INTRODUCTION
Producing speech is one of the most common actions in which we engage. Like any other action, fluent and correct speaking requires self-monitoring. For many years, the dominant theory of monitoring in speech production has been the Perceptual Loop Theory (Levelt, 1983), in which signals from different stages of production planning are sent through the comprehension system and compared to an intended message. Recent research and modeling has suggested, however, that conflict signals arising within the production itself might serve as a critical cue to self-monitoring (Nozari et al., 2011). Such signals have been well-documented outside of language, with one prominent framework suggesting that all action domains might be monitored via a domain-general mechanism sensitive to detecting response conflict (e.g., Botvinick et al., 2001). This latter framework, however, has almost exclusively been investigated using manual responding in non-linguistic tasks. The goal of the present study is to test the viability of monitoring for response conflict in production by assessing whether similar electrophysiological and behavioral signatures of conflict are present in a speech production task and a standard, non-linguistic measure of conflict, the flanker task.
The Perceptual Loop account of monitoring in speech production provides a relatively simple and elegant framework as it does not require any additional machinery beyond that which is already provided by the comprehension system. According to this account, individuals monitor themselves after three stages of the speech production process: message retrieval, phonological encoding, and articulation. The Perceptual Loop account has recently been criticized, however, both on the grounds that internal self-monitoring does not seem to engage the speech comprehension (Huettig and Hartsuiker, 2010), and because of patterns of dissociations in patients who show impairments in comprehension but intact speech monitoring (see Nozari et al., 2011). Such criticisms have led numerous researchers to suggest that there may be specific monitors for different stages of production planning (see Postma, 2000 for a review).
An alternative to individual monitors at different stages of production planning is a more general account put forward by Nozari et al. (2011) who hypothesized that the production system might be monitored via domain-general, action monitoring systems that are sensitive to response conflict. Such conflict occurs when two or more representations are highly active prior to responding. In the case of language production planning, such conflict might arise, for instance, when two words with a similar meaning are both active at the same time (e.g., couch, sofa), or two speech sounds are concurrently active (e.g., /p/ and /b/). Using the two-step model of speech production of Dell and O’Seaghdha (1991), Nozari et al. (2011) were able to demonstrate that signals generated as the difference between the two most active elements at semantic and phonological levels of representation were good predictors of whether the model generated a lexical or phonological error. Moreover, the magnitude of these signals served as a good predictor of the speech production performance of a number of aphasic patients and the likelihood that they would detect an error. The authors tied this notion of conflict to a more overarching, action-monitoring framework provided by the response conflict model (Botvinick et al., 2001; Yeung et al., 2004).
According to the response conflict model, a region of the medial prefrontal cortex, the dorsal anterior cingulate cortex (ACC) is sensitive to situations where two competing responses are both active at the same time. This response conflict signal serves as a cue to regions of the dorsal lateral prefrontal cortex (DLPFC), which maintains task-relevant information and exerts control by biasing signals to task-specific brain regions to resolve the conflict. The response conflict model is based upon a relatively large cognitive control literature indicating ACC involvement in choice response tasks with manipulations of stimulus-response congruency such as the Flanker (Eriksen and Eriksen, 1974), Simon (Simon, 1969), and Stroop (Stroop, 1935) tasks. In these tasks, response conflict is induced either due to the incongruency between an intended response and the surrounding material (Flanker), a spatial position (Simon), or a pre-potent response (Stroop). Although there are a number of competing hypotheses about what the ACC is doing in such tasks (Holroyd and Coles, 2002; Rushworth et al., 2004; Aarts et al., 2008; Alexander and Brown, 2011), stable and replicable electrophysiological signatures of errors of commission and stimulus-response congruency have been established.
