Modelling adjective phrase inversion as an instance of functional specialization in non-locative inversion

Heidrun Dorgeloh & Gero Kunter
Heinrich Heine University Düsseldorf

Abstract

English non-locative inversion is a semantic subtype of full inversion in which the fronted constituent is an adjective phrase (most disturbing is…) or an ascriptive noun phrase (an exception to this rule is…). In the history of English, this type of inversion has emerged relatively late and is subject to specific constraints and discourse functions of the left clause periphery.

Based on the assumption that elements at the left periphery have a connective and a subjective function, we present data from a historical corpus to explain the development of this type of inversion over the last 200 years. Our statistical models find different developments for different semantic subtypes. Adjective phrases with a subjective meaning, which have been frequent in the 19th century, show a decreasing frequency during the last century. By contrast, adjective phrases with a connective meaning have been increasing in frequency over the last 100 years, and our models predict that they are likely to do so in the future.

This rather fine-grained analysis of word order variation suggests that the study of language change calls for an integration of semantic and syntactic categories, and that the possibilities for predicting language change also depend on a suitable selection of the scale of classification.

1. Introduction

Non-locative inversion is a semantic subtype of full inversion, which means it is a declarative sentence in which “the subject occurs in postposed position while some other dependent of the verb is preposed” (Ward, Birner & Huddleston 2002: 1385). In contrast to subject-auxiliary inversion, which follows initial negative or restrictive elements, in full inversion the subject follows the full, lexical verb. This happens together with the preposing of a sentence element that typically functions as a verbal complement, often a predicative complement of the verb be. Examples of this typical pattern of full inversion are given in (1) and (2). [1]

(1) It starts and ends in the Playhouse, and in the middle is a ‘Wizard of Oz’ - like made-up land. (NEWS_AP_2001)
(2) […] Morford is " essentially " telling the truth. Even more interesting is that he thinks Satan is readying a room for old Bill […]. (NEWS_SanFran_2005)

The preposed element can also be an adjunct, as in (3), which is why these inversion constructions have been described as “triggered”, but “optional” word order alternations (Huddleston & Pullum 2002: 96).

(3) His quest began in 1895, when as a 16-year-old he imagined what it would // be like to ride alongside a light beam. A decade later came his miracle year, described in the letter above […]. (NF_EinsteinHisLife_2007)

Long-term historical studies agree that the history of full inversion is closely linked to the change of English from a verb-second to a subject-verb language (Nevalainen 1997; Bækken 2000; Warner 2007; Los 2009). There is further consensus that among the inversion constructions that survived after the Middle English period, at least up to Early Modern English, there were “vestiges of the old verb-second rule” (Nevalainen 1997: 213). Verb-subject order following deictic elements is one example of this. This deictic referencing followed by inversion has been available throughout the history of English as one of the “late subject” constructions (Warner 2007: 94), which serve “to establish textual cohesion” (Los 2009: 106). The cohesive link is typically a deictic adverb, as in (4), or a demonstrative element in a phrase, as in (5):

(4) Fæder her is cumen an enuchus of cinges hirede (OE; from Los 2009: 103)
FATHER HERE IS COME A EUNUCH OF KING-POSS HOUSEHOLD
‘Father, a eunuch from the King’s household has arrived’
(5) Of þese seune heudes comen alle manere of synnes (ME; from Warner 2007: 94)
FROM THESE SEVEN HEADS COME ALL MANNER OF SINS
‘From these seven heads spring all manner of sins’

In contrast to these vestiges of Old and Middle English word order variability, full inversions following fronted complements of predicative, rather than deictic, meaning have shown less continuity in their long-term development. They are discussed in the research literature as instances of a “functional renewal” (Brinton & Stein 1995; Nevalainen 1997; Stockwell 1984) because they seem to have re-emerged only after the end of the Middle English period. Diachronically, this emergence has been described as the “functionalization of a focusing strategy” (Brinton & Stein 1995: 39) and, synchronically, as the extension of the “deictic prototype” of inversion to a “lexical” type following predicative complements (Dorgeloh 1997: 67–74).

However, since predicative complements often also contain a deictic element, the two processes and the distinction they are based on are not entirely unrelated. In (1), for example, the inversion implicitly means “…in the middle [of this Playhouse] is a ‘Wizard of Oz’”. For that reason, describing the historical processes along which inversion has developed has always also been a classificatory challenge.

From a modern perspective, the semantic distinction of locative inversion such as (1) and non-locative inversion such as (2) thus cuts across the historical classification of deictic, “late-subject” inversion and lexical inversion with a new, focusing function. The distinction constitutes two main types of inversion which differ in several important respects (Ward, Birner & Huddleston 2002: 1386; see section 2.1). What they share is that, except when following deictic adverbs, the construction is nowadays bi-focal because, against the fixed word order of Modern English, the front-position for a verbal complement creates as much extra-focus as does the late position of the subject (Dorgeloh 1997: 79). Still, the locative semantic type is more like the vestiges from Old English because it replaces the deictic reference in the fronted constituent with the “lexical information necessary to construct a mental model of the reference situation” (Drubig 1988: 91).

