1. Introduction
English has a class of transitive verb-particle combinations that exhibit word order alternation. In particle verbs like clean out, the particle may remain joined with the verb; the particle verb's object follows, as your gutters follows clean out in (1a).Footnote 1 Alternatively, the particle verb may be split by its object; in (1b), the verb clean and the particle out are linearly separated from each other by the gutters.
The English particle verb alternation is one of a few well-studied examples of word order variation (Robinson and MacKenzie Reference Robinson and MacKenzie2019). Because previous work has established which/how internal linguistic factors matter, this alternation is fertile ground for investigating whether external factors condition syntactic variation. In an influential early variationist study, Kroch and Small (Reference Kroch, Small and Sankoff1978) contend that they do. Specifically, they argue that radio show hosts and in-studio guests show greater sensitivity to prescriptive norms, and that this makes them more likely to use the joined variant of the alternation than listeners calling in to the show. Our study attempts to replicate this finding with a much larger data set and more modern statistical methods, but finds it does not hold up: social roles predicted to be associated with adherence to prescriptive language ideologies do not condition the rate of joined variant use for particle verbs in our data.
2. Background
Kroch and Small (Reference Kroch, Small and Sankoff1978) investigate the role of prescriptive norms in conditioning syntactic variation. They hypothesize that speakers who occupy different social roles will differ in their adherence to the grammatical ideology of the standard language, and that this can be observed in how they produce syntactic structures that vary in their “correspondence between propositional form and surface syntax” (Kroch and Small Reference Kroch, Small and Sankoff1978: 48). One such variable structure is the English verb-particle construction, for which “the surface syntactic configuration V Prt NP would correspond to propositional form more directly than would V NP Prt because only the first reflects in its word order the semantic unity of the verb and particle (compare the split infinitive)” (Kroch and Small Reference Kroch, Small and Sankoff1978: 49). In other words, the prediction is that speakers who are more sensitive to prescriptive norms will be more likely to use what we are calling the joined variant (Kroch and Small's “V Prt NP”).
Kroch and Small test this with a small corpus study of talk radio broadcasts from Philadelphia. To operationalize the factor of adherence to standard language ideology, they divide speakers in their radio data into two sociolinguistic groups: those who have called in to a talk show (“callers”) on one hand, and, on the other, the show hosts and their studio guests (“hosts/guests”). They further assume that “for reasons both of social status and role in the radio talk-show interaction, the host/guest group generally speaks a more standard English than the caller group” (Kroch and Small Reference Kroch, Small and Sankoff1978: 48). They find that this is borne out quantitatively: the host/guest group is found to use the joined variant of the verb-particle alternation at a rate of 63% (N = 167), significantly different from the caller group's 47% rate for the same variant (N = 138) via a chi-squared test (p < 0.01).Footnote 2 Further analysis subsetting the data by linguistic factors (length of object, degree of semantic transparency of verb-particle construction, stress placement) and comparing caller and host/guest rates within each level of the linguistic factor under study generally confirms the basic pattern quantitatively (though in section 5.2 we cast some doubt on these findings). Finally, Kroch and Small report the results of a judgment task carried out by Temple University students (N = 32): 44% of students report that the sentence John called up Mary is “substantially” more correct than the alternant John called Mary up. This further supports the notion that language users are sensitive to prescriptive norms against the split variant of this alternation.
Kroch and Small's findings have been influential. They continue to be cited in discussions of the role of prescriptive norms and standard language ideology in conditioning language variation (Romaine Reference Romaine1981, Irvine Reference Irvine1985, Guy and Bayley Reference Guy and Bayley1995, Johnstone and Bean Reference Johnstone and Bean1997, Cameron Reference Cameron2000, Díaz-Peralta and Almeida Reference Díaz-Peralta and Almeida2000, Bresnan and Ford Reference Bresnan and Ford2010, D'Arcy and Tagliamonte Reference D'Arcy and Tagliamonte2010, Adli Reference Adli2013, Wiechmann and Lohmann Reference Wiechmann and Lohmann2013, Bouchard Reference Bouchard2018), in literature reviews of the English verb-particle alternation (Gorlach Reference Gorlach2004, Bleaman Reference Bleaman2020, Haddican, Johnson, Wallenberg, et al. Reference Haddican, Johnson, Wallenberg, Holmberg, Beaman, Buchstaller, Fox and Walker2020, Röthlisberger and Tagliamonte Reference Röthlisberger and Tagliamonte2020), and in considerations of the question of whether syntactic variation more generally can be sensitive to social factors (Meyerhoff Reference Meyerhoff2000, Röthlisberger and Tagliamonte Reference Röthlisberger and Tagliamonte2020). On this latter point, there is a large literature proposing or asserting that syntactic variation is unlikely to show social conditioning (see Levon and Buchstaller Reference Levon and Buchstaller2015 for a recent review); Kroch and Small's host/guest vs. caller difference is sometimes held up as a counterexample to this (Cheshire Reference Cheshire1987, Meyerhoff Reference Meyerhoff1997, Meyerhoff Reference Meyerhoff2000, Röthlisberger and Tagliamonte Reference Röthlisberger and Tagliamonte2020).