One such electrophysiological signature is the ERN, a centrally-distributed negativity that appears between 0 and 100 ms after an error has been committed (Falkenstein et al., 1991; Gehring et al., 1993). The ERN has been source-localized to the ACC in electroencephalography (EEG; Dehaene et al., 1994), and appears to have the same source regardless of output modality (e.g., hand vs. foot; Holroyd et al., 1998). Furthermore, the ERN is relatively stable within an individual across choice reaction time tasks (Riesel et al., 2013). As such, the ERN has been argued to be a marker of a domain-general, performance monitoring system, although researchers vary as to whether ERN marks response conflict (Yeung et al., 2004), reinforcement learning (Holroyd and Coles, 2002), the likelihood that an error will be committed (Brown and Braver, 2005), or the violation of an expected action outcome (Alexander and Brown, 2011).
One argument in favor of the response conflict account (cf. Alexander and Brown, 2011) is that it is highly correlated with the N2, a stimulus-locked negativity that is greater for incongruent relative to congruent trials (see Yeung et al., 2004). The N2 corresponds to the second negative peak in the ERP signal, and occurs within 200–400 ms following a stimulus. It has the same scalp distribution as the ERN, and occurs across a wide range of conflict-inducing tasks (for a review, see Folstein and van Petten, 2008). Furthermore, the N2 is sensitive to the amount of conflict across a block of trials, with the magnitude decreasing as the ratio of incongruent to congruent trials increases (Gratton et al., 1992), and the latency correlated with reaction times (Nieuwenhuis et al., 2003). These sets of results indicate that both the ERN and N2 are consistent markers of incongruency across stimulus-response dimensions and errors of commission, although to date they have been studied almost exclusively outside of language production.
Both markers, however, have been documented in language tasks with varying needs for production. The first demonstration of an ERN in language production came from Masaki et al. (2001) who showed an ERN for incorrect trials in the Stroop task with vocal responding. A series of studies by Ganushchak and Schiller (2006) also demonstrated that an ERN is present for errors in a phoneme-monitoring task while people covertly named pictures, and is also larger for semantically-related compared to unrelated blocks in a blocked-naming paradigm (Ganushchak and Schiller, 2008). This latter result was taken to indicate that the ERN can be sensitive to conflict arising not only from the response-level, but also from more abstract levels of representation (i.e., semantics). Whether one should expect similar brain signatures to representational vs. response conflict has been major issue of study in the cognitive control literature (see Milham et al., 2001; Egner, 2008), and we return to this point in the general discussion.
Recent research has also indicated that ERN-like components are also present during correct production as well. Using a picture naming paradigm, Riés et al. (2011) not only demonstrated an ERN on incorrectly named pictures, but also showed a relatively smaller negativity occurring in the same time window for correct responses. This negativity was argued to be a marker of an error-monitor system that is active both on correct and incorrect trials. A recent study using bilingual picture naming also demonstrated that the magnitude of an ERN-like component is sensitive to response conflict (Acheson et al., 2012). Acheson et al. (2012) showed that a relatively late (100–200 ms) frontal–central negativity was larger when bilingual German/Dutch speakers named pictures that were cognates (e.g., Haus/huis – English: house) relative to those that are not (Frosch/kikker; English – frog). The authors argued that despite having shared semantics and phonology which can lead to faster speech initiation times (i.e., the cognate facilitation effect; Costa et al., 2000), cognates might also induce response conflict as they differ during phonetic encoding and articulation – the response-levels of speech production. The timecourse of this response-locked negativity was somewhat later than is typically observed for the ERN, which may have occurred either because the conflict itself did not occur until later in the word (all began with the same speech sounds), or alternatively, because the negativity reflected something akin to the feedback-related negativity (Holroyd and Coles, 2002).
The study by Acheson et al. (2012) is significant in demonstrating that speakers are sensitive to response conflict during speech production, and that this can occur even after responding has been initiated. In addition to this response-locked signature, stimulus-locked indicators of conflict have also been observed using the SLIPs paradigm (Baars et al., 1975). Using overt-production, Moeller et al. (2007) demonstrated that 400–600 ms after stimulus-onset, a frontal–central negativity was observed that was larger before a speech error occurred relative to when it did not, and this negativity was source localized to an area just superior to the ACC, the pre-supplementary motor area (pre-SMA). More recent research using the same SLIPs paradigm demonstrated that this negativity is larger when the experimentally-induced error is about to produce a taboo word (Severens et al., 2011). Both of these studies were taken to indicate that the system responsible for monitoring speech is sensitive to conflicting responses prior to speaking.
