Word production involves the activation of both a target word and its neighbors that share semantic and/or form-based features. The result is a complex pattern of interference and facilitation arising from the dynamic nature of the processes involved in activation and selection. In this study, we used a word production paradigm in which robust interference effects have been shown to result from the production of semantically related words, investigating whether the same is true for words that share segments. We considered whether mechanisms that have been proposed to account for certain of these effects in lexical selection—the process that identifies a specific word to convey the intended meaning—may also apply in segmental encoding—the process that identifies the segments to express the selected word. Furthermore, we examined whether these findings extend across segment types (phonemes and graphemes) in both spoken and written word production.

Interference and facilitation in lexical selection and segmental encoding

Although word production (e.g., in picture naming or reading) is often facilitated by the prior presentation of a single semantically related word immediately before the target, this benefit is quickly eliminated or masked by the insertion of an intervening item (e.g., Wheeldon & Monsell, 1994). In contrast, the work that we report here was concerned with paradigms in which items are named in the context of multiple similar items, a situation in which robust, long-lived interference and facilitation effects have been consistently reported. In the continuous naming paradigm, in which semantically related and unrelated pictures are interleaved, participants are slower to name each consecutive item from a semantic category, even when many unrelated items intervene (e.g., Howard, Nickels, Coltheart, & Cole-Virtue, 2006; Schnur, 2014). In the blocked cyclic naming paradigm, in which small sets of pictures are named repeatedly, participants are slower to name items from the same semantic category than to name those same items rearranged into unrelated sets (e.g., Crowther & Martin, 2014; Damian, Vigliocco, & Levelt, 2001; Schnur, Schwartz, Brecher, & Hodgson, 2006).

Two main accounts have been proposed to explain the semantic blocking effect (Howard et al., 2006; Oppenheim, Dell, & Schwartz, 2010), both of which rely on incremental learning, as the basis for long-lived changes within the production system, and on competive mechanisms to explain interference effects. We will briefly describe Oppenheim et al.’s account, because it does not rely on competitive selection to account for cumulative semantic interference (see Navarrete, Del Prato, Peressotti, & Mahon, 2014, for arguments against competitive selection), but instead assumes that learning is a competitive process that both strengthens and weakens connections between representations, providing an account of faciliation and interference effects in certain contexts. Specifically, upon successful naming of a picture, connections between the correct lexical entry (e.g., cat) and its semantic features (e.g., furry, pet) are strengthened. For other lexical items that are active on that trial, connections from the lexical entries to the semantic features that activated them are weakened. Because a semantically related item (e.g., dog) is much more likely to become activated through the semantic features it shares with the target than is an unrelated item (e.g., spoon), dog is more likely to undergo weakening of connections to its features than is spoon. Thus, when the target on the next trial is dog, it is at a disadvantage relative to spoon.

Although incremental learning accounts have primarily been directed toward semantic–lexical mapping, they may also apply to lexical–segmental mapping. Like semantic–lexical mapping, lexical–segmental mapping is affected by learning: For example, the frequency effect, which is a hallmark of experience-based adjustment to the production system, strongly affects lexical–segmental mapping (see, e.g., Kittredge, Dell, Verkuilen, & Schwartz, 2008, and the references therein). In one example of incremental learning applied to this stage of processing, Mulatti, Peressotti, Job, Saunders, and Coltheart (2012) found increased response latencies for the production of words that overlapped in rhyme with previously produced words. However, this was demonstrated in reading in an orthographically transparent language (Italian), and the observed interference may have arisen in nonlexical grapheme-to-phoneme mapping. How might competitive incremental learning work in lexical–segmental mapping? When a word is produced, the connections between the lexical entry and its segments are strengthened, and the connections between other lexical entries and their segments are weakened in proportion to their activation levels. Critically, due to feedback, these other active items are likely to be lexical entries that share segments with the target. For example, when cat is the target, the segments /æ/ and /t/ are activated. Through feedback, they activate other lexical entries that share these segments (e.g., mat). These lexical entries, in turn, activate their remaining segments, including those that are not shared with the target (e.g., /m/). When the correct target is selected (e.g., cat), the connections between these nontarget lexical entries and the unshared segments (e.g., mat’s connection to /m/) are weakened. Segments of an unrelated word (e.g., spoon) are much less likely to be activated and to undergo weakening. Therefore, when the next target is mat, it is at a disadvantage relative to spoon. Thus, interference is predicted in the context of segmental overlap. Note that there are important differences between semantic–lexical and lexical–segmental mappings. Whereas semantic–lexical mapping involves connecting many semantic representations to one lexical representation, lexical–semantic mapping requires connecting one lexical representation to many segments. Therefore, although shared features lead to interference in both mappings, the sources of this interference differ. In the case of semantic–lexical mapping, interference is caused by the adjustment of connections between shared semantic features and competing lexical items, whereas in lexical–segmental mapping, the interference results from the adjustment of connections between similar lexical items and their competing features. In both cases, however, the adjustments are intended to decrease the strength of the competitors vis-à-vis the current target word, reducing the availability of those competitors (interference) on subsequent trials.

