Role of cognitive control in resolving two types of conflict during spoken word production

ABSTRACT A theoretically- and clinically-important issue for understanding word retrieval is how speakers resolve conflict during linguistic tasks. This study investigated two types of conflict resolution: prepotent conflict, when one dominant incorrect response must be suppressed; and underdetermined conflict, when multiple reasonable responses compete. The congruency sequence effect paradigm was used to assess trial-to-trial changes in reaction time and accuracy during word production tasks with either prepotent or underdetermined conflict. Pictures were named faster on trials with low-conflict as compared to high-conflict regardless of conflict type. This effect was modulated by the amount of conflict experienced on the previous trial for both tasks. These results suggest that resolution of underdetermined and prepotent conflict may engage the same general cognitive mechanism. This work expands our understanding of the relationship between cognitive control and word production and can inform clinical approaches for people with anomia.


Introduction
Word retrieval difficulty (anomia) is one of the most common deficits people with aphasia experience and remains a persistent problem for many despite receiving traditional speech-language therapy.One proposed source of word production errors, both in aphasia and in unimpaired speakers, is difficulty inhibiting conflicting words that are activated during retrieval.Theoretical accounts of spoken word production suggest speakers select among related words that are simultaneously activated during word retrieval (e.g.Dell, 1986;Levelt et al., 1999).Lexical errors arise when a non-target word is mistakenly selected for production.There is a growing consensus that cognitive control is needed to detect and resolve interference among competing words to facilitate selection of the correct lexical representation (Crowther & Martin, 2014;Hsu & Novick, 2016;Kan et al., 2013;Nozari & Novick, 2017;Schnur et al., 2009;Snyder et al., 2011;Ye & Zhou, 2009b).
Cognitive control, the set of mental processes that detect and resolve conflict, plays an important role in the resolution of two types of conflictprepotent and underdetermined.Prepotent conflict occurs when an individual must override one dominant but contextually inappropriate response (Botvinick et al., 2001;Hussey & Novick, 2012).Specifically, this conflict arises when there is an incompatibility between the current demands of the situation and how the stimulus is typically processed.One classic example of prepotent conflict is the Stroop task, in which the reading response generates a prepotent bias to produce the written colour word and must be controlled in order to process the task-relevant perceptual representation of the text colour (Hussey & Novick, 2012).In word production, prepotent conflict can occur when speakers must override a strongly activated word that is inappropriate for the current context (e.g.referring to a spouse as "girlfriend" or "boyfriend" after marriage; Anderson & Levy, 2007).
In contrast, underdetermined conflict occurs when multiple reasonable and appropriate responses compete and must be suppressed to select the desired target (Botvinick et al., 2001).This occurs frequently during language production, as there are often many ways to express the same intended message.Speakers must select from multiple appropriate and plausible words which could be used (e.g.couch vs. sofa vs. loveseat).When there is a need to detect and resolve conflict, regardless of conflict type, information processing is slower and less accurate as compared to processing information without conflict (Botvinick et al., 2001;Snyder et al., 2014).
In this paper, we investigate the extent to which similar cognitive control mechanisms are engaged in language production tasks involving prepotent and underdetermined conflict.This includes a replication of previous work which demonstrated that cognitive control is engaged within a word production task that involves prepotent conflict (i.e.picture-word interference; Duthoo et al., 2014;Freund & Nozari, 2018;Shitova et al., 2017), as well as a novel study exploring cognitive control in a word production task with underdetermined conflict using the congruency sequence effect paradigm.To motivate this work, the next section provides an overview of the evidence that cognitive control plays a role in resolving different types of conflict.This is followed by a review of methods used to measure cognitive control and a discussion of how these methods will be used in word production experiments to address the research questions of the current study.

