3.1 Findings on Practice Effects: those Aspects of Speech Production Planning which are Affected by Practice
The null hypothesis as stated in 1.3 that there are no practice effects on SOT types has been rejected with respect to various aspects of speech production planning. As shown in 2.3, the SOTs in everyday conversation and those in TV programs clearly demonstrate differing distributions. Under the premise that the differences reflect different characteristics in planning, the results reveal various practice effects on speech production planning.
The practice effects that have been found by the present study are summarized below.
(24) a. Preplanned speech contained a higher frequency of PHONOLOGICAL and SYNTACTIC errors and a lower frequency of LEXICAL errors than
spontaneous speech.
b. Preplanned speech contained a higher frequency of anticipatory and a lower frequency of perseveratory PHONOLOGICAL errors and SYNTACTIC
lexical unit errors than spontaneous speech.
c. Preplanned speech contained a higher frequency of PHONOLOGICAL omission errors, especially telescopings, than spontaneous speech.
d. Preplanned speech contained a lower frequency of lexical unit errors involving OCs and a higher frequency of lexical unit errors involving CCs than
spontaneous speech.
e. Preplanned speech contained lower frequencies of lexical blends and phrase blends than spontaneous speech.
f. Preplanned speech contained a higher frequency of contextual lexical unit errors and a lower frequency of non-contextual lexical unit errors than
spontaneous speech.
g. Preplanned speech contained a higher frequency of PHONOLOGICAL errors with a greater anticipatory error-source distance and a lower frequency of
lexical unit errors with a greater perseveratory error-source distance than spontaneous speech.
h. Preplanned speech contained a lower frequency of errors involving verbs and adjectives than spontaneous speech.
i. Preplanned speech contained a lower frequency of consonant errors and a higher frequency of vowel errors than spontaneous speech.
A word of caution regarding the claim about “effect” is necessary here. As shown in (24), some types of errors are more frequent and others are less frequent in preplanned speech than in spontaneous speech. However, as discussed in 1.2.3, other studies have shown that at least task-specific practice reduces the overall frequency of errors in general. If this applies to content-practice, it might be incorrect to claim that practice increases or decreases particular types of errors. Thus, all one can claim is that practice produced different relative frequencies of various error types.[28] Thus, the differences in frequencies of different types of errors between the two settings are “effects,” because practice changes the way that one makes SOTs. If the relationship between a particular kind of error and other kinds of errors under the same category in the context of speech production planning is taken into account, one can infer what the effect is from this relationship to a considerable degree.
I am now in a position to inquire into each practice effect in detail. The first practice effect concerns at what stage of production planning errors occur. As most speech production planning models maintain, there are different stages of planning processes, which lead from
lexical selection, to syntactic structuring, and then to sound form generation. In 2.1.5.1, SOTs were classified into three broad TYPES, LEXICAL, SYNTACTIC, and PHONOLOGICAL errors, which occur at these stages. With practice, it appears that the earlier stages become more error-free and the later stages become a more prominent locus of errors. In preplanned speech, because the content is pre-determined, the speaker may have already processed some aspects of lexical selection and knows what lexical items to use. This leads to a lower frequency of LEXICAL errors and a higher frequency of PHONOLOGICAL and SYNTACTIC errors in preplanned speech than in spontaneous speech (Table [1]). Even though lexical unit errors occur with about the same frequency in spontaneous speech and in preplanned speech, they are more likely to occur as LEXICAL errors and less likely to occur as SYNTACTIC errors in
spontaneous speech than in preplanned speech (Table [10-A]). Therefore, it is likely that practice decreases the number of LEXICAL errors, with the result that the frequencies of SYNTACTIC and PHONOLOGICAL errors appear to become higher.
In Dell’s model, “the selection of items for a lower representation must await the construction of corresponding structures of the higher representation” (Dell 1986:287). In preplanned speech, where the speaker has decided what kind of speech act to make and what speech content to express, the highest representation, what Bock (1982) calls the “interfacing representation,” is already fixed in content and tends to be already partially constructed.
