Coalescence and contraction of V-to-V_inf sequences in American English – Evidence from spoken language

2.1 The V-to-V_inf construction

The items in question all instantiate a constructional pattern ‘V-to-V_inf’, i. e. a finite verb that subcategorizes a to-infinitive. ^[2] The inflected verb here takes on a modal-like function, modifying the status of the action or state expressed by the to-infinitive verb (cf. Palmer 2001: 1), as in (1a). The verbs that are possible in this slot are restricted by their compatibility with this function (or propensity to be coerced into it), such that the examples in (1b) seem illicit. Due to this function, some typical instances of this construction have been called ‘semi-modals’ (e. g. Biber et al. 1999) or ‘quasi-modals’ (e. g. Hopper and Traugott 1993) – therefore, we will also call this the ‘semi-modal construction’.

There are also other, very similar structures with to-infinitive complements, such as passives (be supposed/believed/known to) and motion verbs (as in I’ll run to buy milk). The constructional relations of all these types are not our main concern here, but we note that a feature of the ‘semi-modal construction’ appears to be that to is attached to the finite verb rather than the following infinitive and thus can be ‘stranded’ (compare [2a] and [2b]).

Some types of V-to-V_inf are idiomatic outcomes of grammaticalization processes (going to, have (got) to, used to) while in others the modality they express derives directly from the core meaning of the verb in question (e. g. try to, forget to). However, in all these instances the component parts of the construction are transparent (i. e. inflected verb, to, infinitive verb). As long as this is the case, all these types should be treated as instantiations of the semi-modal construction, despite the differences between them in frequency and idiomaticity. These differences are part of our investigation, as will be seen.

Our second assumption is that the more frequently a verb occurs in this construction, the more likely it is that ‘V to’ is not treated as a sequence of two separate items, but as a single chunk (cf. Krug 1998; Blumenthal-Dramé 2012: 104), and is then also more likely to undergo phonological reduction. This process of chunking and reduction is also the origin of the forms of to-contraction that have become conventional items (e. g. gonna, wanna; see Figure 1 above). The cases investigated in the present study – have to, used to, trying to, need to – are of interest in this respect, because they fall somewhere in the middle. They are arguably entrenched, show considerable variation in pronunciation, but it is not immediately clear that there is one typical reduced realization that is becoming an independent item. Thus, we can study these items at a stage of variation that may lead to further change. The relevant concepts of coalescence, reduction and contraction are discussed from a cognitive and variationist perspective, as these lead to the central questions of our analysis.

2.2 Coalescence, reduction, contraction

‘Coalescence’ is at times used as a cover term to include related phenomena such as chunking, contraction and cliticization (e. g. Krug 1998). We employ a more restrictive definition, seeing coalescence only as the manifestation of chunking, i. e. that a sequence is represented and produced as a single unit. Thus, coalescence is not necessarily tied to a change in form. Reduction phenomena may then follow from increasing coalescence: “In the course of this process, the elements of the string may lose their independence, boundaries are blurred, and the whole chunk is compressed and reduced” (Diessel 2007: 116).

Reduction in general can be an outcome of automization, which is a function of frequency (Bybee 2006; Diessel 2007), but is also linked to aspects of intonation and prosody, such as speech rate. Articulatory reduction, as well as phonetic assimilation, is a matter of economy, as speech gestures are carried out with less effort (‘hypo-articulation’ in the sense of Lindblom 1990). If we (artificially) separate frequency from speech rate, we may say that both affect pronunciation, but for different reasons. High frequency promotes reduction because of the item’s entrenchment and automization in the speaker’s mind and its easier retrievability on the part of the listener. Rapid speech promotes reduction because it requires articulatory economy. Yet, even in rapid speech, reduction is selective and sensitive to frequency. Greenberg and Fosler-Lussier (2000) report that rapid speech reduction affects high-frequency words more than low-frequency ones.