Conflict resolution during lexical retrieval
There is widespread agreement across theories of spoken word production that conflict arises due to multiple representations that are activated simultaneously at each processing level and that speed and accuracy of word retrieval are influenced by co-activated representations (Belke & Stielow, 2013;Breining et al., 2016;Howard et al., 2006;Nozari et al., 2016;Schriefers et al., 1990).However, word production models differ in their predictions of how and when competition occurs during the selection process (see Dell et al., 2014;Spalek et al., 2013 for reviews).While most models predict that competition occurs at the lexical level, it has been suggested that competition may also occur at a later post-lexical stage (i.e.response exclusion hypothesis; Mahon et al., 2007).While we acknowledge that competition may occur at different points during word production, in this study we focus on competition that occurs at the lexical level.
Competition during lexical selection can explain the semantic interference effect found in a prepotent conflict task such as a picture-word interference (PWI) task (e.g.Glaser & Dungelhoff, 1984;Roelofs, 1992;Rosinski, 1977;Schriefers et al., 1990).In this task, speakers are asked to name pictures while ignoring a written or spoken distractor word.Picture naming latencies are longer when the distractor word is semantically-related to the target word (e.g.picture of a cat with "dog" written on it) as compared to when the distractor word matches the picture (e.g.picture of a cat with "cat" written on it) or is unrelated (e.g.picture of a cat with "pen" written on it).Naming speed is slower in the semantic condition because the target picture and the distractor word activate competing lexical representations (Levelt et al., 1999;Schriefers et al., 1990;cf. Dell'Acqua et al., 2007).This task is an example of prepotent competition during language production because the speaker must override one strongly competing lexical item to accurately name the picture.
Findings from tasks that evoke underdetermined conflict also support the role of competition during lexical retrieval.In these tasks, there are multiple reasonable and appropriate word responses that compete for selection.Naming latencies are longer when the number of possible responses increases.For example, speakers are slower to name pictures that have multiple names as compared to pictures associated with only one name (Kan & Thompson-Schill, 2004;Madden et al., 2019).Speakers are also slower to complete sentences that have multiple possible endings as compared to sentences that are highly constrained (Barker et al., 2022;Nathaniel-James, 2002) and slower to produce a verb related to a written noun when that noun could be associated with many verbs (e.g.ballhit, kick, throw, etc.) as compared to a noun associated with one primary verb (e.g.scissorscut; Snyder & Munakata, 2008;Thompson-Schill et al., 1998).
Converging evidence from behavioural and neuropsychological studies suggests that cognitive control enables speakers to resolve interference during lexical selection which results in fluent speech with relatively few errors (e.g.Crowther & Martin, 2014;Hsu & Novick, 2016;Kan et al., 2013;Schnur et al., 2009;Snyder et al., 2011;Ye & Zhou, 2009a).The role of cognitive control has been explored in word production tasks that involve prepotent conflict (e.g.Stroop in Crowther & Martin, 2014;picture-word interference in Shao et al., 2013) since these paradigms reliability elicit competition and are relatively simple to manipulate control demands.However, it is also crucial to gain an understanding of the role of cognitive control in resolving underdetermined conflict.Tasks that invoke underdetermined conflict in the linguistic domain may have greater ecologically validity as they more closely resemble conflict that arises when speakers select from many alterative (and competing) words when communicating their thoughts during spontaneous speech (Kurland et al., 2014).
Neuroimaging studies suggest that prepotent and underdetermined conflict may be resolved using different cognitive mechanisms and neural regions.For example, Snyder et al. (2014) found that overlapping but partially independent areas within the prefrontal cortex were engaged when resolving different types of conflict.The ventrolateral prefrontal cortex was active during resolution of underdetermined conflict while the dorsolateral prefrontal was cortex active during resolution of both underdetermined and prepotent conflict.The authors argue that competitive lateral inhibition within the left ventrolateral prefrontal cortex (VLPFC) facilitates the resolution of underdetermined conflict while the dorsolateral prefrontal cortex (DLPFC) is the primary source that controls the resolution of prepotent conflict and provides top-down support for task-relevant representation when competition is high.There is also evidence from lesions studies that support the role of VLPFC in the resolution of underdetermined conflict.Robinson et al. (1998) and Robinson et al. (2005) report detailed investigations of two individuals with damage and/or atrophy to the left VLPFC who present with selective impairments in tasks that involve resolution of underdetermined conflict.For example, during a task where individuals were instructed to generate a sentence from a single noun, performance was better on when there was a prepotent, dominant response (e.g, proper nounsgenerate a sentence about "Mona Lisa") as compared to generating a sentence when there was multiple, competing responses (e.g. common nounsgenerate a sentence about "table ").
Additionally, different cognitive control processes may be involved at different timepoints during the resolution of prepotent and underdetermined conflict.The dual-mechanism control (DMC) framework suggests that proactive control is activated before the conflict arises while reactive control is engaged after a stimulus is processed and conflict is detected (Braver, 2012).Prepotent conflict may be resolved, at least in part, by proactive control when the task-relevant dimension is constant across trials (e.g.name the picture, ignore the word in PWI).Speakers can anticipate prepotent conflict and activate control processes in advance which help to sustain the goal representation until the conflict is resolved (Kalanthroff et al., 2018;Spinelli & Lupker, 2021;c.f. Yang & Pourtois, 2022).This process is facilitated by the DLPFC which can hold the task set in working memory and bias attention towards task-relevant information (Banich, 2019).In contrast, undetermined conflict can be resolved only in a reactive fashion, after the stimulus is present and alternative responses are already competing for selection.
Taken together, there is evidence that both types of conflict occur in language production as they do in non-linguistic cognitive processing.However, it remains unclear whether conflict resolution is achieved in a similar way.In this study, we evaluated whether prepotent and underdetermined conflict during word production are resolved using similar cognitive control mechanisms based on behavioural data collected using the congruency sequence effect (CSE) paradigm.This paradigm, originally borrowed from the psychology literature but more recently applied to study cognitive control in language processing, is explained in the next section.