The stimulus-locked N2 has also been found in numerous ERP studies of language. For instance, many studies have combined covert production with a Go/No-Go task requiring button presses. Such studies have shown that the timing of the N2 difference between Go/No-Go trials is sensitive to the timecourse of production planning, with the N2 for semantic decisions preceding phonological ones (e.g., Schmitt et al., 2000; Rodriguez-Fornells et al., 2002). Numerous studies using bilingual language paradigms have also found evidence of N2 modulation (see Rodriguez-Fornells et al., 2006 for a review). For example, the Go/No-Go paradigm has been used to demonstrate cross language activation at the phonological level (e.g., Rodriguez-Fornells et al., 2005), and an N2 has been observed in language switching paradigms requiring overt production (e.g., Christoffels et al., 2007; Verhoef et al., 2009).
The picture-word interference paradigm has also provided evidence of ACC involvement in language production. For instance, the N2 has been implicated in the picture-word interference paradigm for the need to inhibit non-linguistic distractors (e.g., “XXXX”; Hirschfeld et al., 2008), although later components seem to show sensitivity to linguistic distractors. For instance, phonologically-related distractors produce a larger frontal–central negativity than unrelated distractors starting around 400 ms after stimulus onset (Dell’Acqua et al., 2010). FMRI research has also revealed that portions of the ACC are sensitive to phonological relative to unrelated distractor words in the picture-word interference paradigm (de Zubicaray et al., 2001), although it should be noted that the region of the ACC observed in this study was more anterior and inferior to that which has been observed in the studies of using choice response tasks. What is noteworthy about the studies indicating that the N2 is sensitive to linguistic manipulations is that they all involve tasks in which participants must inhibit a response, thus it is unclear whether these effects reflect conflict monitoring or inhibition.
The research reviewed above thus highlights a number of similarities between mechanisms and signatures of response conflict monitoring outside of language and those that might occur within the production system itself. Although the brain-region responsible for monitoring conflict – the ACC – is purported to be domain-general, to date, no direct comparison has been made to test for comparable ACC engagement in language production and conflict-inducing, choice-response tasks. The present research was designed to address whether similar signatures of response conflict are present across these domains, and furthermore, whether these signatures are related to each other.
In order to induce response conflict in language production, native speakers of Dutch rapidly read phonotactically legal nonwords organized in tongue twister (TT) and non-TT patterns (see Wilshire, 1999 and Methods section below).The choice of nonwords in the present study was designed to focus on the phonological encoding and articulation-levels of production planning while limiting influences from lexical and semantic representations. The TT pattern causes more errors than non-TT, and was thus taken to be our high-conflict condition. Based on the above-described research, a number of predictions can be made if the same mechanisms of conflict monitoring are present in speech and non-speech tasks. First, errors should elicit an ERN in the response-locked analysis in both tasks, and their scalp topography should be the same. Second, differences between high and low conflict conditions should be observed in the N2 time window, with high conflict conditions more negative than low conflict conditions (i.e., TT > non-TT; flanker incongruent > flanker congruent). Third, there should be correlations between people’s behavioral performance and EEG signatures of conflict within both tasks. Finally, if the process of response conflict monitoring and error detection is domain-general, then there should be correlations between EEG signatures and people’s behavioral performance across both tasks.
GENERAL DISCUSSION
The present study was designed to test whether monitoring in language production might be achieved through a domain-general, response conflict monitoring system. EEG was recorded while people performed the TT and flanker tasks, linguistic and non-linguistic tasks designed to elicit varying degrees of response conflict. In order to test whether similar mechanisms were present in both tasks, EEG and behavioral signatures of response conflict and error commission were quantified and compared both qualitatively and quantitatively. Results showed that both the flanker and TT tasks elicited a centrally-distributed ERN. Replicating previous results, incongruent trials in the flanker task elicited a larger N2 than congruent trials. Critically, no such signature was present in the TT task, in which a broadly distributed frontal negativity was observed for the difference between TT and non-TT conditions. ERP-behavior correlations showed some correlations within a task, but few between the two tasks. Taken together, although these results do not necessarily speak against conflict-monitoring as a mechanism for monitoring in speech production, they do not support a common-locus to this mechanism across task domains. These results have implications for the breadth of the conflict-monitoring hypothesis generally, and monitoring within language-production more specifically.