The lasting influence of their findings notwithstanding, Kroch and Small were careful to point out the need to replicate their study with more data and with multivariate analysis (Reference Kroch, Small and Sankoff1978: 51). There has been considerable recent large-scale corpus work on the particle verb alternation which has demonstrated that it has robust social correlates, including register, regional variety, speaker age, and social class (Gries Reference Gries2003; Haddican and Johnson Reference Haddican and Johnson2012; Röthlisberger and Tagliamonte Reference Röthlisberger and Tagliamonte2020). Here, we ask: Can Kroch and Small's host/guest vs. caller difference be replicated on a larger data set using more modern statistical methods?
3. Methods
This study extracted, annotated, and analyzed 10,521 tokens of particle verb constructions.
3.1 The corpus
The data used in this study come from the RadioTalk Corpus (Beeferman et al. Reference Beeferman, Brannon and Roy2019). RadioTalk consists of 2.8 billion words of automatic speech recognition transcripts. The transcribed snippets were sampled from 284,000 hours of talk radio broadcasts that aired in the continental United States between October 2018 and March 2019 and were subsequently made available online. The majority of sampled stations were either news, talk, or public radio stations; as a result, the corpus includes conversational speech both from radio show hosts and from listeners who called in to their shows. Using unsupervised spectral clustering to judge “whether the underlying audio came from a telephone or studio audio equipment,” a process that mostly relied on “the narrow frequency range of telephone audio,” Beeferman et al. estimate that 10.0% of the corpus is telephone speech (Beeferman et al. Reference Beeferman, Brannon and Roy2021).
3.2 Data extraction and coding
We used Google BigQuery Standard SQL to extract instances of 249 transitive verb-particle combinations in either the joined or split orders. Because RadioTalk was automatically transcribed and not checked by human annotators, our query excluded any snippets for which Beeferman et al.'s speech recognizer had less than 98% confidence in its own transcription accuracy. Other incorrect transcriptions whose actual values could not be recovered, as judged during our own data annotation process, were also excluded.
Rate of use of the joined variant is known to increase with the length of the object, becoming nearly categorical with objects above six words in length (Grafmiller and Szmrecsanyi Reference Grafmiller and Szmrecsanyi2018: 389). Reflecting this, our query also excluded any matches where more than four words intervene between the verb and the particle; objects of greater than four words found with particle verbs in the joined order were excluded by hand. We opted for a maximum object length of four words because exploratory queries that expanded the window between verb and particle any further resulted in an excessively high proportion of false string matches relative to legitimate split particle verb tokens (e.g., eating clean is eating less junk and working out is being less sedentary). False string matches of this type were removed by hand.
Pronominal objects are nearly categorical in requiring the split order, in both our data (n = 5339/5350, 99.8%) and previous work (Grafmiller and Szmrecsanyi Reference Grafmiller and Szmrecsanyi2018: 389). When the particle in a particle verb is modified by elements like right, straight, or back, the split order is again highly preferred (n = 30/32, 93.8%, including only non-pronominal objects). Fronted objects, as in passives or relative clauses, are also non-alternating. We excluded all such non-alternating items by hand.
3.3 Coding for conditioning variables
Following these exclusions, the dataset contained 10,521 tokens of particle verb constructions. Each token was annotated for our critical variable of speaker modality, and for five control variables that previous work or our own explanatory data analysis found to affect word order.
Speaker modality: The core finding of Kroch and Small (Reference Kroch, Small and Sankoff1978) that we hope to test is the effect of the social roles occupied by various talkers on radio show broadcasts. Radio show hosts and in-studio guests are hypothesized to adhere more closely to prescriptive language ideologies — use more joined variants — than listeners calling in to the show. This hypothesized difference could stem from differences in the groups’ ranges of occupations and their attendant participation in the linguistic market (Sankoff and Laberge Reference Sankoff, Laberge and Sankoff1978), as well as from the asymmetric power dynamic of the host-caller interaction itself.