Measuring adaptive changes to examine role of cognitive control in lexical selection
Although neuroimaging, neuropsychological, and correlational studies demonstrate that cognitive control is involved during language production, these methods provide only indirect evidence of the need for control during lexical selection for production (Freund & Nozari, 2018;Hussey et al., 2017).There has been a recent shift toward methods that demonstrate a causal link between cognitive control and language processing.This has included assessing changes in language abilities after behavioural training targeting cognitive control with or without neuromodulation (e.g.Hussey et al., 2015;Hussey et al., 2017) and analysing online adaptive changes during language processing (e.g.Freund & Nozari, 2018;Shitova et al., 2017).
Paradigms that measure trial-by-trial effects are the current "gold standard" for analysing whether a task initiates engagement of the control system to regulate online performance in conflict situations (Freund & Nozari, 2018).The difference in reaction time between trials without conflict (i.e.task-relevant and irrelevant stimulus features match; congruent) and trials with conflict (i.e.mismatch between task-relevant and irrelevant stimulus features; incongruent) is often used as an index of adaptive control (Egner, 2017).A smaller difference (i.e.congruency effect) is associated with stronger implementation of control while a larger difference indicates that there was a greater processing cost (i.e. more time was needed to detect and resolve conflict).Gratton et al. (1992) were the first to demonstrate that the congruency effect is modulated by the amount of conflict in the preceding trial.Performance on a trial with conflict is faster and more accurate if it preceded by another conflict trial, as compared to a non-conflict trial.Thus, the congruency effect is smaller following high-conflict trials than on trials following no (or low) conflict trials (Figure 1).This finding, referred to as the congruency sequence effect (CSE), has been replicated across many conflict tasks requiring manual responses (e.g.Botvinick et al., 2001;Gratton et al., 1992;Kerns et al., 2004;Notebaert et al., 2006;Stürmer et al., 2002) and verbal responses (Blais et al., 2014;Duthoo et al., 2014;Lamers & Roelofs, 2011;Shitova et al., 2017;van Maanen & van Rijn, 2010).More recently, the canonical CSE pattern has been replicated in several studies using a linguistic task with prepotent conflict (e.g.picture-word interference; Duthoo et al., 2014;Freund & Nozari, 2018;Shitova et al., 2017;Spinelli et al., 2019;van Maanen & van Rijn, 2010).This supports the notion that during word production, speakers engage cognitive control processes to monitor their productions and regulate their behaviour online in an adaptive manner (Freund & Nozari, 2018).

The current study
The current study examined the extent to which unimpaired speakers implement cognitive control processes to resolve different types of conflict during word production.Because dynamic adjustments in response to conflict (i.e.faster, more accurate performance following conflict detection) are thought to reflect implementation of cognitive control, a CSE paradigm was used to assess online trial-to-trial changes in reaction time and accuracy during performance of a word production task that required resolution of either prepotent or underdetermined conflict.
The first aim of this study was to replicate the CSE in a prepotent language task (e.g.picture-word interference) to provide additional support for the engagement of cognitive control processes to resolve prepotent conflict during word production (Duthoo et al., 2014;Freund & Nozari, 2018;Shitova et al., 2017;Spinelli et al., 2019;van Maanen & van Rijn, 2010).We hypothesised that unimpaired speakers would monitor their productions and regulate their behaviour online based upon information generated when prepotent conflict arises within the language production system.Thus, there would be a stronger implementation of cognitive control (i.e.smaller congruency effect) following trials with highconflict than on those following low-conflict trials, replicating the canonical CSE pattern (see Figure 1).
We also explored whether unimpaired speakers engage similar cognitive control processes to resolve underdetermined conflict during word production using the CSE paradigm.Since information processing is slower and less accurate as compared to processing information without conflict, regardless of conflict type, (Botvinick et al., 2001), it is possible that participants could engage cognitive control to resolve conflict in a CSE paradigm including production tasks with underdetermined conflict, in a similar fashion as tasks with prepotent conflict.That is, stronger implementation of cognitive control would be found following trials with conflict (i.e.canonical CSE pattern).However, different neural regions may be used to engage cognitive control (Cipolotti et al., 2016;Robinson et al., 1998;Robinson et al., 2005;Snyder et al., 2014) and different types of control (Braver, 2012) may be engaged when resolving underdetermined conflict.Therefore, the canonical CSE pattern may not be found if cognitive control is implemented differently in response to underdetermined conflict within a word production task.

Method
Participants Sixty participants recruited from the New York University campus were included in this study.Six participants were excluded from analysis (three due to computer error; one due to low task accuracy; two reported history of neurological problems), leaving 54 participants in the analysis (39 female; mean age = 23 years, 4 months; range = 18;4-35;10 years; months).Twentysix participants were included in the PWI analysis and 28 participants were included in the Name Agreement analysis.Participants were randomly assigned to the PWI task or Name Agreement task.
All participants reported they were native speakers of American English.Participants reported that English was the first language they learned (i.e.no simultaneous bilingual speakers were included) and was their primary language used for daily communication.Additionally, all included participants reported no history of neurological, speech, or language problems and/or previous speech-language therapy on a written questionnaire.Participants were tested individually in a quiet room and received $15 upon completion of the study.All tasks were completed in one session lasting approximately an hour.This study received approval from the Institutional Review Board at New York University (FY-IRB2018-1192).Informed consent was obtained from all participants at the beginning of the study.Participants completed a written questionnaire about their language background before performing experimental tasks.

Picture word interference (PWI)
In this task, speakers were asked to name pictures while ignoring an overlapping written distractor word.In the low-conflict (i.e.congruent) condition, the distractor word matched the picture (e.g.picture of a cat with "cat" written on it).In the high-conflict (i.e.incongruent condition), the distractor word was semantically related to the target word (e.g.picture of a cat with "dog" written on it).
To create the stimuli for the PWI task, a list of 120 target-distractor word pairs was compiled (Appendix A).The targets were evenly distributed across 15 categories.All distractors are semantically related, matched for length (within three letters) and frequency (using SUBTLEX; Brysbaert & New, 2009) and had minimal phonological overlap with the target.Seven additional target-distractor word pairs were created to use for practice and as filler trials for use at the first trial of each block.
Black-and-white line drawing pictures corresponding to the target items were selected from the International Picture Naming Project (IPNP; Szekely et al., 2004) and open source internet sources.All pictures were 300-by-300 pixels and presented in the center of the computer screen.The word (e.g.distractor for high-conflict trials or target for low-conflict trials) was presented simultaneously in the center of the image in red, uppercase 20-point font to create 120 high-conflict and 120 lowconflict stimuli for this task.