Our study only explores non-locative inversion. We further limit the study to inversion following fronted adjective phrases as exemplified by (2), since an adjective in the fronted phrase constitutes a clear instance of an element with lexical, predicative meaning: Even more interesting… as a point of departure for the inversion does not involve deictic referencing. [2] Our research aim is to describe relevant aspects of the path of the development that this instance of the re-newed type of inversion has taken after the (re-)emergence of inversion with a focusing function.

Within this strictly confined category, and for a relatively limited period of 200 years, we will look for similarities with previously known historical processes, particularly those having to do with fronting and the sentence-initial position. We expect that the patterns of adjective phrase inversion follow a functional specialization in line with general principles of development at the left clause periphery. At the same time, we attempt to demonstrate that the possibility of predicting language change largely depends on choosing a suitable scale of classification.

We present the background for the classification used in this study and for the functional factors that are typical for processes at the left clause periphery in section 2. In section 3, we discuss our methodology, and we present our data analysis and the results in section 4. The final discussion in section 5 will show that the relevance of our findings results from our purposefully fine-grained look at inversion, but also from an integration of syntactic, semantic, and discourse-functional categories. We will argue that the statistical methods that have been used for modeling language change can benefit from a careful selection of the linguistic categories that they employ, and that the scale of generalization that is used for the categorization of the linguistic data may improve their usefulness for uncovering and actually predicting the gradient nature of developments that are more short-term.

2. Background

2.1 The distinction between locative and non-locative inversion

A variety of phrase types undergo full inversion in Modern English. A classification according to the constituent type of the dependent leads to the widely used types AdvP-inversion (now is the time…), PP-inversion (among the members of the meeting were…), VP-inversion (gone are the days…), AdjP-inversion (most disturbing is…), and NP-inversion (an exception to this rule is…) (see, for example, Dorgeloh 1997: 23–26). Of these, corpus studies have shown PP-inversion and NP-inversion to be the two most frequent categories, while the other three types are “fairly rare” (Kreyer 2006: 121).

As to the general function of inversion, the accepted analysis is that it is a syntactic strategy available for information structuring: the preposed constituent typically contains information that is more given, or at least assumed to be more recoverable, than the one represented by the postposed constituent (Birner 1992, 1994). What is important to note in this context is that the sentence-initial position can even cue the reader to evoke this information status (Birner 1994: 255).

If the preposed constituent expresses a spatial location, path, or direction, the inversion that it triggers has been termed locative inversion (Bresnan 1994). It is most typically associated with the fronting of an adverb or prepositional phrase, although the “correlation of locativity […] and preposed constituent is imperfect” (Birner 1992: 51–52). For example, inversion following fronted participles can fall into this category because these can also have locative meaning (e.g. Crashing through the woods came a wild bear…; Bresnan 1994: 75). By contrast, there are prepositional phrases that do not carry locative meaning and therefore do not belong to locative inversion (such as Of greater concern is …; Dorgeloh 1997: 79).

Apparently, non-locative inversion does not differ from locative inversion in phrase type, but in being “generally restricted to the verb be” (Bresnan 1994: 76). This criterion limits non-locative inversion to adjective phrases and other ascriptive phrases, exemplified in (6) and (7): [3]

(6) […] whiskey advertisements also are to be discouraged. More significant is the promise to give up altogether […]. (MAG_Nation_1910)
(7) […] this difficulty will be eliminated, officials believe. Another problem is the matter of reports from industry. (NEWS_WallStJrnl_1941)

Furthermore, in comparison to locative inversion, non-locative inversion has a closer connection to the preceding context: it does not only underlie the pragmatic constraints of relative discourse familiarity that apply to subject and dependent across all inversion types, but also requires that there should be an “open proposition” (Ward, Birner & Huddleston 2002: 1386). Such a proposition must be derivable from the preceding context – except for the postposed subject, which the inversion signals as being discourse-new. For instance, (6) and (7) have (8) and (9) as contextually derivable propositions:

(8) Discouraging whiskey advertisements is significant.
(9) This difficulty is a problem.

