequency of response variants and elaborations. Notably, the authors also described the use of certain highly unusual word labels by PWS participants to name common picture stimuli. For example, boy was labelled androgyne, and knee was termed patella. Though the incidence of such responses was low, and the overall number of variant responses was not significantly different between groups, the authors postulated that such lexically creative word choices might reflect word substitution for stuttering avoidance. In the present study, we wished to explore the possibility of capturing instances of stuttering avoidance on the basis of word choice.

Newman and Bernstein Ratner’s (2007) findings invite the closer evaluation of confrontation naming response variants, which, if qualified in relation to a criterion, could provide insights into word selection behaviour (rather than simply lexical processing ability) in PWS. These data could be combined with naming latency measurements to examine the time-cost associated with problem-solving stuttering “in the moment.” Admittedly, the confrontation naming paradigm represents a contrived speaking context, whereas verbal avoidance is a feature of spontaneous conversation. Yet, the particular salience of word substitution as an avoidance strategy among PWS provides certain grounds for taking this approach, with its word-production focus, as an exploratory point of departure.

1.3The present study: picture naming task

Evidence from self-report sources demonstrates that verbal avoidance, of which word substitution is a key strategy, can significantly impact the quality of communication experienced by PWS (Cream et al., 2003, 2004; Plexico et al., 2009; Tichenor & Yaruss, 2019). Research suggests that confrontation naming data may provide a means for objectively commenting on the mechanisms of word substitution behaviour, including its frequency, variants, time efficiency, and relationship to speaker characteristics (Newman & Bernstein Ratner, 2007). The present study used a specifically designed picture-naming tool for investigation of verbal avoidance.

The Picture Naming Task (PNT; O’Beirne, 2016) was developed using pictures from Snodgrass and Vanderwart’s (1980) standardized 260-item set. This picture set consists of black-and-white line drawings of common everyday objects, normed for name and image agreement, familiarity, and visual complexity. The picture set was originally intended for research examining picture and word processing, with the normative information providing a means for the experimental control of different name and image characteristics. Updated data on name agreement in English have also been published for Snodgrass and Vanderwart’s pictures (Yoon et al., 2004). The normative data were obtained from 113 healthy American English-speaking adults, who were asked to provide “the first name of the object that comes to mind” for each picture item (p. 641). Using the collected responses, the authors documented the ‘dominant’ or most common noun provided by participants for each picture, as well as all response variants or alternative names, listed in order of their relative frequencies.

Several studies investigating confrontation naming performance in stuttering speakers have used a subset of Snodgrass and Vanderwart’s (1980) pictures, measuring response accuracy against dominant names from the name agreement database (e.g., Anderson, 2008; Maxfield, 2020, 2021; Pellowski, 2011; Pellowski & Conture, 2005). In the work of Pellowski and Conture, naming responses deviating in any way from the picture’s dominant name were considered as errors and excluded from speech reaction latency analyses. Maxfield (2020, 2021), appeared to follow a similar methodology and disregarded alternative word choices (e.g., word substitutions). On the other hand, Anderson accepted as accurate both dominant picture names and close synonyms (e.g., bunny for rabbit), in a study examining semantic and phonological priming effects on picture naming in children with stuttering. It is unclear to what extent Anderson’s synonym acceptability judgements were based on published name agreement norms. No prior research has involved the closer inspection of specific word choices associated with the Snodgrass and Vanderwart (1980) data, including both dominant picture names and word alternatives, for the construction of a scaled picture-naming accuracy criterion. This was crucial to development of the PNT tool used in the current study.

A subset of 64 pictures from Snodgrass and Vanderwart’s (1980) original set were selected for PNT inclusion, based on (1) picture name variability (Yoon et al., 2004), (2) low visual ambiguity, and (3) low likelihood of naming differences among English dialects, as judged by the researchers. Dominant names for included items contained a range of initial phonemes and clusters, but word frequency and phonological factors were not further controlled due to the exploratory nature of the study. An alphabetized list of dominant and most frequent alternative picture names for the selected items is shown in the Appendix (less frequent alternative names were also considered in PNT scoring, but are not listed). The system of documentation employed by Yoon et al. was thus established as an evidence-based accuracy criterion against which new naming responses could be compared, to determine the usuality (or non-usuality) of different word choices. Assuming PWS will anticipate stuttering during a single-word confrontation task, we hypothesized that stuttering anticipation might be reflected by either unusual word choice and/or delayed word production. Accordingly, the following questions were posed:

2.1Participants

Participants were nine adult PWS (3 males, 6 females) and nine PWNS matched for sex and age (±5 years). All participants were native speakers of English, with no reported or observed neurological or speech-language impairment (with the exception of stuttering for PWS). Stuttering was self-reported by each PWS participant, and confirmed by speech-language pathologist observation and assessment using the Stuttering Severity Instrument –Fourth Edition (SSI-4; Riley & Bakker, 2009). All participants provided written consent for their involvement in the study, which was approved by the University of Canterbury Human Ethics Committee (approval number HEC2016/11). Participant characteristics are summarized in Table 1.