Name agreement
Speakers were asked to name pictures that either have high name agreement or low name agreement.To match the terminology used for other task, pictures with high name agreement are described as lowconflict trials and pictures with low name agreement are described as high-conflict trials.
Initially, a total of 120 pictures were selected from the IPNP database (Appendix B) and divided equally into two sets based on the reported name agreement (Szekely et al., 2003;Szekely et al., 2004).All pictures in the high name agreement group had 100% name agreement.Mean name agreement for low name agreement pictures was 54.8% (range: 28-69%).High and low name agreement sets include the same number of pictures from each lexical category (as defined by IPNP).The two sets of pictures were matched for frequency using SUBTLEX (median value: 2.31 for the highconflict items and 2.45 for the low-conflict items; t (114) = 1.16, p = .25).SUBTLEX frequency values were not available for four low-agreement items.All pictures were 300-by-300 pixels and presented in the center of the computer screen with a white background.
Despite this initial categorisation, following data collection analysis of responses from the participants in the current study revealed that several items did not match the original categorisation criteria.For example, the picture labeled "desk" was originally categorised as a low-conflict trial with 100% name agreement based on data from IPNP.However, in the current study, six labels were associated with this item (e.g.drawer, cabinet, chest) and only 82% of the participants identified the picture as "desk".Therefore, prior to analysis, items were reorganised using a new set of criteria to divide the pictures into high/low name agreement groups based on responses from the participants in this study.
To be included in the high name agreement group (i.e.low-conflict trials), pictures had no more than one response associated with them from the participants in the current study.Pictures with at least two responses associated with them were included in the low name agreement group (i.e.high-conflict trials).Thirteen items that were originally categorised as low-conflict were recategorised as high-conflict prior to analysis (see italicised items in Appendix B).

Procedure
Participants were randomly assigned to one CSE task that was comprised of either all PWI trials or all Name Agreement trials.For both tasks, participants were instructed to "name the picture as fast and accurately as possible".For the PWI task, participants were also instructed to ignore the overlaid distractor word.Example sequences for both versions of the CSE task are shown in Figure 2.Both experimental tasks were presented using PsychoPy open-source software (Peirce, 2009) with a Cedrus response box attached to a Dell desktop computer.Verbal responses were recorded via a head-mounted microphone.
The PWI version of the paradigm consisted of four blocks of 120 experimental trials and two initial filler trials which were excluded from analysis.To minimise order effects, stimuli were organised in four different orders and randomly assigned to participants.Within each block, trials were organised with the following constraints: (a) Half the experimental trials were low-conflict (i.e.congruent), and the other half were high-conflict (i.e.incongruent).(b) There were no more than four conflict level repetitions.(c) Each of the four possible conflict sequences (low-low [loLO], low-high [loHI], high-low [hiLO], high-high [hiHI]) occurred with equal probability within each block.(d) The higher-order conflict sequences were balanced.Each of the eight possible sequences of conflict across three trials (e.g.loloLO, lohiLO) occurred with equal frequency.(e) Each target picture appeared one time in each block (four times total in the experiment) and occurred in each conflict sequence across blocks (e.g.loLO in Block 1, loHI in Block 2, hiLO in Block 3, hiHI in Block 4).(f) There were at least 10 unrelated PWI targets between semantically related targets (defined as category membership) to minimise cumulative semantic interference (Freund & Nozari, 2018;Howard et al., 2006).Participants reviewed a slideshow of labeled images of all PWI targets and completed a practice block consisting of 10 trials prior to the experimental blocks.
The Name Agreement version of the paradigm consisted of two blocks of 120 experimental trials and two additional filler trials at the beginning which were excluded from analysis.This task contained only two blocks because each target stimuli can be presented in two unique conflict sequence conditions (i.e. a target picture is either low-conflict or high-conflict and can be presented in loLO and hiLO only or loHI and hiHI only) unlike the targets in the PWI task which can be presented in four unique sequence conditions (i.e.same picture can be in low-conflict and high-conflict conditionse.g,loLO, hiLO, loHI, hiHI).The trials within each block were organised with the following constraints based on the initial categorisation from the IPNP database: (a) Half the experimental trials were low-conflict, and the other half high-conflict.(b) There were no more than four conflict level repetitions.(c) Each of the four possible conflict sequences (low-low [loLO], low-high [loHI], high-low [hiLO], high-high [hiHI]) occurred with equal probability within each block.(d) The higher-order congruency sequences were balanced.Each of the eight possible sequences of conflict across three trials (e.g.loloLO, lohiLO) occurred with equal frequency.(e) Each target picture occurred one time in each block and occur in both conflict sequences across block (e.g.loLO and hiLO, or loHI and hiHI).Participants completed a practice block consisting of 10 trials prior to the experimental blocks, but did not review a slideshow of labeled images prior to the experimental trials.
In both tasks, each trial began with a fixation cross presented for 250 ms.A blank screen briefly appeared (50 ms) before the target picture (and written distractor for PWI) were presented.The picture stimuli remained on the screen for 2200 ms.The next trial started after an inter-trial interval of 1000 ms.