Regarding reduced forms, we will make a (somewhat stipulated) distinction between articulatory reduction and reduced pronunciation variants. By ‘articulatory reduction’ we mean any on-line mechanical simplification of the articulatory gesture; by ‘reduced pronunciation variant’ we mean a phonetic form stored in the speaker’s memory and available as a target for pronunciation. Presumably, such reduced pronunciation variants emerge out of repeated articulatory reduction (Bybee 2010, 2013) and may gain independent status if their relative frequency increases (Lorenz 2013a: 232). A contraction then, is a reduced pronunciation variant of a coalesced sequence in which the elements of the sequence are no longer clearly separable (as in want to > wanna). Broadbent and Sifaki (2013) identify /t/-lenition as the defining feature of to-contraction. This works well in cases where /t/ follows a vowel (e. g. got to) or an alveolar nasal (want to, goin’ to), that is, in a position that favors reduction to a flap, once the word boundaries have been weakened through frequent occurrence of the string (cf. Gregory et al. 1999). However, in other frequent V-to-V_inf strings, such as have to and used to, a lenition (or elision) of the /t/ is less easily achieved due to the preceding fricative, which is not a typical ‘flapping environment’. Yet, coalescence and reduction are to be expected in these items due to their frequency. We address this issue from a constructional perspective investigating four frequent types of V-to-V_inf: What realization variants do we find in speech? What factors affect the reduction/contraction of these items? Do reduced forms of different types show a common pattern of to-contraction?

2.3 Realization variants

The study presented here analyses the following V-to-V_inf strings: have/has to, used to, trying to, need to. The coalescence of a string can be evidenced by its pronunciation. In need to and trying to this evidence would be a lenition of the medial /t/ (or omission in “tryna”). ^[3] This /t/-lenition is to be understood as any reduction of the articulatory gesture in the voiceless alveolar stop, usually to a voiced sound or a flap. The variants tryna and needa have at times been mentioned as instances of to-contraction (Andrews 1978; Krug 2000: 211). For have/has to and used to coalescence can, in principle, have two different effects. Speakers can either devoice the fricative in assimilation to the following /t/, rendering the type /hæftə/ (“hafta”) or the initial /t/ of to can undergo lenition or complete elision. We hold that /t/-lenition is the stronger indicator of coalescence, firstly because it is less likely in a separated to, whereas devoicing of a fricative tends to also occur word-finally (rendering /hæf/ for have); and secondly because it parallels the precedent cases of gonna, gotta, wanna. We will thus follow Broadbent and Sifaki (2013) in defining to-contraction by /t/-lenition. The option of having both (voiceless fricative and reduced or omitted /t/) is also conceivable, rendering forms of the type /hæfə/ (“haffa”). This form could occur as a reduced form of “hafta”, perhaps in analogy to wanna (and in some ways similar to often). Finally, it should be noted that coalescence need not be reflected in pronunciation at all. A perfect citation-form realization of have + to can nevertheless be represented as a single unit in the speaker’s mind. However, across individual instances and speakers, coalescence and contraction are gradient, and detailed quantitative data can evidence to what degree they are actualized in the language at large. Therefore, in this paper we explore what realizations occur in American English corpus data and what the conditions of their occurrence reveal about potential variant representations.

3.1 Speech rate

Speech rate is an important factor with respect to reduced forms. We consider that a form that is produced only or predominantly in rapid speech is an outcome of articulatory reduction, and not the target of the planned production of a stored variant. Two measures of speech rate are employed: item duration and environment speech rate. The first is on the target items themselves, that is, the duration of the sequence have/used/trying/need to. This measure holds the danger of circularity, in that an item may be reduced because it is uttered rapidly, or an item may have shorter duration because it is phonologically reduced. Therefore, we will present the durational differences between the variant forms but use the second measure in the further analysis. This second measure is the speech rate on the environment, that is, the conversational turn in which the item occurs, excluding the item itself. Thus, the specific phonological form of the item is irrelevant to this measure. Our basic assumption is that rapid speech environments promote articulatory reduction, whereas the repeated occurrence of a reduced form in slow speech indicates that this form is more firmly represented in the speaker’s mind. In our analysis we employ the measure of syllables per second (syll/sec) ^[6] and use logarithmized values in the regression models. The mean environment speech rates are 5.98 syll/sec in the have to/used to data set, and 5.56 syll/sec in trying to/need to.