2.2Procedure

All procedures and analyses were completed by the first author. Each participant individually completed the PNT in a quiet room. Prior to beginning the task, participants were instructed to “name each picture as fast as you can, saying just the name for the picture”. PWS participants were additionally encouraged to “use your natural speaking voice, not a speech technique you may have learned in therapy”. This was to elicit habitual patterns of speech and language rather than learned behaviours from treatment.

Pictures were presented individually and in random order in the centre of a computer screen. Each picture measured approximately 500×500 pixels. The appearance of each picture coincided with a simultaneous audible acoustic tone, following which the participant was required to respond. Screen shots illustrating examples of the picture presentation procedure are shown in Fig. 1. Movement from one picture to the next was manually controlled by pressing a key on the computer keyboard. The task began with 10 practice items, consisting of pictures from Snodgrass and Vanderwart’s (1980) set that were not included in the actual test. The purpose of the practice period was to familiarize participants with the procedure and to provide feedback regarding performance expectations. Participants were informed when the practical trials had ended and the task was about to begin.

Fig. 1

Screenshot examples from the Picture Naming Task (O’Beirne, 2016). Picture stimuli are from Snodgrass and Vanderwart (1980). The dominant names for these pictures are fly (nondominant name is mosquito) and suitcase (nondominant name is briefcase).

Responses to the 64 test items were automatically audio-recorded by the software using the computer’s built-in microphone, and saved as.wav audio files. These were later imported into Praat 6.0 (Boersma & Weenink, 2015) and displayed as amplitude-by-time waveforms. Vertical cursors were superimposed on each waveform and positioned at the onset of the audible acoustic tone (occurring simultaneous to picture appearance) and the onset of the verbal acoustic signal. The time period between the cursors was taken as the verbal response latency. Responses that were preceded by filled pauses, stuttering moments, comments, or other inappropriate vocalizations, and non-responses, were marked as errors and excluded from response latency analysis. Verbal response latencies that fell more than two standard deviations above or below the mean response latency for each participant were marked as outliers and also excluded (Hennessey et al., 2008). Intra-rater measurement reliability was determined for 10%of the total audio data. The Pearson product-moment correlation coefficient between the original and repeat measurements for verbal response latency was 0.98, and the mean absolute difference between the two sets of measures was 0.03 s, indicating a high level of reliability.

All naming responses (including those responses that were judged to have verbal latency errors & outliers), were analysed for accuracy by comparison to the normative data listed by Yoon et al. (2004). Responses were scored ‘match’ if they corresponded to the dominant name for a stimulus item, ‘other’ if they corresponded to any alternative (nondominant) name listed for that item, and ‘none’ if they represented a word label not listed for that item within the normative database.

2.3Statistical analysis

Two-tailed paired t-tests were used for the overall comparison of performance means between the PWS and PWNS groups on error frequency, response latency, and response accuracy for the PNT. Regression analysis was also conducted to model the effects of different variables on response latency and response accuracy, with p-values obtained through likelihood ratio testing.

3.1Group differences

3.2Regression modelling for response latency

R version 3.2.2 (R Core Team, 2015) and the lme4 package (Bates et al., 2015) were used to construct a linear mixed effects model for PNT response latency, in order to identify the relative effects of response accuracy and different participant variables on response latency (Field et al., 2012). Response latency values were first log-transformed to obtain a normal distribution. The mixed effects model was then constructed using a backward stepwise method. The final model contained response accuracy (Accuracy), participant age (Age), and PWS or PWNS group status (Group) as fixed effects, with an Age x Group interaction term. Intercepts for Item and Participant were entered as random effects, to account for variation related to a particular stimulus item or individual. Visual inspection of residual plots did not reveal any obvious deviations from homoscedasticity or normality. P-values for each effect were obtained by likelihood ratio testing of the full model against the model without the effect in question.

Coefficients (with confidence intervals), standard errors, and p-values for the fitted model are displayed in Table 2. There was a significant main effect of Accuracy, whereby other responses (M = 1.15 s, SD = 0.48), F(2,8) = 11.31, b = 0.06, t(8) = 2.917, p < 0.001, and none responses (M = 1.35 s, SD = 0.43), F(2,8) = 11.31, b = 0.14, t(8) = 4.255, p < 0.001, were both associated with increased response latency, as compared to match (M = 1.07 s, SD = 0.42) responses. The effect of Accuracy is depicted in Fig. 2.

