3.1 Selection criteria

For the final data set, a token was considered interchangeable if it could easily be converted from one variant to the other without the insertion or deletion of any lexical items from either the possessor or possessum. The one exception to this was the obligatory deletion of the determiner in the possessum, as in converting the principles of the country to the country's principles. Aside from this intuitive rule, several additional heuristics aided in classifying the genitive tokens. Genitives not meeting the interchangeability requirement fell into several classes, all of which have been discussed extensively in previous literature (e.g. Quirk et al. Reference Quirk, Greenbaum, Leech and Svartvik1985; Biber et al. Reference Biber, Johansson, Leech, Conrad and Finegan1999; Rosenbach Reference Rosenbach2002, Reference Rosenbach2006; Kreyer Reference Kreyer2003; Hinrichs & Szmrecsanyi Reference Hinrichs and Szmrecsanyi2007).

1. (a) Elliptical genitives. Constructions lacking an overt possessum phrase: I’ll meet you at Pat’s, Her eyes are like a hawk’s

2. (b) Descriptive genitives. Constructions in which the possessor is not itself referential, but acts as a classifier that ‘contributes to the denotation of the head noun’ (Rosenbach Reference Rosenbach2006: 81): women's magazines, smoker's cough

3. (c) Fixed expressions. Phrases that have become conventionalized in one construction or the other: arm's reach, Murphy's law, President of the United States, the law of the land

4. (d) Authored works. Titles of books, films, musical pieces, etc., that are pre-modified by a possessor denoting their creator (Hinrichs & Szmrecsanyi Reference Hinrichs and Szmrecsanyi2007): Allen Ginsburg’s Howl, Racine’s Phèdre

5. (e) Any possessum headed by a determiner other than the, e.g. demonstratives, quantifiers and indefinite determiners: nineteen years of age, this responsibility of the teacher, a characteristic of Trevelyan's prose

6. (f) Collective possessums. Of-genitive constructions that refer to collections of individuals: the group of students, the jury of eight women and four men

Finally, all tokens involving pronominal possessors were excluded from the data set. Jucker (Reference Jucker and Jucker1993) observes in his corpus of newspaper prose that pronominal possessors are nearly categorical in their preference for the prenominal position, and this trend is characteristic of English more generally (Huddleston & Pullum et al. Reference Huddleston and Pullum2002; Rosenbach Reference Rosenbach2002). Furthermore, pronominal of-genitives are proscribed quite stringently in the prescriptive literature (Murphy Reference Murphy1997: 128; Quinion Reference Quinion2005). Such proscriptions may not play much of a role in casual speech, but in professionally edited writing they may be taken more seriously, resulting in a higher-level editorial purging of pronominal of-genitives.

3.2 Corpus results

Selection of the interchangeable genitives yielded a data set containing 3,612 genitive tokens (1,115 spoken; 2,497 written). In the present data, the of-genitive construction is more frequent than the s-genitive across both spoken and written varieties.Footnote ³ Within the written data, the of-genitive is more frequent than the s-genitive for each of the five genres (table 2). Though the relative frequencies are similar, there are highly significant differences between the proportions of the two constructions in the general fiction,χ² (1,358) = 25.74, p < 0.0001, non-fiction,χ² (1,919) = 126.70, p < 0.0001, and learned genres, χ² (1, 404) = 129.96, p < 0.0001. Even in western fiction there is a marginally significant preference for the of-genitive, χ² (1, 399) = 4.63, p < 0.05. No difference in the proportion of genitives emerges in the press genre, χ²(1, 433) = 1.02, p = 0.31, suggesting there is something special about newspaper prose with respect to this particular linguistic variable – a finding that is certainly not new (Jucker Reference Jucker and Jucker1993; Hundt & Mair Reference Hundt and Mair1999; Szmrecsanyi & Hinrichs Reference Szmrecsanyi, Hinrichs, Nevalainen, Taavitsainen, Pahta and Korhonen2008).

Table 2.