The current study replicates and extends previous research in a number of ways. With regards to the flanker task, we replicated the N2 difference for congruent and incongruent trials, the ERN for correct and incorrect trials, and importantly, the correlation between these two EEG signatures. This latter result is consistent with the conflict monitoring hypothesis (see Yeung et al., 2004), but also a more recent account that attributes the N2 and ERN signals to an action-outcome prediction process (Alexander and Brown, 2011), a theoretical point we return to below.
Although the TT task has not previously been studied with EEG, the current pattern of results also bears similarity to previous studies using other error elicitation paradigms. Similar to the Masaki et al. (2001) study using a vocal version of the Stroop task, we also observed an ERN for errors compared to correct trials. The timing and maximal scalp topography of the ERN in the present investigation is very similar to that which was observed in by Masaki et al. (2001) and is somewhat broader than the ERN observed for the flanker task. Importantly, the ERN in the current study was positively correlated with the TT effect for item ordering errors, and marginally with people’s overall error rates, confirming that there is a relationship between the generator of the ERN and people’s self-monitoring abilities. The combination of these results suggests that the ERN in language production may reflect the activity of a continuous performance monitoring system.
In addition to the ERN, there was a sustained difference between TT and non-TT trials prior to responding. This sustained difference parallels that which has previously been observed preceding error trials in the SLIPs paradigm (Moeller et al., 2007; Severens et al., 2011). In contrast to the studies of the SLIPs paradigm, the time course of the differences began earlier (50 ms) and were sustained up until just before speech initiation began. This difference likely reflects the summation of multiple potentials, as evidenced by the scalp distribution of the CSD.
The timecourse and distributed nature of TT effect make it difficult to pinpoint when differences between TT conditions were occurring during production planning. Despite the TT conditions being matched across participants (i.e., different participants saw the same items in tongue-twister and non-TT versions), it remains possible that some of the differences reflect differences in reading the pseudowords. Although it is beyond the scope of the current study to fully address this issue, it is possible that some of the very early effects reflect differential sensitivity in the N170 to visual properties during reading, and later differences with accessing the phonological properties of words in the N320 time window (see Bentin et al., 1999 for a review). Regardless of the specific source (i.e., in reading or production planning), these results suggest that the difference between the two TT conditions are unlikely to be driven solely by the same generators that are implicated in the flanker task. In turn, this suggests that the phonological nature of the conflict in the TT task that was occurring prior to responding may reflect a form of representational conflict instead of response conflict. We return to this discussion below, but first address one of the central issues for which this study was designed: whether monitoring in production is accomplished via a domain-general, response conflict monitor.
One of the central predictions of conflict-monitoring hypothesis is that a region of the medial prefrontal cortex, the ACC, serves as a domain-general conflict monitor (Botvinick et al., 2001). The domain-generality of the ACC functionality has been adopted by recent models of monitoring in language production (Nozari et al., 2011), and also has been emphasized in a number of papers studying language production as well (e.g., Rodriguez-Fornells et al., 2006; Hirschfeld et al., 2008; Riés et al., 2011; Acheson et al., 2012). The relationship of the ACC to the electrophysiological markers used in this study (the N2 and ERN) has been well-established outside of language (see Yeung et al., 2004), thus the current research provides an important test of this domain-generality. Previous research that has implicated similar electrophysiological markers of ACC response have used testing conditions that either used the same task with different motor effectors (e.g., Holroyd et al., 1998), or different tasks with the same motor effectors (e.g., Riesel et al., 2013). Similar to these previous results, there was a correlation between the N2 and ERN in the flanker task. Importantly, however, the ERNs were not correlated across tasks.