Fig. 2

Boxplot of Picture Naming Task response latency, with box width representing the number of responses scored as match (n = 782), other (n = 257), and none (n = 113), and mean values indicated within the boxes.

There was no main effect of Age or Group, indicating that response latency was similar for both the PWS (M = 1.15 s, SD = 0.17) and PWNS groups (M = 1.21 s, SD = 0.30). However, there was a small but significant Age x Group interaction, F(1,8) = 8.06, b = 0.01, t(8) = 2.839, p = 0.010, suggesting that increased age was associated with increased response latency for the AWS group only.

3.3Regression modelling for response accuracy

R version 3.2.2 (R Core Team, 2015) and the mlogit package (Croissant, 2013) were used to perform a multinomial logistic regression analysis for PNT response accuracy. A multinomial logistic approach was taken because accuracy was measured as a nominal variable having three possible outcomes (i.e., match, other, none) (Field et al., 2012). The model was constructed using a forward stepwise method. The final model contained only Item and Error (i.e., the production of filled pauses, comments, other inappropriate vocalizations, or moments of stuttering prior to the intended verbal response, and non-responses) as predictors for response accuracy (R² = 0.35). P-values for the predictors were obtained by likelihood ratio testing of the model in question against the baseline model.

Instances of error/outlier verbal latency production were found to be significantly associated with increased likelihood of none responses, F(1,8) = 16.00, b = 1.31, t(8) = 4.00, p < 0.001, OR = 3.71, 95%CI. These were instances where the word label provided by the participant was not listed for a given item within Yoon et al.’s (2004) normative database. No similar relationship was observed for other responses, F(1,8) = 0.21, b = 0.13, t(8) = 0.46, p = 0.647. A tendency towards increased odds of none responses was also observed for certain items on the PNT, but the effect of Item did not reach statistical significance.

PNT response accuracy was then modelled separately for the PWS group using the same method, with Item and stuttering severity as predictors in the final model (R² = 0.43). More severe stuttering, as measured by SSI-4 total score (Riley & Bakker, 2009), was found to be significantly associated with slightly reduced odds of both none responses, F(1,8) = 4.67, b = –0.04, t(8) = –2.16, p = 0.031, OR = 0.96, 95%CI [0.93, 1.00], and other responses, F(1,8) = 11.49, b = –0.04, t(8) = –3.39, p < 0.001, OR = 0.96, 95%CI [0.93, 0.98]. Again, the effect of Item did not reach statistical significance.

4.1Future directions and conclusion

We emphasize that the present study was exploratory in nature and design. In addition to study replication with a larger sample size, further development of the PNT would improve its sensitivity to target behaviours and its utility in this area of investigation. There is a clear role of linguistic complexity in moderating fine aspects of picture-naming performance in PWS (Hennessey et al., 2008, 2014; Newman & Bernstein Ratner, 2007; Tsiamtsiouris & Cairns, 2009). Current PNT stimuli consisted of pictured nouns only, and the word frequency, familiarity, and phonological features of the items selected from Snodgrass and Vanderwart (1980) were not carefully controlled. Expansion of the stimulus set to include verbs, which represent a less simple word class than nouns (Bloodstein et al., 2021; Maxfield, 2021), would introduce an added complexity element. This would reveal subtler differences in word choice and naming behaviour between PWS and PWNS (Prins et al., 1997). Similarly, increasing experimental control of stimulus characteristics using the wider range of normative data published by Snodgrass and Vanderwart would minimize any confounding item effects. Item effects in the current study, though not statistically significant, explained a certain proportion of the variation in response accuracy and latency. While this variation was to some extent accounted for through the use of mixed effects modelling, further standardization of task stimuli would enable stuttering-related word choices to be discerned with greater clarity.

To conclude, use of the PNT has provided evidence of (1) a time-cost associated with word substitution, (2) the effects of processing demands on speech-language performance in PWS, and (3) the independence of stuttering impact from severity of actual speech disfluency. These results confirm the need for a greater understanding of verbal avoidance behaviour in AWS, based on objective sources of evidence. The PNT paradigm, with its focus on word choice in the context of timed responding, provides an original methodological model for exploring the mechanisms and demands of word substitution in these speakers. The current version of the task did not, however, permit the specific evaluation of word substitution behaviour in PWS, in terms of response variant and latency characteristics within a confrontation naming context. Continued research and task development will enlighten aspects of verbal avoidance central to the effective management of stuttering impact.