Distribution of genitive constructions across genres

4.1 Semantic and pragmatic factors

animacy. Rosenbach (Reference Rosenbach2006: 105−6) notes that the genitive alternation is sensitive to at least a four-way animacy ranking (human > animal > collective > inanimate). This is the coding used by Hinrichs & Szmrecsanyi (Reference Hinrichs and Szmrecsanyi2007) and Szmrecsanyi & Hinrichs (Reference Szmrecsanyi, Hinrichs, Nevalainen, Taavitsainen, Pahta and Korhonen2008) in their analyses, which confirmed Rosenbach's rankings, while Tagliamonte & Jarmasz (Reference Tagliamonte and Jarmasz2008) find similar results using a simple human/nonhuman distinction. Since factors other than animacy are the primary concern of the present study, animacy was coded as a binary living/non-living distinction: living beings (humans, animals and organizations) were coded as animate, everything else was coded as inanimate. Representative examples are shown in (3) and (4).

1. (3) Animate possessors:

   Eileen's address, the death of the officer, Papa's footsteps, the truest societies of Christians, the hoot of an owl

2. (4) Inanimate possessors:

   the value of voting, the edge of my chair, the winners of last year, today's society, Utah's weather, society's ills

semantic relation. Classifying the various semantic relations encoded by the English genitive construction can be quite challenging (e.g. Taylor Reference Taylor1996: 339−48; Payne & Berlage Reference Payne and Berlage2011). Semantic relation was coded according to the schema used by Rosenbach (Reference Rosenbach2002: 120−3) in her experiments on English genitives. Rosenbach collapsed several classes of relations into a single category of ‘prototypical’ genitives, which favor the s-genitive, and all others into a category marked simply as ‘non-prototypical’. Prototypical genitives were any examples that fell into one four subclasses, listed in table 3.

Table 3.

Categorization of semantic relations

Every observed token of the genitive construction was manually annotated for the type of semantic relation, and the codings for each were cross-checked among three different annotators.

information status. The influence of possessor information status is still somewhat controversial. While some have claimed that given possessors favor the s-genitive (Biber et al. Reference Biber, Johansson, Leech, Conrad and Finegan1999), others have found no significant effect of possessor givenness on English genitives (Hinrichs & Szmrecsanyi Reference Hinrichs and Szmrecsanyi2007). To avoid the risk of (unintentional) bias from using automated coding (see Hinrichs & Szmrecsanyi Reference Hinrichs and Szmrecsanyi2007: 451) each possessor was manually annotated as either given or new. Givenness was defined as explicit reference to the possessor in any manner in the preceding ten lines.

thematicity. Osselton (Reference Osselton, Nixon and Honey1988) examined the tendency of topical or ‘thematic’ possessors to favor the s-genitive construction, even when they are otherwise disfavored. Thematic possessors are those

As a metric of thematicity, the (log) frequency of the head noun of the possessor NP in each text or conversation was used (Hinrichs & Szmrecsanyi Reference Hinrichs and Szmrecsanyi2007).

4.2 Phonological factors

final sibilant. It is well established that the presence of a final sibilant in the possessor NP disfavors the use of the s-genitive (Quirk et al. Reference Quirk, Greenbaum, Leech and Svartvik1985; Biber et al. Reference Biber, Johansson, Leech, Conrad and Finegan1999; Kreyer Reference Kreyer2003; Hinrichs & Szmrecsanyi Reference Hinrichs and Szmrecsanyi2007, inter alia).Footnote ⁴ All possessor NPs were automatically annotated for the presence or absence of a final sibilant, using the phonetic transcriptions in the Carnegie Mellon University (CMU) Pronunciation Dictionary (v. 0.7a). Words that were not found in the CMU dictionary were subsequently hand-coded. Six phones were classified as sibilant: [s] as in house, [z] as in news, [ʃ] as in Bush, [tʃ] as in avalanche, [dʒ] as in bridge and [ʒ] as in garage.