Previous accounts of inversion have emphasized that without locative meaning, inversion serves other discourse functions than when following locative constituents (e.g. Chen 2003; Kreyer 2006; Prado-Alonso & Acuña Fariña 2010). In particular, the fronting of noun phrases and adjective phrases is not due to “experiential iconicity”, but to their role within the text (Prado-Alonso & Acuña Fariña 2010: 537). As a result, different genres show different preferences for locative inversion. Since locative inversion resembles the presentative prototype, i.e. inversion with deictic adverbs, it has a natural affinity to occur in the immediacy of spoken discourse (Chafe 1992, 1994), or in the displaced immediacy of narration and description (Dorgeloh 2006). By contrast, non-locative inversion is not rooted in the experience of locative relations, but in the organization of the discourse (Dorgeloh 2006: 110–111). This explains why non-locative inversion possesses a strong affinity with more argumentative and expository genres which are primarily written, for instance scientific and official texts (Prado-Alonso & Acuna-Farina 2010: 537).

Data from a historical corpus study, which can only come from written discourse, are therefore more likely to model the overall pattern of development of inversion if they are limited to non-locative inversion. Instead of looking at both NP-inversion and AdjP-inversion, we limit our study to the latter type for two main reasons. First, sentences with NP-inversion have the same NP + be + NP structure as canonical English sentences. In principle, a sentence can be tested for the presence of inversion by way of a corresponding interrogative, as illustrated in (10) (from Ward, Birner & Huddleston 2002: 1385):

(10) Also a nice woman is our next guest […]
*Isn’t also a nice woman our next guest?

As the interrogative is not legal, a nice woman cannot be the subject of the sentence, and therefore has to be an inverted element (cf. Ward, Birner & Huddleston 2002: 1385).

However, tests of this kind quickly become unwieldy with any larger corpus. For instance, COHA contains more than 1600 sentence-initial combinations of another + noun. Obviously, a manual inspection of all potential candidates of NP-inversion is hardly practical. This problem does not arise with fronted adjective phrases.

Second, NP-inversion is about as common as locative inversion, at least in non-fictional genres (Kreyer 2006: 121–123). It has also received some recent discussion in terms of its discourse function, such as its text-structuring role (Kreyer 2006: 161–192; Prado-Alonso 2011). By contrast, adjective phrase inversion has appeared as a relatively infrequent construction in previous studies and has received little attention in its own right. It is therefore a research category that on the one hand can be identified with sufficient precision as to be available for automatic extraction from a corpus, and which is, on the other hand, a construction whose discourse function is still relatively unexplored. All this makes it an ideal candidate for investigation. We will now discuss the patterns that we expect for it in the light of the discourse function of the sentence-initial position, to which the adjective phrase moves.

2.2 The role of the left clause periphery in inversion

The long-term history of inversion with fronted adjective phrases is part of the larger rise of inversion following fronted predicative complements. Previous historical studies see this rise in close correspondence with the presentational nature of the construction, i.e. with its function “to introduce referents” (Brinton & Stein 1995: 39) or “a new subject on the scene” (Bækken 2000: 416). However, in contrast to the situation in Old English, the new types of inversion which emerge with the Early Modern English period have the fronted constituent in a non-canonical position and due to this receive their “bifocal structure” (Brinton & Stein 1995: 39). This property suggests that in these inversions, the sentence-initial element deserves an analysis in its own right.

Most theories about the function of the left clause periphery agree in the claim that it is the position in the sentence for subjective elements and “discourse-coherence markers” (Traugott 2012: 7). To a certain extent, this function arises quite naturally from sentence beginnings being “utterance launchers” (Biber et al. 1999: 1073), in the sense of the physical point of departure of the upcoming clause. Such launching-devices can be extra-expressions, i.e. real prefaces added to the locution of the clause, but they also arise from topicalizations by way of fronting, i.e. through word order giving thematic prominence to a sentence element (Biber et al. 1999: 1074).

The adjective phrase triggering an inversion thereby carries a bi-fold function: as a verbal complement it constitutes an essential component of the predication, but as an utterance launcher it connects the new sentence to the preceding discourse. Elements in that position are prone to presenting a personal stance (Biber et al. 1999: 1003) or serving discourse connectivity, which is why they often have an “ambivalent grammatical status” (Biber et al. 1999: 1076). For the same reason, they undergo developments of subjectification (Detges & Waltereit 2011; Traugott 2010, 2012), which means, in the case of less central sentence elements, that they become pragmaticalized (Diewald 2011; Heine 2013). Well-known cases are the development of in fact, no doubt, or surely, which have developed out of a clause-internal element into a sentence adverbial and, ultimately, a discourse marker (Traugott 1997; Degand & Simon-Vandenbergen 2011).

Adjective phrase inversion follows the fronting of a verbal complement and not of an adjunct, but one might expect that the elements now positioned in the left periphery also prefer patterns that are in accordance with the discourse functions of that position. One might thus predict that the processes of change that the construction undergoes are in line with the discourse-pragmatic functions of expressing connectivity and/or subjectivity. It is against this background that we hypothesize for our corpus study that non-locative inversion, limited here to adjective phrase inversion, shows a functional specialization in its development over the last 200 years. More specifically, we predict that the construction favors adjective phrases with a connective function or subjective meaning as the preferred pattern. The next section will introduce how we tested this prediction.