Audio from listeners calling in to radio shows via telephone is limited to narrowband frequencies (300–3400 Hz), whereas audio recorded in the studio (i.e., radio show hosts and in-person guests) is not subject to that limitation. Unsupervised spectral clustering software made use of this difference to automatically supply “a flag for whether a given utterance was recorded in a studio or came from a telephone call-in” (Beeferman et al. Reference Beeferman, Brannon and Roy2019: 2).
Length of the direct object: Increasing the length of the object has been found to increase the likelihood of using the joined verb-particle order, as mentioned in section 3.2. Longer, prosodically “heavier” objects, which may also be more complex (see Wasow Reference Wasow1997, Wasow and Arnold Reference Wasow, Arnold, Rohdenburg and Mondorf2011 for discussion of endweight), tend to occur at the ends of clauses. We operationalize object length as number of orthographic words (following Gries Reference Gries2003, Grafmiller and Szmrecsanyi Reference Grafmiller and Szmrecsanyi2018, see also Röthlisberger and Tagliamonte Reference Röthlisberger and Tagliamonte2020), ranging from one to four.
NPR show as a proxy for written register: Some previous studies on British English have shown an effect of written vs. spoken language, where written particle verb constructions are more likely to be found in the joined order (Gries Reference Gries2003, Cappelle Reference Cappelle2005). As a proxy for written register, we search for the string “NPR” in the name of the show associated with each token. NPR shows like Morning Edition often include hosts reading prepared or scripted news briefs, which we take to be more like written language. We therefore expect particle verbs coded as coming from an NPR show to be more likely to be joined.
Meaning frequency: Particle-verb collocations are highly polysemous. Each token was coded for its specific meaning, as demonstrated in example (2) for meanings of knock off.
Since measures of particle verb frequency from other corpora are not specified for these meanings, we operationalized meaning frequency using our own data set. We counted all occurrences of each meaning, including in non-alternating contexts (e.g., the piece of Mars that the asteroid knocked off), and log10-transformed the resulting meaning counts to arrive at a measure of frequency.
Semantic compositionality: English particle verbs’ meanings range from entirely predictable from their component pieces’ literal meanings (e.g., blow away the feather) to not at all compositional (e.g., carry out a task). That degree of compositionality has been argued to robustly affect a given particle verb's rate of alternation, with less transparently predictable, more idiomatic particle verb meanings corresponding to higher rates of the joined order (Fraser Reference Fraser1976, Kroch and Small Reference Kroch, Small and Sankoff1978, Chen Reference Chen1986, Dehé Reference Dehé2002, Gries Reference Gries2003).
We hand-coded each specific particle verb meaning as either transparently compositional or non-compositional (Grafmiller and Szmrecsanyi Reference Grafmiller and Szmrecsanyi2018, Röthlisberger and Tagliamonte Reference Röthlisberger and Tagliamonte2020). Following previous corpus studies, we use Lohse et al. (Reference Lohse, Hawkins and Wasow2004)'s diagnostic that a particle verb is only compositional if both the bare transitive (3a, 4a) and predication of the particle over the object (3b, 4b) are entailed. Thus, knock off meaning ‘literally strike, causing to fall’ is compositional, but knock off meaning ‘stop some behaviour’, is not compositional.
Particle prosody: Finally, exploratory visualization of word order variation by particle revealed that particle verbs containing away and around, the only two iambic particles, had visibly lower rates of joined usage than the rest of the particles (i.e., down, in, off, on, out, over, and up). We therefore investigate the prosody of the particle as a possible factor, along the lines of previous work that has explored whether phrasal prosodic considerations such as satisfying endweight (Ryan Reference Ryan2019) or conforming to alternating stress (Shih Reference Shih, Gribanova and Shih2017, Grafmiller and Szmrecsanyi Reference Grafmiller and Szmrecsanyi2018) could affect word order alternations.
3.4 Statistical modeling
Particle verb order is a binary measure that is analyzed using logistic mixed-effects models with speaker modality as the main predictor, treatment-coded with studio as the baseline. The model includes control predictors of object length, NPR, log10 of meaning frequency, particle verb compositionality, and particle iambicity. The model also includes a random intercept for each particle verb meaning, and was fit in R (R Core Team 2020) using the glmer function in the lmer4 package, which provides p-values based on asymptotic Wald tests (Bates et al. Reference Bates, Mächler, Bolker and Walker2015: 101).