rhythmicity. The comparative rhythmicity between the two genitive constructions was coded according to the methods developed by Shih et al. (forthcoming), who demonstrated that preferred genitive construction will be the one resulting in a stress pattern that most closely accords with an ideal weak–strong stress alternation. Each genitive token was coded for the lexical stresses of each word, obtained from the CMU dictionary. For each token, two stress distance measures were calculated: one for the s-genitive, and one for the of-genitive. For the s-genitive construction, the distance was the number of unstressed syllables between the final stressed syllable of the possessor NP and the first stressed syllable of the possessum noun (phrase). Conversely, the distance in the of-genitive alternate was the number of syllables between the final stress of the possessum NP and the first stress of the possessor NP. Within our data, s-genitive distance ranged from 0 to 4, while of-genitive distance ranged from 1 to 6.

From these two measures, a measure of the comparative rhythmicity between the two constructions was computed using the formula in (5) (see Shih et al. forthcoming).

1. (5) Rhythm measure:

   rhythm = | of-genitive distance – 1 | – | s-genitive distance – 1|

When the of-genitive distance is greater than the s-genitive, rhythm is positive, and the greater the value, the more rhythmic the s-genitive is relative to the of-genitive. When rhythm is negative, the s-genitive distance is greater than the of-genitive, and the smaller the negative value of rhythm, the more rhythmic the of-genitive is compared to the alternative s-genitive. In short, the higher rhythm is above 0, the more the s-genitive should be preferred, and the lower rhythm is below 0, the more the of-genitive should be favored. When rhythm is 0, neither construction is preferred to the other.

4.3 Processing-related factors

end-weight. There is a rich literature examining the effects of ‘end-weight’ (Wasow Reference Wasow2002) on syntactic in a variety of syntactic phenomena (e.g. Altenberg Reference Altenberg1982; Hawkins Reference Hawkins1994; Rosenbach Reference Rosenbach2005; Bresnan et al. Reference Bresnan, Cueni, Nikitina, Harald, Boume, Kraemer and Zwarts2007). In the genitive construction, it is known that heavier, i.e. longer, possessors favor the of-genitive construction, since the possessor is placed second in such instances. Following Bresnan & Ford (Reference Bresnan and Ford2010) and Grafmiller & Shih (Reference Grafmiller and Shih2011), a comparative measure of end weight was calculated for each token using the ratio of the number of orthographic words in the possessor NP to the number of words in the possessum.

persistence. The phenomenon known as ‘structural parallelism’ (Weiner & Labov Reference Weiner and Labov1983) or ‘persistence’ (Szmrecsanyi Reference Szmrecsanyi2006) refers to the influence of speakers’ prior choice of a given construction on their uses of the same construction in subsequent contexts. Examples of s-genitive persistence in spoken and written language can be seen in (6).

1. (6)

   1. (a) Some of the, some of the women's roles, I think, are almost for the worse, because we’re losing out on some things, going back to work, but I think if we can, if we can expand the men's roles at the same time, . . .<Switchboard A.2370>

   2. (b) . . .which was his nickname for a messenger who had worked in the White House since Teddy Roosevelt's administration, and discuss the welfare of some one of the animals. It was part of Little Jack's work to look after the dogs. <Brown G41>

Following Hinrichs & Szmrecsanyi (Reference Hinrichs and Szmrecsanyi2007), each genitive token was coded for the use of an s-genitive in the first genitive choice context – if any – immediately preceding it.

4.4 Economy-related factors

type–token ratio. Being the shorter, more ‘compact’ alternative, the s-genitive construction tends to be preferred in environments of high lexical density (Biber et al. Reference Biber, Johansson, Leech, Conrad and Finegan1999; Biber Reference Biber, Aitchison and Lewis2003; Hinrichs & Szmrecsanyi Reference Hinrichs and Szmrecsanyi2007; Szmrecsanyi & Hinrichs Reference Szmrecsanyi, Hinrichs, Nevalainen, Taavitsainen, Pahta and Korhonen2008). To measure lexical richness, the type–token ratio (TTR) over the five sentences preceding and following each token was calculated. With the exception of western fiction, all genres exhibit a significant tendency for s-genitives to occur more in lexically dense environments (One-sided Mann-Whitney test: spokenU = 129721, p < 0.0001; learnedU = 11964, p < 0.05; non-fictionU = 73728, p < 0.001; pressU = 20289.5, p < 0.01; gen. fictionU = 13229.5, p < 0.05; west. fictionU = 19553.5, p = 0.459).