3. Methodology

3.1 Data

The instances of non-locative inversion investigated in this study were obtained from the Corpus of Historical American English COHA (Davies 2010–). The corpus covers about 400 million words from books of fiction (about 50 percent) as well as magazines, newspapers, and non-fiction books. The earliest texts are from the 1810s, and the most recent texts are from 2009. The bar chart in Figure 1 displays the distribution of words across the decades from 1810 to 2000. The average number of words per decade in the whole corpus is 20.3 million words. As Figure 1 shows, the text size for the first decade is much smaller than for the remaining decades (1.2 million words). In order to avoid possibly skewed results due to this relatively small number of words, texts from that decade were not considered in this study.

Figure 1. Distribution of words per decade in COHA.

Figure 1. Distribution of words per decade in COHA.

The number of words in the non-locative adjective phrases that undergo inversion is variable, and probably limited only by functional constraints, as examples (11) (one word) and (12) (four words) illustrate.

(11) Outstanding was the visit to the grave by Austrian officials and ten German mayors, (MAG_Time_1928)
(12) Perhaps even more striking are the conditions in the North Italian Plains, (NF_NewRegionalGeography_1929)

For the purpose of the present study, the scope was delimited to adjective phrases that consisted only of a head adjective preceded by a single modifying non-locative adverb (e.g. so true, marvellously spectacular, equally galling). Other patterns of non-locative AdjP-inversion were not investigated. This had practical reasons: in the case of adjective phrases without a modifying adverb as in (11), a semantic categorization of the fronted AdjP would have to depend on the meaning of the adjective alone, which is highly variable (for a brief look at the extent of this variability, see the end of the next section). Thus, it seems to be extremely taxing to attempt to disentangle the source of subjectivity in these cases. In the case of adjective phrases with more than one modifying element as in (12), a semantic classification was not applicable.

In addition to this constraint on the number of modifying adverbs, the adjective phrases had to occur sentence-initially, and had to be followed by a word-form of be in order to be considered an instance of non-locative inversion. The resulting three query strings together with each number of hits in COHA are given in (13). All queries start with a full stop in order to ensure that the sequence occurs sentence-initially. The query tokens in squared brackets refer to the part-of-speech labels from the UCREL CLAWS7 tagset. [4] [RR*] and [RG*] query for general and degree adverbs, respectively, and [J*] queries for any adjective. [be] queries for all word-forms of be. The last query string was necessary because sentence-initial most and least appear to be tagged as “superlative after-determiners” in COHA (CLAWS7 tag [DAT]).

(13) . [RR*] [J*] [be] 3027 hits
. [RG*] [J*] [be] 1525 hits
. most|least [J*] [be]   147 hits

The initial data set contained 4699 hits. These hits were manually screened for words that are incorrectly tagged in the corpus as adjectives, such as minute-knowledge or dry-cleaning. 39 hits were removed due to tagging errors. An additional set of 17 hits was removed because the sequence did not represent an adjective phrase in that context, as the sentence-initial perhaps important was in Perhaps important was not the right word: although everyone went to the Bonnelles (COHA MAG_IvoryMischief_1942).

Also excluded were hits that contained either however, how, maybe, or yet as the initial adverb, as these adverbs can not only modify adjectives, but are also very frequently used as sentence-modifying adverbs. As the highlighted example in (14) illustrates, it is not always unambiguously possible to determine the target of the adverb.

(14) we feel that he has fallen from his place as a vital member of the body politic to one of paltry insignificance. Yet stronger is this feeling with reference to those who forsake useful labor for petty government offices (MAG_NorthAmRev_1852)

For the present analysis, a conservative approach was taken, in which the adverbs that can also function as sentence-modifiers were categorically excluded. This reduced the data set by another 1077 hits to the final number of 3566 validated sentence-initial sequences either of the form ADVERBgeneral ADJECTIVE or ADVERBdegree ADJECTIVE, followed by any form of the verb be. Some examples are given in (15)–(19), illustrating the variability of the attested adverbs and adjectives.

(15) […] and to wait[sic] the developments of the succeeding day. Utterly fruitless were all the means used by Claire […] (FIC_TrueRiches_1852)
(16) Usually the attitude to the father-in-law is cordial or neutral. Especially difficult is the situation where the mother-in-law is dominating […] (NF_EngagementMarriage_1953)
(17) […] on featureless-faced mannequins rough-hewn of pinkish beige plaster, some as disproportioned as surrealism. Barely practical are the clothes shown by Paris conservatives […] (MAG_Time_1937)
(18) […] had effects, direct or indirect, on almost every sphere of life. Most important were the industrial and social effects. (NF_MakingModernBritain_1956)
(19) Top sellers include a CD/radio/alarm device, DVD players and digital cameras. Also popular is the novelty Q-Ball toy, which sends out […] (NEWS_Denver_1999)

3.2 Classification

After the data set was compiled and validated, the 3566 adjective phrases were semantically coded using the classification scheme given in Table 1. As suggested by the query strings above in (13), all phrases can be subsumed under the two main types of Degree and General, which each contain several semantic categories as subtypes. The adjective phrases were assigned to these semantic categories on the basis of the semantics of the modifying adverb. The semantic categories of the Degree main type closely follow the description in Mittwoch, Huddleston & Collins (2002: 720–724) for clause adjuncts. The General main type encompasses various semantic categories that are not considered to express degree relations.