3.2 Stress accent

Another prosodic factor of reduction is the degree of accentuation (stress) a syllable receives. Reduction typically occurs on less accented syllables (Greenberg et al. 2002; Shockey 2003: 22; Raymond et al. 2006). In V-to-V_inf constructions, to is always unaccented but may form a unit with the preceding verb (which in turn always receives at least a light accent). Stress accent on this verb is therefore a potentially relevant factor of the realization of the unit (consider We HAVE to go versus We have to GO). Stress accent is gradient, but cannot easily be quantified since it is based on a combination of acoustic cues such as pitch, duration and amplitude (Greenberg et al. 2002: 351; Wichmann 2011: 334). Largely following Greenberg et al.’s (2002) procedure, we have labeled the stress accent by two categories, ‘heavy’ and ‘light’. All tokens were annotated by both researchers separately (initial agreement rate: 75.2 %), and unclear cases then revisited. Overall, ‘heavy’ and ‘light’ accents are almost evenly distributed (56 % ‘heavy’ in have to/used to, 50 % in trying to/need to).

3.3 Following sound

Another speech-related factor is the following sound, that is, the first sound of the verb (or adverb) directly following to. In the V-to-V_inf construction the particle to is originally an infinitive marker linked to the following verb; to-contraction then involves the fusion of to with the preceding item and thus its detachment from the following verb. The effect of the following sound can indicate to what extent the realization of to is phonetically conditioned. As this phonetic conditioning is largely a matter of articulatory ease, stored variants in which to is fused to the preceding verb will be less affected by it. Moreover, a following vowel or pause has been associated with fuller forms (Fox Tree and Clark 1997) – specifically, the function word to has shown less reduction of /ʊ/ to schwa before vowels (Jurafsky et al. 1998), and word-internal /t/-deletion is rare before pauses (Raymond et al. 2006: 71). We categorize the sounds by place of articulation, using four categories: labial/dental (comprising bilabial, labiodental and interdental), alveolar, velar and ‘other’ (comprising /h/, vowels and pauses).

3.4 Speech situation

It is known that reduction is often a phenomenon of informal language, whereas formal registers favor full citation forms. Hollmann and Siewierska (2011: 49) point out “the possibility that certain reduced variants index social identities, and that this may have an impact on the facts of reduction in a given variety”. Even the conventionalized to-contractions gonna, gotta and wanna, while being the default option in spoken language, still carry a connotation of informality and colloquialness (Boas 2004; Lorenz 2013a: 184). We take up the formality of the speech situation as a factor in order to assess if such an informality feature is assigned to any specific variant of the V-to-V_inf types investigated here. Based on the file descriptions in the SBC documentation, we defined three categories of speech situation, ‘private’, ‘professional’ and ‘public’. ‘Private’ comprises conversations (usually face-to-face) among friends or family; conversations or task-related talk in a professional setting (e. g. among colleagues or business partners) were labeled as ‘professional’; the category ‘public’ contains organized public talks or discussions. Out of the 60 recordings in the corpus, 34 fall into the category ‘private’, 13 are ‘professional’ and 13 ‘public’.

3.5 Speaker’s age

In order to account for possible changes in the variations, we take up the speaker’s age as a measure of apparent time. If realization preferences change over the generations, this would indicate that a form is emerging as a stored pronunciation variant. The Santa Barbara Corpus provides the speakers’ age and the dates of the recordings. From this information we derived the year of birth for each speaker. The range is from 1903 to 1984, the mean birth year is 1955. For 15 speakers in our data set, age information is not provided.