Analysis
The primary dependent variable of interest was reaction time on correct trials.Accuracy was also analysed but was expected to be relatively high for both tasks.Verbal responses were recorded as individual sound files beginning at the onset of each trial.Vocal onset Lowercase letters refer to the conflict status of the previous trial ("lo" and "hi" for low-conflict and high-conflict, respectively).Capital letters refer to the conflict status of the current trial ("LO" and "HI" for low-conflict and high-conflict, respectively).For the PWI task, the distractor word matches the picture in low-conflict trials and is semantically related to the target word in high-conflict trials.For the name agreement task, low-conflict trials consist of pictures with high name agreement and high-conflict trials consist of pictures with low name agreement.
of the responses (i.e.reaction time) was calculated using the virtual voice-key script in PsychoPy.Responses with vocal fillers/disfluencies, computer errors (e.g.voice key did not detect onset), or that were greater than 2.5 median absolute deviations from the participants' median reaction time were manually checked to ensure accuracy.Responses with short onset times (below 500 ms) were also manually checked as they were unlikely to reflect accurate responses (e.g.too quick to reflect word retrieval and production, likely to be extraneous noises).In total, 11% (721/6720) of the trials for Name Agreement task and 6% (760/12,480) of the trials for the PWI task were manually coded.
In all reaction time analyses, incorrect trials, trials following incorrect trials (to avoid post-error slowing), trials with computer errors (e.g.not recorded) and trials with verbal fillers (e.g."um", "uh") or disfluencies were excluded.Filler (e.g.initial trial of each block) and outlier (i.e. more or less than 2.5 median absolute deviations from the participant's median) trials were also removed prior to analysis.A total of 16.3% of trials were removed from the PWI analysis and 16.4% of trials were removed from the Name Agreement analysis.

Statistical analysis
All computations were performed using the R software environment (http://www.r-project.org/).Data wrangling and plotting were completed using tidyr (Wickham & Henry, 2019), dplyr (Wickham et al., 2019) and gglpot2 (Wickham, 2016) packages.Linear and logistic mixed-effects models, for reaction time and accuracy respectively, were fit using the lme4 package (Bates et al., 2015) and p values were obtained through the lmerTest package (Kuznetsova et al., 2017).Post-hoc pairwise comparisons were conducted using the emmeans function in the Rmisc package (Hope, 2016;Lenth et al., 2019) to determine significant differences between specific conditions of interest.All reaction time values were log-transformed prior to analysis to better approximate a Gaussian distribution.
Separate models were fit to analyse reaction time and accuracy for each task.In all models, fixed effects included current-trial conflict, previous-trial conflict and their interaction.Since each picture was shown multiple times (once per block; four times total in PWI, two times total for Name Agreement), block was also included to control for stimuli repetitions.Random intercepts were included to control for participant and item.The correlation of the random intercept of participant and random slope of block was included in all models to reflect variability in individuals' performance across the task (e.g.improvement due to repetition, decline due to fatigue).
Model selection was not conducted for the fixed effects as all terms were of interest and related to the design of the experiment.Model selection was conducted to determine the optimal random effects structure for all models following the method outlined in Harel and McAllister (2019).Each dependent variable was modeled as a function of several random effect terms, holding the fixed effects structure constant.Random intercepts of participant and item were deemed necessary to reflect the clustered nature of the data; therefore, models removing these random effects were not considered.Akaike information criterion (AIC) and Bayesian information criterion (BIC) were compared, and models with the lowest AIC/BIC values were selected.
Summary tables with mean reaction time and accuracy by task and condition are provided in Supplemental Materials 1. Summary tables of AIC/BIC values for each model and the full model output for the selected models are provided in Supplemental Materials 2. Sample size considerations and a post-hoc power analysis are provided in Supplemental Materials 3.