4. Results
All control predictors were found to have significant effects on the variation. Complete model output is available in Appendix A, but to summarize: the joined variant is favoured with longer objects, more frequent particle verb meanings, non-compositional particle verbs, and on NPR shows. Object length and compositionality also interact significantly, such that the preference for joined order is stronger for longer objects with non-compositional particle verbs. The split variant is favoured with the iambic particles around and away. However, the predictor of speaker modality (studio vs. telephone) does not reach significance (β = 0.09, SE = 0.11, p = 0.41).
There are many data points in our “studio” category from radio shows that do not themselves have a call-in component. In order to see whether our null effect of speaker modality is somehow due to non-call-in show studio data skewing the results, we narrowed down the studio data to only those data points from shows that also have a telephone component in our data. But modality does not reach significance in this subset either (β = 0.08, SE = 0.11, p = 0.48). Figure 1 plots the rate of use of the joined variant among studio speakers in the full data set, the subset of studio speakers from shows that also have a telephone component (“call-in shows”), and telephone speakers. The three groups have identical rates of use of the joined variant at 81%.
We also checked for the possibility that speaker modality interacts with one of our other predictors. Speaker modality does not interact significantly with compositionality (β = 0.05, SE = 0.30, p = 0.87), object length (β = 0.04, SE = 0.15, p = 0.79), particle prosody (β=−0.05, SE = 0.57, p = 0.93), or meaning frequency (β = 0.03, SE = 0.15, p = 0.86). In all four cases, inclusion of the interaction raises AIC and BIC compared to the model with no interaction, and does not change the non-significant main effect of speaker modality.
5. Discussion
What might explain this failure to replicate Kroch and Small (Reference Kroch, Small and Sankoff1978)? One option is that the coding of studio vs. telephone modality is unreliable in our data. We think this unlikely, given the clear acoustic signature of telephone speech, as discussed in section 3.3. In the rest of this section, we discuss two possible sources of the difference between Kroch and Small (Reference Kroch, Small and Sankoff1978)'s results and ours.
5.1 Changes since the 1970s
It is possible that our null result represents a change in the prescriptive norms surrounding particle verb order since the 1970s. An acceptability study in Haddican, Johnson, and Hilton (Reference Haddican, Johnson and Hilton2016) reveals that younger participants have a weaker preference for the joined order than older participants do. Moreover, Haddican, Johnson, Wallenberg, et al. (Reference Haddican, Johnson, Wallenberg, Holmberg, Beaman, Buchstaller, Fox and Walker2020) find that both US and UK written Englishes are changing toward the split variant in real time, possibly suggesting “the loosening of a prescription against the VOP order in written English” (p. 220), which may or may not have a parallel in spoken English. Finally, in a matched guise study, Robinson and MacKenzie (Reference Robinson and MacKenzie2019) find that participants do not socially evaluate speakers who produce all particle verb tokens in the joined order differently from speakers who use all split.
Conversely, it is possible that prescriptive norms on particle verb order persist, but that radio hosts/in-studio guests do not contrast with callers on their adherence to those norms. This could either be because the two groups are too socially disparate with respect to factors that may matter more for adherence to prescriptivist norms,Footnote 3 or because talk radio's orientation towards those prescriptive norms has changed since the 1970s. The radio talk shows of 2018–19 inhabit and generate different social and interactional contexts than the radio talk shows of 1978. In the 1970s and 1980s, talk show audiences – and callers in particular – were socially peripheral (Turow Reference Turow1974, Bierig and Dimmick Reference Bierig and Dimmick1979, Armstrong and Rubin Reference Armstrong and Rubin1989), but today's talk radio listenership is more integrated into general society (Hofstetter et al. Reference Hofstetter, Donovan, Klauber, Cole, Huie and Yuasa1994, Turow et al. Reference Turow, Cappella and Jamieson1996, Davis and Owen Reference Davis and Owen1998). Moreover, regulatory shifts in the late 1980s and 1990s led to increasingly polarized content appealing to “talk-radio listeners[, who are] fairly united in their strong anti-government sentiments – sentiments that are often conditioned and reinforced by talk-show hosts” (Owen Reference Owen and Craig1996: 134, Davis and Owen Reference Davis and Owen1998). It is possible that today's talk radio hosts’ mistrust of government and other powerful institutions extends to established standards in language, too. Perhaps talk radio hosts no longer orient themselves towards prescriptive linguistic norms in general; perhaps their stance towards dominant language ideologies has changed. If so, then future work with corpora with speech from contexts or speakers more susceptible to prescriptivist pressures may yet uncover social conditioning of particle placement.