4.1 Results for the types have/used to

We now take a closer look at the factors of variation for each of the three realization variables defined above. The following analyses of have to are based on 350 tokens, namely the ones that provide at least two syllables of context and the age of the speaker. Differences between the types (have to vs used to) are considered by including the univariate tests on used to.

In our analysis of have to, logistic regression modeling is applied for the purpose of effect estimation (cf. Harrell 2015: 98–99), that is, for testing the effects of a pre-defined set of factors (as outlined in Section 3), rather than attempting to ‘predict’ the variation as a whole. The factors were tested for interaction effects and relevant interactions taken up in the models. For used to, a meaningful multifactorial modeling is not feasible with the available data, firstly due to the small overall token number, and secondly due to the scarceness of some variant forms (e. g. a voiced fricative occurs only 8 times); we therefore revert to univariate statistical tests for each factor in turn. ^[7]

Tables 3 to 5 provide an overview of the three logistic regression models for have to hand in hand with the univariate screenings for used to. These tables deal, respectively, with fricative devoicing, /t/-lenition and final vowel centering. The logistic regression models are summarized by means of the coefficients, standard errors and p-values with significance rating. ^[8] We discuss the results for each variable in turn.

Beginning with the fricative variable (voiced /v/, /z/ vs devoiced /f/, /s/), the model for have to overall is significantly better than chance, but not very strong in predictive power (p(χ ²) < 0.001, C=0.729, D _xy=0.458). The fricative has a slight tendency to be devoiced (in 53 % of cases), as the intercept reflects. Of the factors considered here, speech rate and an interaction of stress accent and speaker’s age significantly determine the variation for have to. Fricative devoicing is less likely in rapid speech, thus assimilation of the fricative predominantly occurs in slow speech. It is also more likely when a heavy stress accent is on have; this effect, however, declines with increasing year of birth, that is, it holds for older speakers but not for younger ones. It should be noted that overall, there is no change in the frequency of fricative devoicing across generations. Devoicing also does not appear to be conditioned by speech situation or the sound following have to.

In used to, fricative devoicing is much more common (89 %). This leaves only eight tokens with a voiced /z/. None of the effects found for have to can be confirmed on this data.

The model for /t/-lenition in have to is shown in Table 4 (p(χ ²) < 0.001, C=0.757, D _xy=0.514). Overall, /t/-lenition is relatively rare in have to (13 %), hence the strong negative intercept. The significant main effects in the model are for stress accent, ‘other’ following sounds and ‘professional’ speech situations. Thus, /t/-lenition is less likely when have is heavily stressed and when have to occurs before pauses or vowels. It is also dispreferred in the more formal, careful speech of professional settings (which, somewhat surprisingly, does not extend to public speech). Speech rate exerts an influence in interaction with stress accent: rapid speech favors /t/-lenition in stressed items but has no influence when stress accent is light. Lastly, there is no indication of ongoing change (factor ‘year of birth’) for have to.

In the case of used to, only speech rate and stress accent have a clear effect on the form of the /t/ sound: /t/-lenition is more likely in rapid speech and when stress accent on used is light.

The statistical model for reduction to schwa of the final vowel in have to is weak (p(χ ²)=0.017, C=0.631, D _xy=0.261) and the only, but very strong effect is that of following labial/dental sounds, which disfavor vowel reduction. This confirms the impression that the realization of the vowel in to is quite variable and rather gradient. The effect of following labial/dental sounds is perhaps really an assimilation effect. In strings like have to work, anticipation of the lip rounding in /w/ can render the vowel as [ʊ] rather than [ə]. Vowel realization is not conditioned by any of the other factors in the case of have to. Used to behaves differently as regards the reduction of the final vowel, showing significant effects of stress accent and year of birth. Thus, schwa reduction is avoided when used is heavily stressed, and younger speakers are more likely to produce a final schwa in used to.