Therefore, SOTs are less likely to occur at this level. However, since preplanning of content does not affect the lower representations, SOTs can occur freely at these levels. Similarly, in Levelt’s model, where a processing component feeds information to the next component basically in a unidirectional way, in preplanned speech, information is likely to have been processed at the Conceptualizer and at least in part by the Formulator. If the message is already represented at the level of grammatical encoding with lemmas accessed, the speaker will be less likely to make a SOT at this level. Therefore, SOTs are relatively more frequent at later
processing components in preplanned speech than in spontaneous speech.
MacKay (1982:503), who noticed that concept structuring and lexical selection are more conscious and less automatic than the construction of sound forms, claims through introspection, under the node structure theory, that “concept nodes in general receive less prior practice than phonological nodes.” In spontaneous speech, what are automatic are phonological and syntactic processing, and activating concepts is ongoing focus of attention. However, the current study showed that this is not true of preplanned speech; in preplanned speech, the concepts do receive prior practice and become less available for errors.
Second, practice generally seems to increase anticipations and decrease perseverations (Tables [5-B] and [12-A]).[29] This effect applies both to PHONOLOGICAL errors and SYNTACTIC lexical unit errors. This accords with one of the findings about experimental-practice effects by Schwartz et al. (1994) and Dell et al. (1997). It should be noted that this effect is observed not only in PHONOLOGICAL and SYNTACTIC lexical unit errors in general but also in errors of each kind of unit (e.g., consonants, vowels, pitch accents, OCs, CCs),
although it is sometimes only a numerical trend. The difference in the frequency of anticipations seems to be due to the large number of incomplete anticipations rather than complete
anticipations in preplanned speech compared to spontaneous speech (Note that in Table [12-A]
the percentages of complete anticipatory substitutions in the two settings are about the same).
The predominance of incomplete anticipations (Dell et al. 1997:127) appears to be another kind of practice effect: one could hypothesize that speakers of preplanned speech can easily notice
the error and fix it early before they say its source because they can self-monitor their speech better and more effortlessly.
According to Schwartz et al. (1994), the practice effect of strengthening connections can account for the augmentation of anticipatory errors and the reduction of perseveratory errors in practiced speech in their experiment. Their interpretation of this effect is as follows: in less practiced speech, “the weak connections make the system erroneously stick with what it has recently retrieved. So, when a word is incorrectly encoded, the interference is more likely to come from a previous word than an upcoming word. With more practice, the intended word and upcoming words become stronger competitors with the result that perseverations are less favored errors” (p.72).[30] However, they do not explain exactly why the weak connections lead to more interference from what has been recently retrieved and less interference from what is going to be retrieved and why stronger connections due to practice reverse the tendency. The following proposal may be worth consideration. In spontaneous speech, the connections are weak both in what has been recently retrieved and in what is going to be retrieved. In such speech, the speakers have to pay active attention both to what they have just said and what they are going to say. In short, effort is necessary both for self-monitoring and for planning ahead. In preplanned speech, on the other hand, the connections are strong in both directions. If the practice is
successful, speakers are unlikely to make an error, and do not have to pay as much attention to what they have just said, so that previously activated elements decay quickly. More attention can be directed to forthcoming elements.
Dell et al. (1997) call such an effect the “anticipatory practice effect.” According to them, the speech production system “activate[s] the present, deactivate[s] the past, and prepare[s]
to activate the future” (p.123). Anticipation involves the relative activation of the present and the future, whereas perseveration involves that of the present and the past. Since practice
promotes the activation of the present and the future, anticipations increase with practice. As the name of the effect suggests, they take the view that “[p]ractice does not ... have much effect on the deactivation of the past” (p.132). However, the view would not be able to explain why perseverations are less frequent in preplanned speech than in spontaneous speech in the present data. In fact, the “perseveratory practice effect” is even more significant than the “anticipatory practice effect,” as in Tables [5-B] and [12-A] (PHONOLOGICAL errors: anticipations !2=9.48, p<0.05; perseverations !2=19.90, p<0.01; SYNTACTIC lexical unit errors (OCs and CCs):
Anticipations !2=6.63, p<0.05, Perseverations !2=22.70, p<0.01). Dell et al. argue that there is a
“self-inhibition mechanism,” which turns off the past, and that the “turn-off function is effective ... regardless of how activated the past is” (p.132). The perseveratory practice effect found in the present study, however, suggests that the turn-off function may be facilitated after practice and that practice helps deactivate the past (Dagenbach and Carr 1994).