Main type Semantic category Description Examples of adverbs
Degree maximal specifies a degree at the upper extreme of a scale totally, utterly, …
multal specifies a degree between the midpoint and the upper extreme of a scale frequently, strongly, ...
moderate specifies a degree somewhere around the midpoint of a scale quite, rather, …
paucal specifies a degree in the lower range of a scale barely, a little, …
minimal specifies a degree close to the lower extreme of a scale hardly, scarcely, …
approxi­mating indicates proximity to a degree scale almost, nearly, …
comparing specifies a degree in comparison to another point of reference more/most, less/least, equally, …
General descrip­tive specifies a property derived from a descriptive adjective brightly, socially, …
linking relates adjective phrase to preceding proposition also
stance expresses speaker attitude admirably, painfully, …
temporal specifies time eventually, still, ...

Table 1. Semantic categories used to classify the adjective phrases.

For example, the initial adjective phrase very affectionate in the sentence Very affectionate was their parting embrace for the night; and Carrie Sinclair did not sleep for a long while (FIC_EutawASequelThe_1856) was coded as “multal” because the adverb very specifies a degree that is well between the midpoint and the extreme upper end of the affectionateness scale. The semantic category “maximal” may be considered as competing with “multal”. Yet “maximal” is ruled out because very affectionate does not place the referent at the extreme end of an affectionateness scale, not as much as, for instance, completely affectionate. In this manner, most of the modifying adverbs of the adjective phrases in the data set could be unambiguously assigned to exactly one of the semantic categories in Table 1. Consequently, all adjective phrases containing these adverbs were coded as belonging to the semantic category associated with the adverb. The only exception was still, which can be either relative (e.g. Still greater was [...]) or temporal (e.g. Still unexplained was […]). For the 45 adjective phrases in the data set containing this adverb, the classification was done on a case-by-case basis.

As outlined in section 2.2, semantic categories which have a connective function or which are associated with a subjective meaning are particularly interesting with regard to our hypothesis. The connective function is mostly evident in adjective phrases coded as “linking” or “comparing”: sentences starting with relative adjective phrases such as equally important is or more serious are involve a comparison between the preceding and the new proposition. In sentences starting with an adjective phrase, such as also present is or also important is, the new proposition is explicitly linked to the preceding one. In both cases, the adjective phrase implicitly or explicitly connects the preceding and the new proposition. Note that also is the only occurring “linking” adverb in our data set.

Adjective phrase types that are particularly associated with speaker subjectivity are those types which involve a high degree of speaker commitment to the property expressed in the adjective phrase and which thereby indicate the stance or belief of the speaker with regard to that commitment. For the purpose of our hypothesis, we considered “maximal”, “multal”, and “minimal” as categories linked with the highest degree of subjectivity: by using an adjective phrase of the first three categories in Table 1, something is placed at or close to the extreme ends of a scale, which requires a higher degree of speaker commitment. In this sense, the maximal adjective phrase totally enjoyable is considered to be more subjective than the paucal adjective phrase somewhat enjoyable, since the maximal adjective phrase places something at the extreme end of the enjoyability scale. [5] This requires a stronger commitment of the speaker to the truth value of the proposition than the one with the second, much more vague adjective phrase.

Semantic category N %
multal (subjective) 2544 71.3
comparing (connective) 709 19.9
linking (connective) 112 3.1
maximal (subjective) 77 2.2
moderate   38 1.1
temporal   28 0.8
stance   21 0.6
descriptive   15 0.4
approximating   11 0.3
minimal (subjective) 7 0.2
paucal   4 0.1
Total   3566 100.0

Table 2. Distribution of semantic categories in the data set. Categories with a subjective meaning or a connective function are highlighted in blue and orange, respectively.

There are considerable frequency differences between the semantic categories. As Table 2 shows, multal adjective phrases are by far the most frequent category, accounting for 71.3% of all observations in our data set. Also fairly frequent categories are comparing, linking, and maximal phrases (25.2% altogether), while the remaining categories occur in less than 3.5% of the instances. In the following analysis, we will focus on those semantic categories which are labelled either as “subjective” or “connective” in the table, as these are the categories for which we predict an increasing number of observations. Incidentally, almost 97% of all observed cases of non-locative AdjP-inversion belong to one of these categories.