Returning to the variants listed in Table 2 above, we can see how these are aligned with respect to environment speech rate and stress accent. Figures 4 (speech rate) and 5 (stress accent) show that variants with a devoiced fricative are aligned with slow speech and a heavy stress accent; in used to, stressed items also tend to have a final [ʊ]. The forms with /t/-lenition largely occur in rapid speech and are more frequent where the stress accent is light.

Figure 4:

Environment speech rate of realization variants of have to and used to.

While /t/-lenition usually comes with a voiced fricative, there are also the few cases of a devoiced fricative and lenited /t/ (see Table 1; predominantly the form “uSeda”). These 13 items show a mean environment speech rate of 6.96 syll/sec, a clear increase from the grand average of 5.98 syll/sec, and even from that of the ‘/t/-lenition’ group (6.44 syll/sec). The differences shown in Figure 4 suggest that any form with a voiced fricative (including the citation form [v]-[t]) is already a product of accelerated speech. It appears, then, that very high speech rates can produce /t/-lenition in spite of a devoiced fricative and in spite of heavy stress accent (recall the interaction effect on have to in Table 4 above).

4.2 The types trying to/need to in comparison

For the realizations of trying to and need to, the question of fricative devoicing obviously does not apply, but the situation regarding /t/ and the final vowel is more complex (viz. the frequent form “trying [ɾʊ]” and frequent /t/-elision in “tryna” which has no equivalent in need to, see Table 1). The token counts are clearly much lower than for have to. In order to compare the cases, we will therefore present univariate tests of those factors that have shown relevant effects on have to and used to.

The types trying to and need to present a useful case for comparison, since they provide the environment typical of /t/-flapping in American English (preceding vowel or nasal). Consequently, the rate of /t/-lenition is high, at least in trying to (60.4 %). A reduced /t/ in need to (“needa”) is rarer, as is /t/-elision in trying to (“tryna”). These forms have very similar relative frequencies (“needa” 21.9 %, “tryna” 23.6 %).

In the factor speech rate we see another effect of the ‘/t/-flapping environment’ in trying/need to as opposed to have/used to. There is no statistically significant effect of speech rate on either /t/-lenition or vowel reduction in trying to or need to. ^[9] There is, however, a trend for the most reduced variants (“needa”, “tryna”) to occur in faster speech (see Figure 6).

Figure 5:

Stress accent and realization variants of have to and used to.

Figure 6:

Environment speech rate of realization variants of trying to and need to.

For the following sounds, the data on trying/need to is too sparse to usefully account for the separate categories. We only present a comparison of realizations before pauses or disruptions, which have been shown to disfavor /t/-lenition in have/used to above. Here we can observe a similar trend with trying/need to, albeit on a low level, as there are only a few tokens of this case. As Figure 7 shows, /t/-lenition (as well as elision) is extremely rare before speech disruptions in have to, used to (see Table 4 above) and need to, whereas in trying to a following disruption only disfavors elision but not lenition of /t/. That is, the form “tryin’ [ɾʊ]” freely occurs in this context which otherwise disfavors reduction of any kind.

Figure 7:

/t/ realization in have/used to and trying/need to with and without following speech disruption.

Since more formal (in particular ‘professional’) speech situations were found to inhibit /t/-lenition in have to and used to, we also test for this on trying to and need to. The distributions are presented in Figure 8. There appears to be a slight trend for fewer reduced realizations in ‘professional’ and ‘public’ settings, but the differences are not statistically significant in either set (Fisher’s exact test; p(trying to)=0.2619, p(need to)=0.0801). Thus, from what can be gleaned from the data at hand, /t/-lenition in trying to and need to does not generally carry a sense of casualness or informality.

Figure 8:

/t/ realization in trying to and need to by speech situation.