There are, of course, a number of reasons why we may have failed to find significant correlations. For instance, the ERN generated in the production task was likely jittered in time compared to the flanker task, given that errors occurred at multiple points in a nonword. Another possibility is that the signal-to-noise ratio was worse in the TT task as the ERN was calculated from fewer error trials. A third possibility is that the lack of correlation is a result of different response modalities (i.e., verbal vs. manual; but see Holroyd et al., 1998). These less interesting points aside, one intriguing possibility is that people may be differentially sensitive to different types of errors, which would suggest some amount of domain-specificity to the error monitoring process. Of particular note here is the Laplacian transformation (i.e., CSD), which showed a standard, frontal–central ERN in the flanker task, but more distributed frontal and central generators of the ERN in the TT task. While this suggests some differences between the two generators, CSD is not an optimal measure for finding deep sources such as the ACC, and the current study was not optimized to directly test for differences in source location. If the observed differences are genuine, however, the current results have important implications for theories of action monitoring as the ERN has been taken as a biomarker of error-processing outside of language. Not only does the lack of correlation in the ERN across manual and vocal responding fail to support a domain-general conflict monitor, it is also not consistent with a single, domain-general, error-monitoring system (e.g., Holroyd and Coles, 2002). While this result is potentially interesting, it is a null result, and will need to be tested in future experimentation using a more spatially-sensitive measure (e.g., fMRI), and possibly in tasks that utilize the same output modality.
The present investigation adds to a growing body of evidence that has failed to find evidence of a domain-general, conflict monitoring system. Research using standard cognitive control tasks, for instance, have failed to find that correlations in people’s behavioral response to conflict across a range of conflict-inducing tasks (e.g., Stroop; Fan et al., 2003; Keye et al., 2009). Furthermore, attempts to find cross-task, conflict-adaptation effects have often failed to find any spillover effects from one task onto another (see Egner, 2007), as have those that have tried to find cross-task interference by virtue of factorial designs between two conflict-inducing tasks (e.g., a combined Simon-Flanker task; see Egner, 2008). Finally, attempts to relate the neural signatures of responses of conflict have revealed that while there may be some spatial overlap in the brain response to conflict across tasks within the ACC, there are also noteworthy differences in the spatial location of these responses (Fan et al., 2003).Taken together, these results have led researchers to conclude that while the mechanisms of conflict monitoring might be domain-general, the actual brain regions implementing this monitoring and exerting control may be domain-specific (Milham et al., 2001; Egner, 2008). An alternative view could be accommodated within Alexander and Brown’s (2011) PRO model, which equates ACC activation to a process of comparing predictions about likely action outcomes to the action that is taken. Here, one could assume that this process of prediction and comparison is domain-general, but that the actual action domains over which possible outcomes are being predicted are domain specific.
Such a view would be broadly consistent with neuroimaging studies that have implicated different regions of the ACC for conflict- and error-monitoring (Garavan et al., 2003), as well as studies that have demonstrated different functional subdivisions within the ACC depending on task and motor output (Paus et al., 1998). Critically, the activity of the ACC seems to be related to situations in which there is response- but not representational conflict (Milham et al., 2001; but see Aarts et al., 2008). Within linguistic representation and responses, for instance, a study using a lexical-decision task with bilinguals found that lateral prefrontal regions were sensitive to stimulus-based (i.e., representational) conflict, while the ACC became active only when this conflict led to different responses (van Heuven et al., 2008). Representational conflict here is a form of conflict that would arise pre-response, for instance, as a result of lexical or semantic interference between color words and ink color in the Stroop task. Such pre-response, representational conflict may be functionally equivalent to the competition that is at the heart of selection mechanisms in models of language production (e.g., Levelt et al., 1999). Numerous studies have implicated that other frontal brain regions, such as the inferior frontal cortex (IFG) are sensitive to representational conflict arising due to incompatible stimulus representations (e.g., Stroop interference) or situations of under-determined responding (e.g., in verb generation; see Novick et al., 2005). With regards to monitoring and control operations more generally, these studies point to a network of brain regions involved in the monitoring, detection and response to conflicting stimulus and response dimensions. These networks, in turn, may be organized as separable systems depending on the representational source (e.g., visual, verbal) and the motor effectors (e.g., manual, articulatory) influenced by conflict. The tasks used in the current study maximally differentiated both representational and response conflict, thus future experimentation might be focused on manipulating only one of these two sources (e.g., using verbal responding across different sources of representational conflict), or through conflict adaptation paradigms, such as task switching or factorial manipulations of similar sources of conflict.