5.1 Model 1: the combined data

5.1.1 Model summary

Summary statistics of Model 1 are presented in table 4. Predictive performance of the model is excellent, as indicated by its concordance probability (c index), computed by taking the proportion of all possible -s-genitive and -of-genitive pairs in the data for which the model correctly assigns a higher probability to the s-genitive.Footnote ⁵

Table 4.

Model 1: summary statistics

The predictive accuracy, presented visually in figure 1, provides a more intuitive measure of model performance. The x axis represents the probabilities assigned by Model 1 (binned into deciles), and the y axis the proportions of s-genitives in the data for each of the bins predicted by the model. The closer the points are to a straight line, the more accurate the model.

Figure 1.

Model 1 predictions by observed corpus tokens

5.1.2 Main effects

Statistics for the main effect sizes and explanatory power of the individual predictors in Model 1 are presented in table 8. Model predictions are for the s-genitive, so factors with positive coefficients favor the s-genitive, while negative coefficients reflect a preference for the of-genitive.

Table 5.

Main effects of individual predictors in Model 1. Model predictions are for the s-genitive

For levels of genre, the model coefficients are to be interpreted relative to the baseline level spoken, which is to say, the model's main effects represent those for the spoken data. Numerical predictors were centered and standardized by dividing by twice the standard deviation. This protects against harmful effects of data multi-collinearity, and normalizing by two standard deviations enables direct comparison between effect sizes (coefficients) of numerical and binary predictors (Gelman Reference Gelman2008). For ease of comparison, model estimates (in linear log odds scale) for the significant main predictors are shown in figure 2.

Figure 2.

Estimates of significant main effects of Model 1 (with 95 percent confidence intervals). Genre levels are relative to baseline level spoken

Of all the factors, animacy has the largest effect size by far. Averaging over the entire dataset, animate possessors are almost sixty times more likely to be used in the s-genitive.Footnote ⁶ Prototypical genitive relations, e.g. kinship or part-whole relations, also favor the s-genitive, though by a much smaller factor of about 2.5. The presence of a final sibilant on the possessor has a significant effect, though in the opposite direction. Possessor NPs ending in a sibilant are roughly 4 times less likely to occur in the s-genitive. Similarly, increasing the possessor/possessum length ratio also disfavors the s-genitive. For example, moving from a ratio of 1/2 to a ratio of 3 results in a nearly two-and-a-half-fold decrease in s-genitive likelihood.Footnote ⁷ Lexically richer environments, as measured by TTR, also significantly favor the use of the s-genitive, though the small effect is difficult to interpret meaningfully.

When it comes to the stylistic effects of genre, the most important finding is that, in general, individuals are significantly less likely to use s-genitives in writing than in spoken conversation. The lone exception to this is press. A potential genitive is about 1.8 times more likely to be an s-genitive in newspaper writing than it is in speech. Newspaper prose therefore exhibits a stark contrast to other genres in its preference for the s-genitive construction, in accordance with previous findings (Jucker Reference Jucker and Jucker1993; Leech & Smith Reference Leech, Smith, Renouf and Kehoe2006; Szmrecsanyi & Hinrichs Reference Szmrecsanyi, Hinrichs, Nevalainen, Taavitsainen, Pahta and Korhonen2008).