Levelt (1989) does not talk about anticipations and perseverations in the context of a general cognitive mechanism. Although he does not mention contextual lexical unit errors, he discusses how anticipatory and perseveratory phonological substitutions occur in terms of a selection step and a checkoff step (2.1.6). Anticipatory substitutions are characterized as errors where a wrong item, which should fill a slot later, is selected from the filler candidates. In a complete anticipation, the wrong item is not checked off; the error is due both to misselection and checkoff failure. In an incomplete anticipation, the misselection is noticed by
self-monitoring, which detects that the incorrectly selected item is missing from the filler candidates;
the error is owing only to misselection. Perseveratory substitutions are those errors where an item is not checked off and is selected again for a subsequent slot. In other words, in
anticipatory substitutions, a forthcoming item is activated highly enough to be selected from later in the utterance, and in perseveratory substitutions, an item is activated highly enough not to be checked off and to be selected again later in the utterance. This description can explain why incomplete anticipations are more frequent in preplanned speech than in spontaneous speech. In preplanned speech, the checkoff step is functioning properly with the help of efficient self-monitoring; however, since much attention can be directed to candidates for later slots after practice, they may become activated early and contaminate the current selection.
Third, phonological omissions, including telescopings, were more common in preplanned speech than in spontaneous speech (Table [4]). Lexical omissions were, though only
numerically, more frequent in preplanned speech than in spontaneous speech (Table [11]).[31]
Although neither Dell’s model nor Levelt’s model address the relationship between omissions and practice, one can hypothesize that omissions, at least those with no directionality as well as anticipatory ones, are kinds of anticipations in essence, which are generally more frequent in preplanned speech than in spontaneous speech.
Fourth, the distributions of lexical unit errors involving OCs and those involving CCs vary depending on practice (Tables [8] and [10-B]). I would argue that, in spontaneous speech, since the content to be expressed has not been preprocessed and thus OCs have not been selected in advance, they are prone to errors. In preplanned speech, on the other hand, OCs with which to express the preplanned content have been preselected to a large extent, but CCs, which express structure, are susceptible to errors. The higher frequency of CC errors in preplanned speech than in spontaneous speech is mainly due to the large number of errors involving particles (Table [15]). The lower frequency of OC errors in preplanned speech than in spontaneous speech is mainly due to the relatively small number of errors involving verbs (Table [15]). The difference in the number of verb errors is large in SYNTACTIC lexical unit errors, while errors involving other lexical categories show similar distributions.
This is a practice effect that Schwartz et al. (1994) and Dell et al. (1997) did not find, or rather would not have been able to find. It would be impossible to examine the frequencies of OCs and CCs in experiments that use a limited number of expressions. For example, there were no prepositions or conjunctions used in the tongue twisters in their experiments.
Dell does not explicitly address the effect above, but it seems to be able to account for it.
In his model, connections associated with CCs should be strong due to acquisition-practice.
Hence, generally, errors involving CCs are less common than those involving OCs. Content-practice strengthens connections associated with OCs. Since it is OCs expressing content that are practiced during content-practice, the increase in connection strength is greater with OCs than CCs, whose connections are generally stronger than those of OCs. Therefore, while OC errors are reduced, CCs errors are not.
The model by Bock and Levelt (1994), however, seems to be able to predict this effect.
They divide grammatical encoding (Levelt 1989) into two stages, “functional processing” and
“positional processing” (also, Garrett 1975), and argue that syntactic functions are assigned to selected OC lexical items at “functional assignment,” a component of functional processing, where CCs are selected (pp.960-968). Functional assignment occurs after “lexical selection,”
another component of functional processing, and before “positional processing,” where
constituents are arranged in order. Practice facilitates selection of OCs and reduces the number of errors involving them, while this effect does not apply to CCs.
Fifth, there were fewer lexical blends and phrase blends in preplanned speech than in spontaneous speech (Tables [9], [11]). Although errors involving CCs were generally more
frequent in preplanned speech than in spontaneous speech, lexical blends of CCs did not occur in preplanned speech, while there were five examples in spontaneous speech. Blend errors occur when more than one (normally two) lexical item or phrase that express similar or related concepts compete for a single syntagmatic slot. Two lexical items or phrases can be selected more often in spontaneous speech than in preplanned speech, as practice prevents this from happening.