Discussion
This aim of this study was to determine whether unimpaired speakers engaged cognitive control processes in a similar manner to resolve two types of linguistic conflict during word production: prepotent conflict (e.g.PWI task) in which one dominant, but context inappropriate, response must be suppressed to select the correct target; and underdetermined conflict (e.g.name agreement task) when multiple reasonable responses compete and must be suppressed to select the desired target (Botvinick et al., 2001;Snyder & Munakata, 2008).The predicted CSE was found for both tasks albeit with small effect sizes; reaction time in word production tasks was influenced by the amount of conflict experienced on the previous trial (i.e.larger effect following low-conflict trials as compared to high-conflict trials), regardless of conflict type.This data suggests that resolution of prepotent and underdetermined conflict may engage the same general cognitive mechanism which is sensitive to the amount of conflict during processing.
While the pattern of behavioural data in our study is consistent with a CSE in both tasks, reaction time analysis represents only the end-product of a series of processes which are involved in conflict resolution (e.g.bias towards task-relevant information, select relevant information from working memory, resolving competition, response selection; Banich, 2019) and cannot provide information about similarities or differences in how these processes work at a neural level to resolve prepotent and underdetermined conflict.Therefore, neuroimaging is needed to identify if there are differences in when cognitive processes and which neural mechanisms are recruited to resolve prepotent and underdetermined competition during language production.
There is already some evidence to suggest that the resolution of different types of conflict rely on separable cognitive mechanisms and neural regions.Using fMRI studies of healthy speakers and neural network model simulations, Snyder et al. (2014) found that overlapping but partially independent areas within the prefrontal cortex are engaged when resolving different types of conflict.More specifically, they proposed that the competitive lateral inhibition within the left VLPFC facilitates the resolution of underdetermined conflict.That is, the most active representation suppresses the activation of multiple alternative responses that are activated during underdetermined conflict.Their results also suggest that the DLPFC is the primary source that controls the resolution of prepotent conflict and provides top-down support for task-relevant representation when competition is high.It is important to note, that since top-down biasing increase activation of all responses within a task set, this may increase underdetermined competition and slow response time after high-conflict trials which could result in the opposite of the canonical CSE pattern (Snyder et al., 2014).
There is also evidence that different types of conflict have distinct neural underpinnings from studies of people with prefrontal cortex damage.Investigations of two individuals with damage to the left inferior frontal gyrus (LIFG, also known as VLPFC) revealed deficits in tasks where the stimuli activated multiple representations that competed for each other (i.e.underdetermined conflict) but not when the stimuli activated one dominant response (i.e.prepotent conflict) (Robinson et al., 1998;Robinson et al., 2005).For example, in Robinson et al. (2005), patient CH's ability to complete sentences with a single word was better when there was one dominant response, but his performance was impaired when there were many possible appropriate and competing responses.
Differences in the neural underpinnings between conflict types are also supported by findings from Robinson et al. (2010) which revealed that individuals with frontal lobe damage that include the LIFG had selective impairments in a sentence completion task only when that task elicited multiple plausible responses that competed for selection.This deficit was not observed for individuals with damage in the frontal lobe that did not include LIFG and those who had lesions in the posterior regions of the brain.Cipolotti et al. (2016) also found that lesion location (right vs. left prefrontal cortex) was correlated with performance on tasks that engaged different types of conflict.Participants with left frontal damage had poorer performance on a prepotent conflict task (e.g.colour-word Stroop task) while participants with right frontal damage had impaired performance on a sentence completion tasks that elicits underdetermined conflict (i.e.part 2 of the Hayling task).
While there is some evidence which suggests that different neural underpinnings are involved in resolving different types of conflict during word production (e.g.Cipolotti et al., 2016;Robinson et al., 1998Robinson et al., , 2005;;Snyder et al., 2014), and that these differences may not be identified in behavioural analyses, there are also methodological factors to consider.Competition processes are difficult to study in unimpaired speakers because they monitor their speech efficiently and accurately and will make very few errors when performing tasks that are typically used with people with language impairments (e.g.confrontation picture naming; Nozari et al., 2016).Therefore, experimental tasks with overt distractors and less natural conflict such as the PWI task are frequently used.Tasks that invoke underdetermined conflict may have greater ecologically validity, even if they elicit single words, since they more closely resemble conflict that arises when speakers select from multiple plausible and competing words during spontaneous speech (Kurland et al., 2014).However, it can be challenging to design tasks which can reliability elicit this type of conflict.Sentence-level tasks (e.g.sentence completion) that vary conflict demands by including high and low conflict conditions have been shown to be sensitive to selection demands in healthy speakers (Barker et al., 2022) and in people with prefrontal cortex damage (Robinson et al., 1998(Robinson et al., , 2005)).For example, Barker et al. (2022) demonstrated that healthy participants were slower to complete sentences when there was high conflict demands (more possible options to select from) as compared to when there was lower demands (constrained sentences).In the current study, the name agreement task was chosen because it was a picture naming task with nouns that was more comparable to the PWI task than other tasks used to represent underdetermined conflict and a large stimulus set (120 pictures) could be more easily created to include in the CSE paradigm.Future studies should consider examining underdetermined conflict during word production using the CSE paradigm with additional tasks such as sentence completion.
Another methodological factor to consider is that tasks that elicit underdetermined conflict such as the name agreement task, the number of lexical representations which are activated and require processing and inhibition may vary across participants resulting in differences in the amount of conflict experienced on each trial between individuals.This differs from prepotent conflict in a task such as PWI in which it is known that all participants must process and override a task-irrelevant response on each trial (Aron et al., 2014).We will return to this issue in the section on underdetermined conflict during spoken word production.