Stress accent also affects the realization of trying to and need to (see Figure 9). The form “tryna” (/t/-elision) makes up 34 % (16/47) of tokens with light stress but only 15 % (9/59) of those with heavy stress accent. The effect of stress accent on realization of trying to is on the threshold of statistical significance (χ ²=5.808, df=2, p=0.055). With need to, the reduced form (“needa”) occurs at a rate of 8 % (3/39) in heavy stress accent, and 33 % (18/36) in light stress. This difference is statistically significant (χ ²=8.012, df=1, p=0.005). Thus, in both trying to and need to heavy stress accent disfavors reduction; again, it is /t/-elision in trying to and /t/-lenition in need to that are subject to this effect, while /t/-lenition in trying to is less affected.

Figure 9:

/t/ realization in trying to and need to by stress accent.

5.1 Pronunciation variants and articulatory reduction

Returning now to the different pronunciation variants of have to, used to, trying to and need to (as listed in Table 2 above), the following picture emerges. We find pronunciation variants that indicate coalescence but are not fully contracted: “haf to”/“hafta”, “uSed to”/“uSed ta”, “tryindu”/“tryinda”. They follow from general phonological processes (assimilation, /t/-lenition) and are highly frequent; judging from our data, they are entrenched in language users’ representations and may be targeted in production. ^[11] We also find reduced forms that qualify as full contractions but that occur as outcomes of on-line articulatory reduction: “havda”/“hava”, “uzda”, “tryna”, “needa”. These forms have a lower degree of entrenchment and are not or rarely targeted in production. We propose the following account of these sets of variants. Language users tend to avoid unnecessary articulatory effort in speech production (cf. e. g. Lindblom 1990); however, this drive for ease of articulation is always balanced against the listener’s need for explicitness (cf. Beckner et al. 2009: 16). Therefore, speakers will initially (and frequently) make articulatory adaptations (assimilation or reduction) that are conservative in terms of transparency. These are adaptations that signal coalescence and reduce the articulatory effort for the speaker but still allow morphological parsing of the structure on behalf of the listener. They do not require the listener to have a single-unit (coalesced) representation of the string, as the individual elements can still be identified. This rings with accounts of complex words in morphology, where “the whole takes precedence over its parts” (Hay and Baayen 2005: 344), but the compositional parts are still activated (Marslen-Wilson 2001). The resulting pronunciation variants are relatively unconstrained in their use, such as “haf to”/“uSed to” (devoiced fricative) and “tryinda” (/t/-lenition in flapping environment). In contrast, reductions that obscure the original morphological structure can only be communicatively successful if the listener has and applies a non-compositional representation and hence does not expect individual morphemes to be discernible. These reductions are (partly) constrained by speech rate, speech fluency, stress accent or speech situation and thus lead to reduced forms that are restricted in use, such as “havda”, “hava”, “uzda”, “tryna”, “needa”. ^[12] Thus, when “boundaries are blurred” (Diessel 2007: 116) in coalescence, it is due not only to frequency but to other reduction-favoring conditions as well.

These reduced forms are contractions, analogous to the well-established gonna, wanna, gotta – the difference is that the latter have become emancipated and have largely outgrown the constraints on their use (Lorenz 2013a). It is conceivable for “havda”/“hava”, “uzda”, “tryna”, “needa” to take a similar path, if they get perpetuated and increase in frequency and entrenchment strength. However, our data show no sign of such developments. Lorenz (2013a: 232–235) proposes a diachronic model of contraction and emancipation, in which reduced forms proceed through the stages ‘on-line phonetic reduction’, ‘on-line morpho-phonological fusion’, ‘stored pronunciation variant’, ‘stored lexical variant’ (see Figure 1 above). In this model, we may place forms like “hava” and “tryna” at the stage of ‘on-line morpho-phonological fusion’, where “due to the chunking of a sequence, a morphologically non-transparent form becomes available, but is not yet conventional at this stage” (Lorenz 2013a: 234), and verging on ‘stored pronunciation variant’, where language users have a stored representation of the form.