Our description is not intended to imply that there is necessarily a single representation at the lexical level. Our proposal is equally compatible with several architectural alternatives with regard to lexical representations, including systems with abstract lemmas only (with lexical syntactic information), with only modality-specific phonological and orthographic lexemes (with links to syntactic features), or with both lemmas and lexemes (see Caramazza, 1997, for a review of models with and without lemmas). In a system with both lemmas and lexemes, the incremental learning mechanism that we discussed above would apply to the connections between lexemes and segments.

Because the effects of segmental overlap are somewhat ambiguous in the current literature, it is not clear that this type of competitive learning mechanism actually applies in segmental encoding. Both facilitation and interference effects have been reported for segmentally overlapping words (e.g., Damian & Dumay, 2009; for a review, see Sevald & Dell, 1994, and the references therein). Importantly, situations that typically produce interference for semantically related items, including the blocked cyclic naming paradigm, have not shown robust interference for form-related items. When items in a block consistently share onset segments, production is facilitated (e.g., Damian, 2003; Roelofs, 1999; Schnur et al., 2009). Shared-onset facilitation is widely attributed to the words’ high predictability allowing strategic preparation (e.g., Damian, 2003; Meyer, 1991). This likely arises outside the language production system (O’Séaghdha & Frazer, 2014), and could mask interference effects generated within the system. Evidence generally consistent with the prediction that removing predictability reveals underlying interference from segmental overlap has come from Belke and Meyer (2007), who reported that facilitation effects on response latencies disappeared when multiple onset-related items were named within a single trial, and that gaze durations to onset-related items increased when items were named quickly. However, overt interference effects comparable to those in semantically related naming paradigms have not been reported. This could mean that competitive incremental learning is not operational in segmental encoding, or that the predicted interference is masked by the strategic preparation that is possible in conditions with predictable initial-segment overlap.

In the present study, we investigated these possibilities by examining the consequences of segmental overlap when it is distributed unpredictably across positions in words, as opposed to being limited to the first position (e.g., pill in the context of pig, peg, pot, log, leg), in a blocked cyclic naming paradigm. We examined picture naming because it necessarily includes both lexical selection and segmental encoding. Although some evidence of segmental overlap interference has been shown in reading (e.g., Mulatti et al., 2012) and repetition (e.g., O’Séaghdha & Marin, 2000; Sevald & Dell, 1994), these results may have been due to similarity in the input or to nonlexical processing in the tasks, making these tasks less appropriate for investigating questions concerning lexical selection and segmental encoding. Removing predictability reduces opportunities for strategic preparation, which allowed us to evaluate whether principles such as competitive incremental learning operate during lexical selection and segmental encoding.

We examined both spoken and written word production. Because similar organizational principles have been observed across the two modalities (e.g., Bonin & Fayol, 2000; Rapp & Fischer-Baum, 2014; Shen, Damian, & Stadthagen-Gonzalez, 2013), the extension of the work to written production provides an opportunity to replicate findings with orthographic segmental encoding processes that, presumably, should operate according to principles similar to those of phonological segmental encoding. Furthermore, given that spoken and written production differ considerably in terms of response execution, with speaking taking place over a shorter time course and in a more parallel fashion than writing, examining both modalities provided an opportunity to evaluate the robustness of findings across these considerable task differences.

Experiment 1: Distributed segmental overlap in spoken word production

In Experiment 1, we investigated the effects of unpredictable segmental overlap in a spoken blocked cyclic naming paradigm, investigating whether interference occurs in this situation. Such a finding would suggest that similar principles underlie lexical selection and segmental encoding.