Dell’s model would predict this effect. In this model, blend errors occur when two nodes are “nearly equally active” and are “selected and tagged for the same position” (Dell 1986:118).
After practice, the connections between elements within the intended utterance are strengthened, but the connections to other semantically related items are not, or may even be weakened.
Therefore, since it is very unlikely that outside elements would become activated to the same degree as the intended ones, blends are much less frequent in preplanned speech than in spontaneous speech. Levelt’s model would also predict this effect. In preplanned speech, concepts to be expressed are already set up and disturbance from related concepts is unlikely to occur. Hence, the frequency of blend due to conceptual intrusion in (23a) is low in preplanned speech.
The sixth finding concerns the contextuality of lexical unit errors. There are more contextual lexical unit errors and fewer non-contextual lexical unit errors in preplanned speech than in spontaneous speech (Table [13]). There seem to be two factors in causing this effect.
One is the large number of SYNTACTIC errors in preplanned speech, which are caused by the preplanning of the content, as discussed for the first finding. The other is the low frequency of paradigmatic lexical substitutions of OCs and that of lexical blends in preplanned speech.
The low frequency of paradigmatic lexical substitutions of OCs in preplanned speech can also be explained by the two models, by means of the same account as the one given of the practice effect on lexical blends. Practice prevents intended OC items from being susceptible to intrusion into from outside of the linguistic context, because of already established concepts or lemmas.
The seventh finding is that the forward planning distance in PHONOLOGICAL errors appears to be greater in preplanned speech than spontaneous speech. A contextual non-pitch-accent PHONOLOGICAL error is more likely to be influenced by a source in a different word in preplanned speech than in spontaneous speech (Table [6-A]). On the other hand, it is difficult to make an argument about PHONOLOGICAL perseverations because they are not very common in preplanned speech, even though within-word perseverations were more common in
preplanned speech than in spontaneous speech.
Dell’s model does not handle the practice effect on forward planning distance in
PHONOLOGICAL errors. However, the proposal made for the practice effect on anticipations and perseverations, which seems to be applicable to the comparable finding by Schwartz et al.
(1994) and Dell et al. (1997), may be apt. In preplanned speech, the connections are strong before and after the ongoing utterance, and more of the speaker’s attention is directed to
forthcoming elements than to past elements. Connections are strengthened farther into elements that occur later. Levelt’s model contrasts automatic and controlled processing (2.1.6).
Automatic processing tends to be parallel and to have low capacity demands imposed on it, whereas controlled processing tends to be serial and to demand attentional resources. Since content-practice leads to automatic processing, such processing enables the speaker to plan longer distances with a greater memory capacity than controlled processing.
SYNTACTIC lexical unit errors show a pattern slightly different from
PHONOLOGICAL errors (Table [12-B-2]). Forward planning distance is not necessarily greater in preplanned speech than in spontaneous speech. A source that has already appeared in the previous discourse is likely to influence the error in spontaneous speech and thus this increases the number of perseverations. Thus, the distance between a perseveratory error and its source tended to be greater in spontaneous speech than in preplanned speech. This tendency can be explained in terms of quick decay of previously activated elements in preplanned speech, as discussed for the perseveratory practice effect. However, a question still remains as to why forward planning distance is not great in preplanned speech compared to spontaneous speech.
This is probably because the speaker can only plan one, or two at most, syntactic units like phrases or clauses at a time; it is impossible to plan farther unless the speaker plan syntactically parallel structures.
Eighth, there are fewer lexical unit errors involving verbs and adjectives in preplanned speech than in spontaneous speech (Table [15]). This effect appeared both among LEXICAL errors and SYNTACTIC errors. The lower frequency of lexical errors involving verbs in preplanned speech suggests that predicates are selected before their arguments are selected and assigned. It seems to support the “centrality of the verb to higher level production processes”
(Bock and Levelt 1994:968), which Bock and Levelt list as one of the possible reasons for the insusceptibility of verbs to substitutions. If this is the case, it will also support the presupposition made by most current syntactic theories that the predicate is the center of the semantic
representation of a clause and specifies its thematic relations, the relationship between the predicate and its argument(s) (Fillmore 1968, Chomsky 1981, Foley and Van Valin 1984, Van Valin 1993, Van Valin and LaPolla 1997 among others). Levelt’s model adopts this view (Levelt 1989:192-193).