Engaging cognitive control to resolve prepotent conflict during spoken word production
Previous work has demonstrated that cognitive control is engaged within word production tasks that involve prepotent conflict (e.g.picture-word interference; Duthoo et al., 2014;Freund & Nozari, 2018;Shitova et al., 2017;Spinelli et al., 2019).The current study replicated this finding and supports the notion that during a word production, speakers monitor their productions and regulate their behaviour online when prepotent conflict arises (Freund & Nozari, 2018).According to the conflict monitoring theory (Botvinick et al., 2001), the amount of top-down control can be adjusted based on the amount of conflict experienced on each trial.Resolving conflict on a high-conflict trial engages the control system which reduces the processing cost for subsequent conflict.Thus, participants who completed the PWI task were faster on high-conflict trials that followed other high-conflict trials as compared to those following low-conflict trials.Alternatively, topdown control is relaxed when experiencing low or noconflict trials resulting in less efficient processing on following trials (i.e.larger congruency effects; Berger et al., 2019; Figure 1).In the current study, participants were slower on low-conflict trials that followed high-conflict trials than those that followed low-conflict trials.Although there are alternative explanations proposed to account for this pattern of results (e.g.contingency learning, feature integration), the current study followed recommended procedures to minimise the contribution of alternative mechanisms so it was more likely that the findings resulted from an adaptive control process (Braem et al., 2019).
Although the contribution of bottom-up processes cannot be ruled out, procedures were followed to minimise their influence so that the findings could be more likely to be attributed to conflict monitoring theories (Braem et al., 2019;Schmidt, 2019).Stimulus-response learning was minimised in the PWI task as a large set of stimuli and responses were used, which resulted in every trial being a new stimulus (within each block).The number of consecutive low-conflict or highconflict trials was also limited to no more than four to minimise the participants' anticipating the congruency of the next trials.Although it would be better to have fewer consecutive repetitions of each trial type, the design was constrained by creating orders that were balanced in probability for both consecutive conflict sequences (i.e.loLO, loHI, hiLO, hiHI) and higher-order conflict sequences (e.g.loloLO, hiloHI).
The picture-word interference task has been used to study cognitive control because it is a simple and reliable paradigm to elicit competition and manipulate control demands during a word production task (Freund & Nozari, 2018).Prepotent competition arises during this task because the speaker must override one strongly competing lexical item to accurately name the picture (Schriefers et al., 1990).EEG results suggest that this effect emerges during the word planning stage (e.g.lexical selection; N400 component) after conceptual encoding is completed and prior to onset of articulation (Shitova et al., 2017).Thus, it has been argued that even though the interference during the PWI task is different than what speakers typically encounter during conversational speech, for purposes of studying cognitive control, it can engage the same processing level as "natural" word production (Freund & Nozari, 2018).While this is true, it is also important to consider how other types of competition, such as underdetermined conflict, are regulated during spoken word production.

Underdetermined conflict during spoken word production
This study examined the engagement of cognitive control during word production with underdetermined conflict using a CSE paradigm.Although a small but significant interaction between current-and previous-trial conflict was found, it was driven by changes in reaction time only for the low-conflict trials.Participants were slower to name pictures on low-conflict trials when they followed high-conflict trials, but the opposite finding did not reach significance.There may be three possible explanations for this pattern.First, it is possible that the difference in the two low-conflict conditions in the name agreement task is driven by factors that are unrelated to the CSE.For example, it is possible that the activation of multiple words carries over to the next trial, causing the low-conflict trials that follow high-conflict trials to be more like a high-conflict trial.However, this explanation would also predict higher reaction times on the high-conflict trials following high-conflict trials, which is not consistent with our data (see Figure 5).Second, it is possible that the effect is smaller for this task than for the prepotent conflict tasks that were used to power the study, and thus the effect for the high-conflict trials (which is often smaller than the low-conflict trial effect within the CSE paradigm) did not reach significance due to the number of trials being investigated.Third, studies of prepotent conflict have suggested that these effects may be dissociable and result from two distinct processing pathways (Erb & Marcovitch, 2018).The direct processing pathway leads to the primary, dominant response which in high-conflict trials generates a competing response.As a result, the control-demanding pathway must be engaged only for high-conflict trials to facilitate selection of the appropriate response (Erb & Marcovitch, 2018).It also is possible that different cognitive pathways are engaged for high and low-conflict trials in underdetermined conflict tasks.
The name agreement task was chosen in the current study to represent underdetermined conflict for several reasons: (a) It involves components of spoken word production including conceptualisation, lexical activation and selection, phonological encoding and articulation (Levelt et al., 1999).(b) It was comprised of a picture naming task with nouns that was more comparable to the PWI task than other tasks used to represent underdetermined conflict (e.g.verb generation task).(c) It is a relatively simple paradigm that reliably elicits competition effects (i.e.faster and more accurate responses for high name agreement/low-conflict items than low name agreement/high-conflict items; e.g.Kan & Thompson-Schill, 2004;Madden et al., 2019).
However, there are several limitations of the name agreement task.First, it uses an across participant measure to define name agreement for each picture.When responses from the participants in the current study were analysed, several items did not match the original categorisation criteria determined by the participants in the IPNP.This demonstrates there is variability in name agreement; a picture which may have many labels associated with it for one person (i.e.low name agreement), may only have one or two labels associated with it for another person.As a result, items which activate several responses for one person and be considered high-conflict, may only activate a single response for another person and be considered low-conflict.
In addition, the current study did not differentiate between the source of name disagreement within the low-agreement group (e.g.multiple correct names, abbreviations, incorrect names associated with the picture).Vitkovitch and Tyrrell (1995) found that the source of disagreement differentially affected naming latencies; only pictures with multiple correct names or incorrect names were named slower than the highagreement pictures.In a second experiment, they demonstrated that slower reaction times for the low agreement pictures with incorrect names were due to difficulty identifying the object while slow reaction times for low agreement pictures with multiple correct names were due to conflict after structural recognition of the objects.Taken together with later behavioural and neuropsychological studies, Madden et al. (2019) suggest that slower naming latencies for pictures with multiple plausible names is related to increased competition during selection among activated multiple lexical items, while slower reaction times for low name agreement due to other sources is not.
Another concern is that there were fewer trials in the name agreement task than the PWI task because each target stimuli can be presented only in two unique congruency sequence conditions (i.e. a target picture is either low-conflict or high-conflict and can be presented in loLO and hiLO only or loHI and hiHI only) unlike the targets in the PWI task which can be presented in four unique sequence conditions (i.e.same picture can be in low-conflict and high-conflict conditionse.g,loLO, loHI, hiLO, hiHI).The CSE is calculated as the difference between two differences (i.e. the congruency effect; Braem et al., 2019) and these effects tend to be small.Therefore, this design may not have had enough trials to have sufficient power to detect small effect sizes.For example, Freund et al., (2016) initially found a significant CSE effect for a language production task (PWI) that was driven by changes in reaction time only in lowconflict trials.However, when the number of trials were doubled in a subsequent experiment, they found significant effects for both low-conflict and highconflict trials.
Other individual participant factors such as age and vocabulary size have also been shown to influence picture naming latencies during the name agreement task (Madden et al., 2019;Paesen & Leijten, 2019).Although not pertinent to the current study because all participants in the current study were between 18 and 35 years old, the same age range considered "younger adults" in Madden et al. (2019), it would be important to consider these participant factors in future work which examines underdetermined conflict in clinical populations, as these participants tend to be older.Additionally, it has been shown that the size of the CSE can vary from person to person and may be related to individual differences in cognitive or affective processing (Weissman et al., 2015) or differences in application of strategy to resolve conflict in these tasks.For example, Duthoo et al. (2014) found the CSE for one group of participants but not in another group, even when the groups performed the same task.Future studies should consider assessing CSE using a within-subject design (e.g.Weissman et al., 2015) to minimize this potential confound.