Method

Participants

A group of 24 individuals (13 female, 11 male; mean age 19.8, 20 right-handed) participated. For both experiments, the participants were native English-speaking undergraduate students who gave informed consent and received course credit for participation.

Stimuli

Pictures corresponding to 36 monosyllabic three- to six-letter words were selected to create six homogeneous lists with high position-independent phonological overlap. They were rearranged to form heterogeneous lists (Table 1). Position-independent phonological overlap, defined as the total number of phonemes shared by two strings, regardless of position, divided by the total number of phonemes in the two strings (Goldrick, Folk, & Rapp, 2010), was significantly greater in the homogeneous (mean = .44) than in the heterogeneous (mean = .08) lists, t(5.39) = –10.96, p < .001, 95% CI of difference = .28–.45. The stimuli were black-and-white line drawings of objects freely available online. Using Amazon Mechanical Turk, a separate group of participants (N = 40) rated the visual similarity of all 180 pair-wise picture combinations on a scale ranging from 1 (not at all similar) to 5 (very similar), following Schnur et al. (2006). The ratings for pairs from the homogeneous block (mean = 1.4) were not significantly different from those from the heterogeneous blocks (mean = 1.3), t(478) = –0.83, p = .41, 95% CI = –0.13–0.05), so visual similarity was not included in the models.

Table 1 Items used in Experiments 1 and 2
Full size table

Procedure

The experiment was run using E-Prime 2 Professional (Psychology Software Tools, Pittsburgh, PA). Picture stimuli (7 in. by 5 in.) were displayed at the center of a 19-in. × 12-in. monitor approximately 20 in. in front of participants, who responded by speaking into a microphone (E-Prime’s stimulus–response box voice key captured the response times [RTs]). In the familiarization phase, participants saw each line drawing and silently read its provided label.

Following familiarization, participants completed 36 practice trials in which they named each item aloud once, receiving corrective feedback. On each trial, a fixation cross appeared on the screen for 1,000 ms, followed by a stimulus picture that remained on the screen until participants initiated the response (RT) or for 3,000 ms, if no response was made. Recordings of the responses were used to score accuracy. At the end of each trial, a button was pressed to continue, followed by a fixed 3-s intertrial interval.

Next, participants completed six homogeneous and six heterogeneous blocks, in pseudorandom order, with periodic breaks. Each block consisted of four cycles of the six pictures from one list presented in random order. The trial structure was the same as in the practice session.

Analysis

Error responses in which participants did not correctly produce the intended label and outliers more than 2.5 standard deviations from each participant’s overall mean were removed from further analysis. For the analysis of RTs, only Cycles 2–4 were considered, because past results had indicated that block type effects typically emerge only after the first cycle (e.g., Belke, Meyer, & Damian, 2005). Repeated measures 2 × 3 analyses of variance (ANOVAs) were conducted using IBM SPSS (version 21), including Block Type (homogeneous or heterogeneous) and Cycle (2, 3, 4) as within-subjects factors. The dependent variable was mean RT, calculated for each participant in the F 1 or for each item in the F 2 analysis for each of the six block-type-by-cycle conditions. In a secondary analysis, we conducted the same by-participants ANOVA, but included only the 21 items that shared their first segment with at least half of the items in their homogeneous block.

Results and discussion

Of the total 6,912 trials, 8% were removed due to technical errors, incorrect responses, or outlier status. Participants were slower to initiate the production of items in homogeneous than in heterogeneous blocks, F 1(1, 23) = 29.90, p < .001, mean difference = 17.0, SE = 3.11, 95% CI = 10.6–23.4, η p 2 = .57; F 2(1, 35) = 25.21, p < .001, mean difference = 18.4, SE = 3.67, 95% CI = 11.0–25.9, η p 2 = .42. No significant main effect of cycle was apparent, F 1(2, 46) = 0.16, p = .80, η p 2 = .01; F 2(2, 70) = 0.66, p = .52, η p 2 = .02, nor was an interaction of cycle and block type, F 1(2, 46) = 1.12, p = .33, η p 2 = .05; F 2(2, 70) = 2.40, p = .10, η p 2 = .06 (Fig. 1a). Thus, distributed segmental overlap resulted in interference in spoken production.