5.2 Data and statistical problems with Kroch and Small (Reference Kroch, Small and Sankoff1978)
Finally, we have to recognize the possibility that the original effect in Kroch and Small (Reference Kroch, Small and Sankoff1978) was spurious. Specifically, their finding could be driven by differences in the length of direct objects used by each speaker group. As discussed above in section 3, particle verb constructions become nearly categorical in requiring the joined order with objects of greater than six words (Grafmiller and Szmrecsanyi Reference Grafmiller and Szmrecsanyi2018: 389), and we excluded tokens with objects of greater than four words for this reason. Kroch and Small, on the other hand, split tokens into those with short “L1” objects (<three words in length) and those with longer “L2” objects (three+ words). Longer objects which would categorically require the joined order seem to be included in that “L2” category (see, for instance, the underlined six-word object in their example (10c) Harry mulled over the idea that his wife suggested). Although the difference between speaker groups in the “L2” (three+ words) condition is not significant, this is likely due to the relative sparsity of caller data (36 tokens, 58% joined) vs. host/guest data (64 tokens, 73% joined).Footnote 4 It is therefore possible that the strongly joined host/guest “L2” sample could contain many very-long objects, which might then drive the overall effect of modality within the full dataset.
Another factor complicating efforts to replicate Kroch and Small's finding is lack of clarity concerning the envelope of variation and the operationalization of predictors in their study. For one, it is possible that the verb-particle constructions that Kroch and Small studied do not fully overlap with our definition of the phenomenon. In exemplifying their semantic dependence (compositionality) predictor, Kroch and Small write that “Cases where the particle is an adverb and functions as such in nonparticle as well as particle constructions (e.g., bring home the bacon) were excluded from the analysis of this factor” (emphasis ours), which seems to imply that they included bring home elsewhere in the analysis (Reference Kroch, Small and Sankoff1978: 50). In our data, only constructions where around, away, down, in, off, on, out, over, or up function as the particle, and not elements like home, are included; it is possible that differences in defining the notion of particle verb explain the divergence between Kroch and Small's findings and our own.
Another such possible difference in defining the phenomenon under study arises from Kroch and Small's “D0” semantic category of particle verbs, which they describe as cases where “the particle has no semantic content except as part of the verb” (p. 50). This category is both the best-represented in their data (153 (62%) out of the 246 tokens they coded for compositionality), and the only semantic category where the host/guest vs. caller distinction is significant (their Table 4). However, it is not clear which particle verbs exactly map to this category. Kroch and Small represent “D0” using their example (13) He called the mayor up, and subsequent work has argued compellingly that up (as used in call up) and other “aspectual particles” actually do have semi-predictable semantic contributions to a verb's aktionsart (Bolinger Reference Bolinger1971, Brinton Reference Brinton1985, Tenny Reference Tenny1994, Jackendoff Reference Jackendoff, Dehé, Jackendoff, McIntyre and Urban2012, i.a. on such particles as telicity markers; see Cappelle Reference Cappelle2005, Walkova Reference Walkova2013, Larsen Reference Larsen2014 for alternative approaches to the semantics of “aspectual” up). Kroch and Small's definition, however, also seems compatible with non-compositional particle verbs that are more opaque and idiomatic, like carry out (duties) or knock off (behaviour). We include and code such tokens as non-compositional in our own dataset, but it is unclear to us whether Kroch and Small did so in theirs. In any case, our results show no interaction of speaker modality and semantic compositionality, regardless of whether we split aspectual particle verbs like Kroch and Small's example (13) from other types (β = −0.31, SE = 0.27, p = 0.24), or consider them together with all other non-transparently-compositional particle verbs (as described in section 3.3 and reported above). This both shows the indeterminacy of which particle verbs Kroch and Small were including for study, and demonstrates that our analysis is not missing a modality effect that is hidden in a particular syntactic/semantic class of particle verbs.
6. Conclusion
This study shows that Kroch and Small's finding – that social roles (in radio talk show contexts) condition the rate of joined variant use for particle verbs – does not replicate. It is possible that either the specific prescriptive linguistic norm under study has shifted, or that the way that talk radio relates to that norm has changed. Another possibility is that Kroch and Small's original study suffered from statistical flaws or issues with defining the phenomenon consistently.
The results of this study make the search for positive cases of social evaluation of syntactic variation even more pressing. Future work should both seek evidence of social influence on particle verb order in other types of data (including large-scale matched guise studies), and cast a wider net for the social correlates of syntactic variation in a variety of languages, variables, and data sources.
Supplementary material
To view supplementary material for this article, please visit https://doi.org/10.1017/cnj.2023.13
Appendix A: Model results