Before we proceed with the detailed analysis of the developmental patterns of these categories, it is worthwhile to have a brief and rather informal look at the distribution of the adjectives in our data set. All in all, 953 different adjective types occur in the 3566 inverted phrases. Table 3 shows the ten most frequent adjectives with their number of tokens, and the percentage relative to the total number of tokens in the data set. The most frequent adjective, great, accounts for 7.4% of the inverted adjective phrases, and taken together, more than a quarter of all observations (27.1%) can be attributed to the adjectives in Table 3. Of course, as the last column of the table with the total frequencies in COHA shows, all adjectives are generally rather frequent words. Yet, the COHA frequency does not appear to be sufficient to explain the rank of the adjectives in our data set. For instance, the COHA frequency of strong and present is very similar (about 73.500 times), but the former occurs about six times as frequently in an inverted adjective phrase than the latter. It may be promising to investigate whether this is simply due to different preferences of the adjectives to occur in predicative position, or whether the probability to occur in inverted position also reflects functional differences. Such an investigation, however, is outside of the scope of the present study.

Rank Adjective N % COHA frequency
1 great 264 7.4 324.993
2 strong 181 5.1 73.265
3 important 167 4.7 80.401
4 different 113 3.2 94.207
5 true 60 1.7 108.605
6 powerful 40 1.1 26.594
7 successful 38 1.1 24.062
8 serious 37 1.0 40.036
9 present 37 1.0 73.754
10 eager 32 0.9 16.468
  Total 969 27.1  

Table 3. List of the ten most frequent adjectives in the data set. Percentages are relative to the total number of inversions.

4. Analysis and results

Very often, the observable usage pattern of language features that undergo diachronic change cannot be well described by a model with strict parameterization such as linear models, which basically assume that the observed variable increases or decreases monotonously if the independent variables change. As we do not have a hypothesis of the kind of change that we can find in the use of non-locative AdjP-inversion, we opted for non-parametric models (Bowman & Azzalini 1997; R-package “sm”) to examine the relation between time and the frequency of occurrences. In effect, these models abstract over the observed frequency to provide an estimated frequency, levelling out small local changes that may be a random result of the data. The models also calculate a variability band by adding the size of two standard errors above and below the estimated frequency (cf. Bowman & Azzalini 1997: 76). The variability band is conceptually similar to confidence intervals in models that estimate single means, and can be interpreted as an indication of the degree of estimation certainty. A frequency estimation that is based on only few observations (e.g. those at the beginning or the end of the observation period), or that is based on observed values that show a large variability, will co-occur with a relatively wide variability band, because the estimation certainty is relatively low.

The histogram in Figure 2 shows the overall distribution of occurrence of non-locative adjective phrase inversion by decade in COHA. Each bar corresponds to the number of observations per million words in the sub-corpus for that decade. The thick line in the figure is the regression line from the non-parametric model (i.e. the estimated frequency), together with the variability band.

Figure 2. Occurrences per million words of non-locative AdjP-inversion by decade in COHA with non-parametric regression line.

Figure 2. Occurrences per million words of non-locative AdjP-inversion by decade in COHA with non-parametric regression line.

A visual inspection of the figure suggests that the number of inversions remained fairly constant for the first ten decades, but started to decrease after the 1910s. The statistical significance of this pattern was investigated by testing the non-parametric regression model against two reference models: a “no effects” model that assumes that there is no statistically significant relation between the temporal axis and the number of occurrences per million words, and a “linear model” that assumes that the number of occurrences per million words changes for each decade by a constant factor. Each test estimates the probability that there is no statistically significant difference between the regression model under investigation and the reference model.

Both tests are significant for the regression model shown in Figure 2 (“no effects”: p=0.025, “linear”: p=0.046). Apparently, the overall frequency of non-locative inversions has changed notably in COHA over the last 200 years, and the majority of this decrease took place during the 20th century.

Yet, as Figure 3 illustrates, an analysis of the different semantic categories shows that the development is more fine-grained than this general description suggests. The three semantic categories associated with a subjective meaning (multal, maximal, minimal) are given in blue in the first row and the two linking categories (comparing, linking) are given in orange in the second row. The lower right panel shows the distribution for all other categories. The non-parametric regression line is superimposed over each panel.

Figure 3. Occurrences per million words of non-locative AdjP-inversion by semantic category and decade in COHA.

Figure 3. Occurrences per million words of non-locative AdjP-inversion by semantic category and decade in COHA.

Three of the six non-parametric regression models differ significantly from their “no effect” and “linear” reference models: multal (p=0.019 and p=0.040), comparing (p=0.018 and p=0.027), and linking (p=0.019 and p=0.026). The remaining three models (maximal, minimal, and other) differ only marginally from the “no effect” model.