Clinical implications for assessment and treatment of anomia
The current study examined the role of cognitive control in resolving linguistic conflict during language production in people with intact language skills.Although people with aphasia were not included in this study, it can provide some insights on considerations for assessment and treatment of people with language impairments.First, our findings highlight the importance of considering the control demands when using different assessment and treatment tasks.Results from this study demonstrate that speakers engage cognitive control in word production tasks with prepotent conflict most likely using similar computational principles as non-linguistic systems (Freund & Nozari, 2018;Nozari, 2018;Nozari & Novick, 2017).However, for underdetermined conflict a significant difference in reaction time in response to conflict was seen only for lowconflict trials.Therefore, it is important to consider the type and amount (e.g.high vs. low) of conflict each task/trial is invoking and determine how it is influencing the person's performance.If there are differences in the underlying cognitive or neural mechanisms for resolving prepotent and underdetermined conflict which we cannot determine based on behavioural data alone, it is possible that they could be differentially impaired (see Robinson et al. (1998) and Robinson et al. (2005) as examples).Therefore, a person may be able to resolve or compensate for underdetermined conflict more easily than prepotent conflict, or vice versa.
Speech-language pathologists may also capitalise on using conflict to increase or decrease the task difficulty during therapy.For example, a speech-language pathologist could increase task complexity by manipulating the amount of competition generated from a picture stimulus.This could be done in an overt manner (e.g.present along with semantic distractors) or more discretely (e.g.manipulating properties of the picture such as name agreement).Speech-language pathologists should also carefully consider the order of target/stimulus presentation.In the current study, performance on the current trial was influenced by what came in the previous trial for both prepotent and underdetermined conflict.Clinicians should be careful to balance presentation of different targets (and complexity of targets) because improvements (or declines) in performance may be directly related to the type of target in the preceding trial.

Conclusion
This study investigated the role of cognitive control to resolve conflict during spoken work production in unimpaired speakers.Trial-by-trial effects were analysed to determine whether a language task with prepotent conflict or underdetermined conflict engaged the control system to regulate online performance.The current work demonstrated that unimpaired speakers engage cognitive control to resolve conflict during word production with prepotent and underdetermined conflict.For both tasks, reaction time for picture naming varied depending upon the amount of conflict experienced in the current trial and the previous trial.However, in the task with underdetermined conflict this interaction was driven by the change in reaction time for the low-conflict trials only.These findings help us understand the relationship between cognitive control and language production, which can be used to inform clinical approaches for people with word finding difficulties.

Figure 1 .
Figure 1.Visual representation of conflict adaptation effect with hypothetical data.

Figure 2 .
Figure2.Example sequences for the congruency sequence effect paradigm for the picture-word interference (a) and name agreement (b) tasks.Lowercase letters refer to the conflict status of the previous trial ("lo" and "hi" for low-conflict and high-conflict, respectively).Capital letters refer to the conflict status of the current trial ("LO" and "HI" for low-conflict and high-conflict, respectively).For the PWI task, the distractor word matches the picture in low-conflict trials and is semantically related to the target word in high-conflict trials.For the name agreement task, low-conflict trials consist of pictures with high name agreement and high-conflict trials consist of pictures with low name agreement.

Figure 4 .
Figure 4. Mean percent accurate (with standard error bars) as a function of current-trial conflict and previous-trial conflict for PWI.

Figure 5 .
Figure 5. Mean reaction time (with standard error bars) as a function of current-trial conflict and previous-trial conflict for name agreement task.p-value <.05 indicated by asterisk.

Figure 3 .
Figure 3. Mean reaction time (with standard error bars) as a function of current-trial conflict and previous-trial conflict for PWI.p-value <.05 indicated by asterisk.

Figure 6 .
Figure 6.Mean percent accurate (with standard error bars) as a function of current-trial conflict and previous-trial conflict for name agreement task.