The smaller number of errors involving verbs in preplanned speech than in spontaneous speech also seems to be owing to the centrality of the predicate at the SYNTACTIC level. For example, in Role and Reference Grammar, at the syntactic representation, an appropriate syntactic template is selected from the syntactic inventory, under the principle that the “number of syntactic slots for arguments and argument-adjuncts within the core is equal to the number of distinct specified argument positions in the semantic representation of the core” (Van Valin and LaPolla 1997:324).[32] Since the “semantic representation of the core” is determined by the type of predicate, the choice of a syntactic template hinges more on the predicate than its arguments.
Syntactic building procedures in Levelt’s model are also based on a similar principle: “the main verb dictates what arguments have to be checked in the message, and which grammatical
functions will be assigned to them” (p.244).
Dell’s model does not seem to handle this issue at the level of either semantic or syntactic representations, but his model might ascribe the effect above to the centrality of predicates at the level of semantic representation with respect to lexical activation, which in turn leads to the centrality of predicates at the level of syntactic representation.
Why does the centrality of the predicate cause a lower frequency of errors involving verbs and adjectives in preplanned speech than in spontaneous speech? This is partly because the overall proposition in discourse context, which is the main portion of what is planned in preplanned speech, limits the possible choice of predicates, whereas, in spontaneous speech, a speech event where communicative goals are less clear, there may often be more choice about perspectives from which to frame proposition, i.e. different predicates.
Finally, fewer consonant errors and more vowel errors occur in preplanned speech than in spontaneous speech (Table [3]). Speech production models of the connectionist type seem to be able to explain this. Practice reinforces the phonological networks that connect phonological nodes representing consonants on the segmental level with phonological nodes on other levels and causes consonants to be less error-prone. The question is why. The effect seems to support the idea that slots for consonants are determined in syllable templates.[33] Nevertheless, further research is necessary to account for this effect.
As discussed so far, it is only the practice effect on anticipatory and perseveratory errors that accords with a finding by Schwartz et al. (1994) and Dell et al. (1997).[34] Another claim by them is that PHONOLOGICAL errors of a “good” pattern result in actual words in the language.
However, as shown in Table [7], I have obtained results that cast doubt on their claim: the frequency of PHONOLOGICAL errors resulting in words is about the same without respect to practice; in fact, preplanned speech contains a few more PHONOLOGICAL errors resulting in non-words and a slightly fewer PHONOLOGICAL errors resulting in words than spontaneous speech does (not statistically significant).
For the supposed practice effect, Schwartz et al. (1994) state the following.
“... because practice is hypothesized to cause the delivery of more activation to intended word and sound units, there is, consequently, more activation that can participate in positive feedback between words and sounds. Although this extra activation should enhance correct encoding, there may be a relative tendency for more errors that create words over nonwords from the positive feedback” (p.72).
However, it is not clear why such a positive feedback occurs not only between the target word and its sounds but between words and sounds in general. Since the “extra activation should enhance correct encoding,” the connections pertaining to the target are strengthened but those between the target and other phonologically similar words should be weakened, possibly through an inhibitory mechanism.
3.2 Findings on those Aspects of Speech Production Planning which are not Affected by Practice
Despite the practice effects discussed above, there are a number of characteristics of speech production planning that seem to be unsusceptible to practice.
First, phrase unit errors are rare compared to phonological unit and lexical unit errors (Table [2]). Accordingly, lexical substitutions are the most common SYNTACTIC errors. Since phrases are large composites that consist of smaller planning units and are constructed rather than stored and retrieved, they are less likely to serve as units in speech production planning than phonological units and lexical units, regardless of practice.
Second, although PHONOLOGICAL errors involving consonants and those involving vowels show different distributions in spontaneous speech and in preplanned speech (24i),
segments are the most common PHONOLOGICAL error units in both (Table [3]). Regardless of practice, PHONOLOGICAL errors involving larger units and feature errors are rare. As
Kubozono (1985, 1989) argues, there are slightly more mora errors than syllable errors in spontaneous speech, and this is also true of preplanned speech, but it would be difficult to make any argument about this based on my data, because the numbers of these types of errors are too