The figure reveals that the decrease in total frequency of non-locative AdjP-inversion found in Figure 2 is mostly due to the development of the multal subtype in the upper left panel. For the first ten decades, this category was the dominant inversion pattern, with up to 9.24 occurrences per million words. After the 1910s, however, the frequency of the multal type decreases drastically to 2.04 occurrences per million words in the 2000s. The regression line does not show any indication of decelerated decrease.

A very different developmental pattern is found for the comparing subtype in the lower left panel. This category is present right from the beginning of the observation period, although with only 0.34 occurrences per million words for the 1820s. The frequency gradually increases throughout the following decades, with an accelerated growth from the turn of the century onwards. This acceleration accounts for the significant deviation from the “linear” reference model. The estimated frequency reaches its peak for the 1970s and 1980s at about 2.70 occurrences per million words. For the last two decades, the frequency is estimated to decrease again. However, this estimation is associated with an increased uncertainty, which is evidenced by the widening variability band at the right edge of the regression line.

A third pattern still can be seen in the lower middle panel for the linking subtype. AdjP-inversions involving this semantic category are not attested at all for the first eight decades. Yet, with the turn of the century, this type emerges as a new pattern with continually growing frequency that is estimated as 0.92 occurrences per million words in the last decade.

In summary, the analysis of the frequency changes over the course of the observation period reveals a differentiated picture. First of all, multal is the only type of adjective phrase with subjective meaning that is attested with a notable number of occurrences in the corpus. The other two types (maximal and minimal) are very rare in inverted position, and there is no indication of a change for them. The distribution of the multal type, however, shows a clear developmental pattern. For a large part of the observation period, it was almost the only type of AdjP that was used with inversion. During the last century, it was used much less frequently and lost its predominant role. Conversely, both categories with a connective function (comparing and linking) show a notable increase in frequency, particularly from the turn of the 19th to the 20th century onwards.

Apparently, the functional specialization that we find in our data is one in which the connective function of AdjP-inversion strongly gains in importance, while AdjP-inversion is used less and less frequently to express subjectivity. This reversal is illustrated in Figure 4.

Figure 4. Non-parametric analysis of covariance testing equality of multal and comparing subtypes.

Figure 4. Non-parametric analysis of covariance testing equality of multal and comparing subtypes.

For this plot, the three subjective types were combined (blue estimate), as were the two connective types (orange estimate). The grey band is the variability band of an “equality” model. This model tests the assumption that there is no significant difference between the subjective and the connective types. For the most part of the observation period, the two estimates fall clearly outside the grey band. Starting with the 1970s, however, the grey band and the other two estimates overlap. This means that the “equality” model cannot detect a significant difference between the two types any more – apparently, the subjective and connective functions are equally frequent in the subcorpus spanning the last decades.

5. Discussion and conclusion

The results of our corpus study presented in the previous section confirm, as predicted in section 2, that the development of inversion following fronted adjective phrases shows a preference for adjective phrases with a connective function or subjective meaning. The instance of the re-newed, non-locative type of inversion that we investigated indeed favours collocation with subjective degree and linking adverbs, rather than adverbs from the other semantic categories. As the preference for the subjective elements in the AdjP is stronger in the earlier decades, whereas the preference for connective items is stronger in more recent times, our findings also confirm a general grammaticalization path from a more pragmatic to a more grammatical discourse function.

The investigation of this development provides us with a particularly fine-grained look at a process of language change. This approach and the results that it has yielded have two advantages: firstly, they cover only a period of fewer than 200 years, which seems to be comparatively short for a study of word order variation in English. However, where broader temporal perspectives save historical linguists “from the academic uncertainties of the present” (Labov 1994: 10), an analysis of data from a more limited time span grasps speakers’ preferences in language use and thus tends to document changes in progress (e.g. Kiesling 2011; Maguire & McMahon 2011).

Secondly, the confined scope of our corpus study (non-locative inversion of adjective phrases) allowed us to focus on the historical development of a non-canonical construction of present-day English that so far has received only little attention in the literature. Our findings about this non-locative type of inversion are in line with the general assumption that a “bi-focal”, non-deictic type of inversion has become available some time after the Early Modern English period (Brinton & Stein 1995; Nevalainen 1997; Stockwell 1984). We have also been able to show that speakers recruited this construction for various purposes once it had become available, and that these purposes have in turn evolved over time. Since a bi-focal construction strengthens the role of the sentence-initial element these purposes arise out of the discourse-pragmatic function typical of the left clause periphery. This confirms the notion expressed in Kiesling (2011: 172) that one source of language change is the adoption of existing variability because of its expressiveness.