5.2 Analogy and the network of variants

We asked if there was a common pattern of reduction and contraction across these four types of the V-to-V_inf construction. Hollmann and Siewierska (2011), for instance, find a common pattern of reduction of the definite article following prepositions, which leads them to suggest a constructional schema that licenses these reductions. In the present case, the different phonological conditions have evidently led to different strategies of reduction or assimilation. Thus, we do not find a common pattern that would allow us to posit a single reduction schema for these items or the V-to-V_inf construction in general. But there are commonalities, which we think should be described in terms of analogy. Both have to and used to show frequent fricative devoicing, and /t/-lenition at a low rate. It seems plausible that speakers use these forms in analogy, that is, each use of a given variant form of one type (e. g. “hafta”) also strengthens the representation of the corresponding variant of the other type (e. g. “uSedta”). If this is correct, then the strengthening of representations could also increase the likelihood of the corresponding forms of less frequent types by analogy (e. g. “lof ta” for love to). ^[13] This is outside the scope of the present study, and to confirm (or reject) these conjectures, more evidence would clearly be needed, as well as a better understanding of how pronunciation variants of different items are interconnected.

Another analogy worth considering is with the conventionalized to-contractions gonna, wanna and gotta. Here, “tryna” and “needa” seem to profit from the analogy. These forms show somewhat higher relative frequencies than “havda” and “uzda”; their occurrence is probably aided by their phonological properties allowing for a smooth transition from [t] to flap to zero. The analogous forms of have to would be “hava” or “haffa” – yet, the former is rare and the latter inexistent. It seems that /t/ is not easily dropped without the intermediate stage of lenition, which is inhibited in these items. The case of have to shows that this phonological hurdle can be overcome when high frequency and speech-internal factors (light stress, rapid speech) coincide; used to, being less frequent, is one step behind, such that /t/-lenition is possible, but /t/-elision is strongly avoided.

In Figure 10 we summarize the above findings in a sketch of a hypothesized network of constructional types and pronunciation variants, as derived from the results of our study. It incorporates notions from Construction Grammar, in particular that have to/used to/need to/trying to/etc. are all types of one construction (the V-to-V_inf or semi-modal construction), and connected by instantiation links. On the level of instances, it presents the attested pronunciation variants and analogy links between them; the empirical findings of the present study inform this level, which is the crucial part of the network. The items and links are presented by the strength of their mental representation. We follow the convention of using dotted arrows and rounded boxes for weakly entrenched relations and items, while solid arrows and angular boxes stand for firm entrenchment (cf. Langacker 1987; Hollmann and Siewierska 2011: 43). ^[14] The most frequent realization of each type is in bold face. Forms that are only distinguished by the final vowel are represented as co-variants (sharing the same ‘box’), based on our finding that the variation in the final vowel is phonetically conditioned rather than a matter of variant choice; the form “hava” is seen as a further reduction and thus a sub-variant of “havda”. Curved arrows are used to indicate analogy (and dotted curved arrows for weak analogy). The analogy relations reflect the main results of this study: that have to and used to cluster together, that the devoiced forms “hafta” and “uSedta” are entrenched variants, while “havda” and “uzda” are tied to on-line reduction, and that the conventional contractions gonna/wanna/gotta likely exert a greater influence on trying to and need to than on have to and used to due to their phonological similarities. Thus, there is a frequent and entrenched reduced variant “tryinda”; the condensed forms “tryna” and “needa” show similar usage conditions as “havda” and “uzda”, and are therefore also classified as ‘weakly entrenched’. It should be noted that this network model is an attempt at pulling together the results of a detailed corpus study and visualizing their interpretation with regard to the mental representation of variants; it involves some necessary simplification (for example, entrenchment is really gradient, not a matter of weak or full representation) as well as conjecture (e. g. on the influence exerted by gonna/wanna/gotta). The network model thus constitutes a hypothesis which may serve as a basis for future research on similar phenomena. Its merit, apart from providing a visual summary, is in proposing relations between variants of different items, which is something that has hitherto received little coverage in usage-based studies.

Figure 10:

A hypothesized network of construction types and pronunciation variants.