Fig. 1
figure 1

Results of Experiment 1: Effects of distributed segmental overlap on response times (RTs) for spoken picture naming of words in blocks with high segmental overlap (homogeneous) versus low segmental overlap (heterogeneous). Error bars represent the between-subjects standard errors of the means, corrected for repeated measures using the Cousineau (2005) method. Panel A includes the data from all items. Panel B depicts the secondary analysis that included only items that shared their initial segment with at least half of the other items in the homogeneous block

Full size image

In the secondary analysis, including only items that shared their initial segment with at least half of the other items in the homogeneous block, interference was again observed. Participants were slower to initiate production in homogeneous blocks, F(1, 23) = 21.55, p < .001, mean difference = 17.1, SE = 3.68, 95% CI = 9.5–24.7, η p 2 = .48, even when the items shared onsets—the condition that past research has suggested is most likely to yield facilitation in predictable situations (Fig. 1b).

Experiment 2: Distributed segmental overlap in written word production

In Experiment 2, we tested whether the results of Experiment 1 could be replicated in written production. A similar interference effect would indicate that reliable segmental overlap interference occurs regardless of whether the segments are phonemes or graphemes.

Method

Participants

A total of 34 individuals (21 female, 13 male; mean age 19.2, 28 right-handed) participated. The sample was increased in this experiment due to borderline results with the initially planned 24 participants.

Stimuli

These were identical to the stimuli in Experiment 1. The relevant overlap type was orthographic rather than phonological, with significantly greater position-independent letter overlap in the homogeneous (mean = .57) than in the heterogeneous (mean = .16) lists [t(5.474) = –7.52, p < .001, 95% CI = 0.27–0.53].

Procedures

The procedure was identical to that of Experiment 1, except that participants responded by writing their response on a graphics tablet (Wacom Bamboo), which they were trained to use before the experiment. Participants’ responses were constrained to a 2.5-in. × 1-in. rectangle centered at the bottom of the responsive surface. On each trial, participants began with the pen on a marked starting point centered 1 in. below the writing surface. They were instructed to begin writing their response in either cursive or print when they knew the name of the picture. As soon as the writing surface was touched, the picture was replaced by the participant’s pen strokes on the monitor, and the RT was recorded. Screen shots of the completed responses were used to score accuracy. After writing their responses, participants returned the pen to the marked starting point and pressed a button with their nondominant hand to advance to the next trial.

Analysis

The data were analyzed as in Experiment 1. To further compare the interference effects between the written and spoken experiments, the data from Experiments 1 and 2 were entered into the same model, with Experiment as a between-subjects factor.

Results and discussion

Of the 9,792 total trials, 4% were removed due to technical errors, incorrect responses, or outlier status. The analysis revealed a significant effect of block type (Fig. 2a), whereby participants were slower to initiate production of items in homogeneous than in heterogeneous blocks, F 1(1, 33) = 4.96, p = .033, mean difference = 12.0, SE = 5.37, 95% CI = 1.0–22.9, η p 2 = .13; F 2(1, 35) = 5.98, p = .020, mean difference = 11.6, SE = 4.76, 95% CI = 2.0–21.3, η p 2 = .15. We observed no significant main effect of cycle, F 1(2, 66) = 1.25, p = .29, η p 2 = .04; F 2(2, 70) = 1.99, p = .15, η p 2 = .05, nor an interaction of cycle and block type, F 1(2, 66) = 0.46, p = .63, η p 2 = .01; F 2(2, 70) = 0.58, p = .56, η p 2 = .02. As in Experiment 1, these results point to interference induced by distributed segmental overlap. In the model with data sets from both experiments included, the main effect of block type remained significant, F 1(1, 56) = 17.51, p < .001, mean difference = 14.5, SE = 3.46, 95% CI = 7.5–21.4, η p 2 = .24; F 2(1, 70) = 25.02, p < .001, mean difference = 15.0, SE = 3.01, 95% CI = 9.0–21.0, η p 2 = .26. However, there the interaction between experiment type (written vs. spoken) and block type (heterogeneous vs. homogeneous) was not significant, F 1(1, 56) = 0.53, p = .47, η p 2 = .01; F 2(1, 70) = 1.27, p = .26, η p 2 = .02, indicating no reliable differences in interference found in the spoken and written modalities.