Bringing this back to a conflict-monitoring hypothesis for language production, the model put forward by Nozari et al. (2011) suggested that conflict signals can be generated both at the level of lexical-retrieval and phonological encoding by, for instance, calculating the difference between the two most active elements at the time of selection. Although the authors discussed this research within a larger, response conflict framework, the present results suggest that it may be more appropriate to discuss these signals in terms of pre-response (i.e., representational) conflict. This does not diminish the fact that signals within the production system itself might serve as cues to monitoring and the need to increase cognitive control. Rather, there may be more domain-specificity to this signal than was originally emphasized. The current study thus suggests that it is unlikely that all tasks domains are mapping onto the exact same conflict-monitor. What remains possible, however, is that the signal to increase monitoring and control could be domain-general, but that it would originate from domain-specific systems and stages of planning. Put differently, the current research suggests a model of monitoring in speech production in which signals arising at many different levels of the planning process (e.g., message, grammatical, lexical, phonological, articulatory) could serve as cues to increase monitoring and control. Each of these signals would, presumably, originate in the cortical networks responsible for each of these stages, and the signals themselves might be mediated via distinct temporal, medial frontal, lateral frontal, and midbrain loops (see Alexander et al., 1986 for a review of cortical-midbrain loops). Ultimately, the extent to which this might lead to response conflict in particular is dependent upon whether activation of two competing response alternatives cascade all the way to articulatory planning.
Although the current paper and discussion has been motivated by a response conflict monitoring perspective, the present results might also be accommodated within a recent account suggesting that signals for monitoring and increased cognitive control could emerge as a result of prediction errors (see Alexander and Brown, 2011). This latter account differs from a conflict monitoring hypothesis in a number of ways, not the least of which is that it requires some mechanism of comparing intended with actual responding. According to the PRO model, the signal that arises from incongruous conditions does not arise from conflict, per se, but rather, because multiple responses have been predicted, and only one of them was executed. The difference between predicted and actual responding results in a prediction error. This proposed mechanism is, in many ways, similar to recent proposals by Pickering and Garrod (2013), who suggest that prediction serves a central role in production planning and monitoring. The mechanism of prediction in the PRO model is the learned association between a stimulus to possible response outcomes, and not from response to possible sensory consequence. As such, prediction in the PRO model bears more similarity to what Pickering and Garrod (2013) refer to as prediction by association rather than prediction via forward models.
Framing this prediction-based account within the staged process of language production, “response” in this instance could be viewed as what is sent from one planning stage to another. Given that this process is likely to be cascaded, the prediction-error account suggests that partially processed information from one stage of production planning (e.g., message-retrieval) would lead to “predictions” (i.e., partial activation) of items at the next stage of planning (e.g., multiple lexical candidates). Ultimately, one of these candidates is selected, thus leading to a form of prediction error at the following stage of production planning. Whether the above-described comparison process necessarily requires the comprehension system (as proposed in the Perceptual Loop theory; Levelt, 1983) remains to be determined. It does suggest, however, that there should be a high degree of domain-specificity with regards to the predictions that are being made. In the current study, for instance, the use of nonword stimuli emphasized phonological encoding over lexical-semantic representation, which was done to elicit conflict primarily at late stages of the production planning process. Such monitoring may or may not be different from that which is involved in monitoring at the lexical-semantic level, and this remains an important area for future investigation.
To conclude, the current study sought to test whether a domain-general, conflict monitoring mechanism might serve as a cue to monitoring in speech production. By comparing EEG and behavioral signatures of conflict across both speech and non-speech tasks, the present results provide little evidence in favor of a single, conflict monitoring mechanism. Similar findings have been observed outside of language. This pattern of results thus suggest that while the mechanism and signals to increase monitoring and control may be domain general, the representations over which they are generated are likely to be domain-specific.