Turning to the non-significant predictors, there is no main effect of rhythm, persistence or possessor text frequency in the model. Recent studies have found significant, though small, effects for each of these factors (e.g. Hinrichs & Szmrecsanyi Reference Hinrichs and Szmrecsanyi2007; Shih et al. forthcoming), however, the present study differs from these (and others) in its use of mixed-effects models, which treated each individual speaker/author as a source of random variability in the data. Individual speaking and writing styles can vary quite considerably, and so it is possible that the idiosyncratic patterns with respect to individual users would account for enough variability to mask any aggregate effects of these particular predictors here. Nevertheless, despite the lack of significant main effects for these predictors, there are significant interactions involving these and other factors, to which I turn now.

5.1.3 Interactions

Table 6 lists the significant interactions in Model 1, and figure 3 shows a visual profile of the variation in main effects grouped by genre.

Table 6.

Significant interaction effects in Model 1. All genre interactions are interpreted relative to the baseline genre level spoken

Figure 3.

Profile of significant effects by genre

While the model provides little evidence for any interaction effects of phonological or economy-related predictors across spoken and written genitives in general, specific genre interactions offer some insight into the importance of stylistic pressures on two key factors, namely end weight and possessor animacy. For length ratio, the model reports a negative main effect (β = −1.86), interpreted as the effect size of the baseline level spoken. The interaction term for each other genre is positive, indicating a reduced effect of end weight, though this effect is significant only for the general fiction, western fiction, and learned genres. Moreover, for the two fiction genres, the direction of the effect is reversed: a larger possessor/possessum length ratio actually favors the s-genitive. A plausible explanation for this is that these genres are simply less susceptible to the processing demands underlying weight effects. This seems especially true of academic prose (learned), as the genre is well-known (and sometimes parodied) for its complex syntax and larger vocabulary. The other two genres comprise examples of narrative fiction, a diverse category that not only covers a wide range of individual styles – compare, say, Hemingway vs Faulkner – but also often emphasizes aesthetic expression over communicative economy. For these genres, efficiency therefore might take a backseat to artistic style. Furthermore, Hinrichs & Szmrecsanyi (Reference Hinrichs and Szmrecsanyi2007: 465) suggest that the influence of parsing efficiency is operative primarily in more colloquial styles, which would explain the fact that newspaper prose and narrative non-fiction (the latter being perhaps slightly less prone to aesthetic flourishes than fiction) do not differ significantly from speech in this regard.

In contrast to end weight, the effect of animacy is qualitatively the same across all genres, i.e. it has the same direction, but it is quantitatively weaker within some genres, especially newspaper prose, when compared to spoken English. But a closer look reveals that the reasons for this weaker effect are not the same across written genres. General fiction and non-fiction show a marked decrease in log odds for both animate and inanimate possessors, reflecting a general bias for of-genitives in these genres. But this is not the pattern found in newspaper texts. There is indeed a decrease in s-genitive likelihood for animate possessors, but also a marked increase in s-genitive likelihood for inanimate possessors when compared to speech (figure 4). Together with the main effect preference for s-genitives in press styles, the overall reduction in the influence of animacy on genitive choice in newspaper writing suggests a strong stylistic bias in favor of the s-genitive in these texts. While in other written genres, writers strongly prefer inanimate possessors in of-genitives, and animates in s-genitives somewhat less so, journalists are (implicitly) less discriminating. This result is compatible with the notion that reporters’ need for economy of expression is reflected in unconscious adjustments to their probabilistic knowledge (i.e. factor weights).

Figure 4.

Significant interactions of general fiction, non-fiction and press genres and possessor animacy in Model 1

Also of interest are those measures thought to be indicative of either the shifting of the written norm toward more a conversational style, i.e. rhythm, or of increased stylistic pressure to use more compact expressions, i.e. text frequency (thematicity) and TTR (Szmrecsanyi & Hinrichs Reference Szmrecsanyi, Hinrichs, Nevalainen, Taavitsainen, Pahta and Korhonen2008). No significant interaction of genre × rhythm was found, though when it comes to the thematicity and TTR, results from Model 1 point to academic prose (learned) and not newspaper prose as the truly odd duck. General fiction also exhibits a significant interaction with text frequency, though why this should be so is uncertain. It may be due to sampling differences – fiction texts are excerpts of longer material, while newspaper texts constitute an entire article – though this is true of all the non-press genres. It could also simply be random chance that the particular set of texts exhibit this pattern. Further investigation is certainly needed.