Fig. 2
figure 2

Results of Experiment 2: Effects of distributed segmental overlap on response times (RTs) for written picture naming of words in blocks with high segmental overlap (homogeneous) versus low segmental overlap (heterogeneous). Error bars represent the between-subjects standard errors of the means, corrected for repeated measures using the Cousineau (2005) method. Panel A includes the data from all items. Panel B depicts the secondary analysis that included only items that shared their initial segment with at least half of the other items in the homogeneous block

Full size image

As in Experiment 1, we found a numerical interference effect when only items sharing their initial segment with at least half of the items in their homogeneous block were analyzed (Fig. 2b). Although this effect did not reach statistical significance, F(1, 33) = 1.65, p = .21, η p 2 = .05, the effect was not reliably different from that of Experiment 1, indicated by the significant effect of block type, F(1, 56) = 8.43, p = .01, mean difference = 13.1, SE = 4.52, 95% CI = 4.1–22.2, η p 2 = .13, but the nonsignificant interaction of experiment and block type in the analysis of the combined data, F(1, 56) = 0.77, p = .38, η p 2 = .01. In summary, we found comparable robust interference in both modalities for segmentally overlapping words.

General discussion

Using a novel manipulation in which segmental overlap was distributed unpredictably across word positions, we observed interference in spoken and written word production, even when considering only items that shared their initial segment with half of the items in their homogeneous block. Critically, this interference that was observed for items with distributed segmental overlap mirrors the interference previously observed in lexical selection for items with semantic overlap, but not the facilitation found when picture names predictably shared onset segments. The effect was replicated across modalities and was not reliably different between the two, increasing confidence in the stability of the effects across considerable variability in task conditions.

Implications for theories of word production

The results of these experiments have several implications for theories of word production. First, we found evidence that similarity-based interference occurs at both stages of word production. In general, distributed feature overlap creates interference during repeated retrieval, regardless of the nature of the overlap (semantic or segmental), the modality of production (spoken or written), or the locus of selection (lexical items or segments). Although our predictions were framed using Oppenheim et al.’s (2010) model, our data are equally consistent with an incremental learning account that relies on lateral inhibition rather than competitive learning (Howard et al., 2006). Although we believe the interference reflects a similar computation principle, we did not expect the resulting effects to have identical properties at the two stages. For instance, the interference generated during lexical–segmental mapping might be more susceptible to the presence of intervening items than the interference generated during semantic–lexical mapping, which typically survives lags of 10+ items (Schnur, 2014). In future work, it will be important to investigate potential differences in order to more fully characterize the mechanisms at the two stages.

Second, these results also support the claim that the facilitation effects reported for initial-segment overlap arise at least in part outside the word production system, since they disappear when predictability is eliminated. Note that we do not rule out that facilitatory effects of similarity might also arise within the production system itself and be masked by the stronger interference effects. This point underscores that it is important to consider that facilitatory and inhibitory effects coexist in the word production system, and performance reflects the sum of these opposing forces. This interplay is affected by the task, such that semantic similarity typically creates interference when related pictures are named repeatedly (e.g., Damian et al., 2001), but facilitation with the presentation of a single semantically related word (e.g., Wheeldon & Monsell, 1994). Phonological overlap can, similarly, have both facilitatory and inhibitory effects, and phonological neighbors can also induce facilitation or inhibition, depending on how strongly activated they are, which can be task-dependent (Sadat et al., 2014; see also Chen & Mirman, 2012). Furthermore, facilitation and interference may be observed even within the same task: In the blocked cyclic naming paradigm, a large initial facilitation often occurs due to repetition (see, e.g., the RT drop between Cycles 1 and 2 in Exps. 1 and 2), before the interference becomes visible in later cycles. The critical claim of the present work is that the competitive incremental learning guiding both facilitation and interference is similar between semantic–lexical and lexical–segmental mapping.

Third, within the frameworks that we have considered, there must be feedback between segments and lexical representations for shared activation to affect the lexical–segmental mapping (see Dell, Nozari, & Oppenheim 2014; Rapp & Goldrick, 2000). Without feedback, there is no reason to expect shared activation for targets and form-related competitors, and therefore no reason to expect interference from high segmental overlap. This creates a challenge for strictly feedforward models (e.g., Levelt, Roelofs, & Meyer, 1999), in which segmental overlap effects could not arise within the production system itself, but would instead need to be explained in terms of the operation of the monitoring system.

In sum, these findings provide evidence for the generality of the incremental learning mechanisms that apply across semantic and form-based levels of representation, giving rise to the complex patterns of facilitation and interference that we observed in spoken and written word production.