Finally, there are the significant interactions of animacy × rhythm and animacy × semantic relation. The negative interaction coefficient for animacy × rhythm (β = −0.57) shows that as we move from inanimate to animate possessors, the influence of rhythm significantly decreases. In fact, there is no significant influence of rhythm among animate possessors. The model does no better than chance when predicting the effects of rhythm when the possessor is an animate entity. The opposite pattern obtains for the animacy × semantic relation interaction.

5.2 The written data

To answer this question, a second model (Model 2) was run on only the written data, treating press as the baseline level of a five-level genre factor. All other predictors in the model were the same as for Model 1.

Before moving on to the results of Model 2, I return to the distribution of genitives among the different written genres, presented graphically in figure 5. With the exception of press texts, all genres contain significantly higher proportions of of-genitives than s-genitives, with the largest disparity in the learned genre.

Figure 5.

Distribution of written genitives by genre (* = p < .05, ** = p <.01, *** = p < .001)

Writing in these texts tends toward more deliberate word and construction choice, and unlike with other genres of written English, this trend toward greater complexity – and by implication, greater formality – has been growing stronger in academic writing over time (Biber Reference Biber1988, Reference Biber1995). Between the two extremes lie the three other genres: general fiction, western fiction and non-fiction. Biber (Reference Biber1988) situates all these genres (including press and learned) along an ‘informational vs involved focus’ dimension, which marks ‘high informational density and exact informational content versus affective, interactional, and generalized content’ (107). Yet, despite their similar informational focus, the largest divergence in the distribution of genitives and other constructions (see Biber Reference Biber, Aitchison and Lewis2003) is between academic and journalistic prose styles. As I show in section 6.2.2, this divergence is reflected in adjustments to the probabilistic weights of several key factors affecting genitive choice. These results fit with the intuition that the of-genitive is the more formal variant (Rosenbach Reference Rosenbach2002: 39; Tagliamonte & Jarmasz, Reference Tagliamonte and Jarmasz2008).

5.2.1 Model 2 summary and main effects

Statistics for Model 2 are presented in table 7. As with Model 1, Model 2 performs quite well, as indicated by the high c statistic. Model 2 accurately predicts 86 percent of the possible outcomes (see figure 6), and bootstrap cross-validation confirms that it does not over-fit the data. Overall, the explanatory power and fit of the model is excellent.

Table 7.

Model 2: summary statistics

Figure 6.

Predictions for Model 2 fixed effects by observed corpus tokens

Results from Model 2 are mostly consistent with those of Model 1. Animacy is again the most important predictor of genitive choice by far, followed by significant contributions of genre, the presence of a final sibilant, length ratio, possessor givenness, and finally the text frequency of the possessor (figure 7). The details of the main effects are shown in table 11.

Figure 7.

Estimates of significant main effects of Model 2 (with 95 percent confidence intervals). Genre levels are relative to baseline level press

As the findings of Model 1 suggested, there is a great deal of variation between newspaper prose and other writing styles when it comes to genitive choice. Simply knowing the genre in which a possessor–possessum pair occurs tells us more about the author's likely choice of genitive construction than any other single factor besides possessor animacy. In the most extreme example, a genitive is over twenty times more likely to be an s-genitive in a newspaper article than in an academic journal (learned). As expected, final sibilant and length ratio have significant negative effects on s-genitive likelihood, while the text frequency of the possessor has a significant positive influence, unlike in Model 1. Moving from tokens featuring possessors in the lowest frequency range (1 occurrence) to those with possessors in the highest range (> 9 occurrences), the likelihood of the s-genitive increases by a factor of roughly 2.5. Possessor givenness also plays a role in predicting written genitives, with given possessors being 3.26 times more likely to occur in the s-genitive. This is not surprising given the significant positive interactions of givenness and genre observed in Model 1. Lastly, semantic relation, persistence, rhythm and TTR each failed to achieve significance as main effects.

Table 8.

Statistics of individual predictors in Model 2. Model predictions are for the s-genitive

5.2.2 Interactions

With a couple exceptions, the significant interactions in Model 2 largely accorded with the findings of Model 1. In comparing written genres to speech, Model 1 reported significant positive interactions of the genres general fiction, learned, and western fiction with length ratio, but no interaction with press, suggesting that newspaper texts do not differ significantly from speech in this regard. Not surprisingly, Model 2 also reports a significant positive interaction with each of these genres when compared to the baseline press. Similarly, Model 2 finds significant positive interactions of learned × type-token ratio and animacy × semantic relation, just as one would expect given the previous model's results.

Table 9.

Interaction effects in Model 2. Non-significant interactions not shown

Two other sets of interactions deserve additional attention. First, unlike Model 1, every written genre significantly interacts with possessor animacy. The direction of the effect does not change, but the influence of animacy greatly increases for every genre relative to press. Not only is the effect of animacy significantly weaker in most written genres compared to speech, but it is significantly weaker still for newspaper writing. This pattern can be represented as a ranking along a scale according to how strong the effect of possessor animacy is for each genre.

1. (5) Importance of Possessor Animacy on genitive choice:

   Speech > Fiction, Narrative non-fiction, Academic > Press

This pattern is hinted at in Model 1 by the large negative interaction term of press × animacy, but only confirmed through direct comparison in Model 2.

Second, two genre interactions with final sibilant also have significant effects (p < 0.05). Among non-press genres, the presence of a final sibilant on the possessor has the expected effect: possessors with final sibilants are much more likely to be used in the of-genitive. In newspaper prose, however, the presence of a final sibilant has no effect on the choice of construction; either construction is equally likely. Taken together, these results show that the influences of factors which tend to favor the of-genitive in other genres are significantly attenuated in journalistic English.

5.3 A note on the corpora

Before moving on, it should be mentioned that the two corpora used in this study are not only split in modality, but in time as well. The Brown and Switchboard corpora comprise data from the early 1960s and early 1990s respectively, and it has been shown that changes in genitive frequencies have occurred in just that brief period (Leech & Smith Reference Leech, Smith, Renouf and Kehoe2006; Hinrichs & Szmrecsanyi Reference Hinrichs and Szmrecsanyi2007). It is therefore uncertain how much comparing the two directly can tell us about general trends in spoken versus written English, though it is possible to get an inkling of what a genuine synchronic comparison might reveal. Using birthdate information from Switchboard, a simulated apparent time study compared Brown data to Switchboard data from speakers who would have already been adults at the time Brown was collected (1960–1). Results from the time-adjusted corpus (N = 2759; 295 spoken) are very similar to those of the full corpus. Compared to speech, s-genitives in the time-adjusted corpus are significantly more likely in newspaper texts (β = 0.99, p < 0.01), but significantly less likely only in the learned genre (β = −1.38, p < 0.05). This contrasts with results from the full model showing the s-genitive to be significantly less likely in all written genres (other than press) when compared to conversational speech. Together, these results support previous findings that spoken use of the s-genitive has been increasing relative to most written genres over the last few decades (Szmrecsanyi & Hinrichs Reference Szmrecsanyi, Hinrichs, Nevalainen, Taavitsainen, Pahta and Korhonen2008), though in this case it has not quite reached the level in journalistic prose.

The interaction of genre with possessor animacy demonstrates this even more clearly. Only press texts show a significant reduction in the influence of animacy in the time-adjusted corpus (β = −2.44, p < 0.0001), similar to what was found in the full model. With the exception of these texts, older speakers in the Switchboard data do indeed appear to pattern more like writers in the Brown corpus in their choice of genitive construction. In other respects, the time-adjusted model reports the very same significant interactions of genre × givenness and genre × length ratio, adding confidence to the interpretations of Model 1.