Such a purposefully fine-grained analysis provides a more differentiated picture of what previous work termed the “functionalization of a focusing strategy” for the sentence-initial element (Brinton & Stein 1995: 39). Due to the strengthened role of the left sentence periphery in a bi-focal construction, we could assume that adjective phrase inversion would follow a path of functional specialization. In line with the pragmatic functions of sentence beginnings to express both connectivity and subjectivity, we expected non-locative inversion to favor adjective phrases with exactly those kinds of meaning. What the results confirm is indeed an early prominence of subjective meanings, which were frequent in the 19th century, but occurred with decreasing frequency during the last century. There is no indication that this decrease will reverse again.

By contrast, adjective phrases with a connective function have generally been increasing in frequency over the last 100 years, with some divergent patterns for different elements expressing connectivity. Comparing adjective phrases, which constitute an implicit link to the previous discourse (more important = more important than X), were rare at first and gradually increase in frequency – a development that may at first glance resemble the shape of an S-curve often used to describe the course of language change (e.g. Kroch 1989; see Tagliamonte & D’Arcy 2009 for a recent review and discussion). However, the last decades have seen no further increase in the number of occurrences of comparing adjectives, and there may even be a change of trend in the future. It would be difficult to implement such a change of trend in the cumulative models of language change that produce an S-curve.

The other subtype of AdjP that expresses connectivity, the one which uses the explicit linking element also, as in also important is…, shows a developmental path that is related to the comparing subtype, but with some notable differences. Non-locative inversion involving this linking element was not attested in the 19th century, during which the comparing type was already in use, even if the frequencies were relatively low. Both subtypes show a notable increase in frequency at roughly the same time, i.e. at the turn of the 19th century, which suggests that there are at least similar, if not constant, rates of change for these two subtypes of AdjP-inversion. Following the constant rate hypothesis, i.e. the assumption that an ongoing linguistic change affects all available linguistic contexts at the same rate (see Kroch 1989 for a detailed discussion), this would suggest a systematic shift of the function of non-locative AdjP-inversion away from an expression of subjectivity to an expression of connectivity, which affects all subtypes at roughly the same time with similar rates of change.

The sequence in which these developments have occurred suggests that the subjectivity inherent in the left periphery of the clause has been the possible initiator of the original change, but that there has then been a change in the functional preferences for the fronted type of constituent followed by non-canonical word order. If one interprets the connective function as the more (discourse-) grammatical one, while subjectivity is more exclusively pragmatic, such a path of the development parallels the “layering” of grammaticalization processes from lexical to pragmatic to grammatical functions (Kiesling 2011: 159; see also Hopper & Traugott 2003: 124–125).

What the nature of our findings also confirms is that observations on language change and the patterns they describe largely depend on prior decisions of categorization. Much like full inversion having followed a different path of development from subject-auxiliary inversion in the history of English, non-locative inversion has later turned up as an innovative, functionally modified reordering strategy. Our analysis has shown that there are divergent developmental patterns even within this strategy.

More generally speaking, our findings agree with the overall patterns of developments found in Modern English. Leech et al. (2009: 8) note that “relatively few categorical losses or innovations have occurred in the last two centuries”, but that the syntactic changes found are “statistical in nature” instead. This notion is in line with the increasing body of evidence that diachronic changes are generally affected by a host of frequently non-categorical factors (see, for example, Gries & Hilpert 2010; Wolk et al. 2013), which has led to the development of frameworks like Diachronic Probabilistic Grammar (DPG, Szmrecsanyi 2013) that explicitly set out to model the effect of probabilistic constraints on the choice of past language users to explain historical developments. By extending the frequency analysis with a functional categorization, the present analysis of the non-locative AdjP-inversion is compatible with these frameworks. The data from COHA suggest that there are certain functional preferences, which result in different probabilities, and these preferences are still shifting, but not to the extent that a function that was once preferred has categorically disappeared from the language. The functional preferences draw on syntactic (i.e. positional) as well as collocational properties of the different types investigated. As noted elsewhere, such “co-occurrences”, especially with adverbs, in syntactic change often provide us with “clues about how the constructions are functionally distributed” (Kiesling 2011: 161). As with our (sub-) categorization of non-locative inversion following the semantic type of adverb, this brings a semantic component into the study of syntactic change.

Based on the present study, what is our response to the guiding question raised in this special issue, the question whether language change is predictable? It seems that this question entails two distinct aspects. First: is it possible to predict that a particular feature of the grammar of the language will change? Second: is it possible to predict the future development of an ongoing linguistic change?

While the first aspect may be still fairly difficult, we believe that analyses like the present demonstrate that it can indeed be possible to make a plausible prediction on the developmental path of some shifts in the grammar of a language, at least for the near future. However, in line with approaches such as DPG (Szmrecsanyi 2013), these attempts have to go beyond an exclusive focus on the change of frequency counts. The case of non-locative AdjP inversion has shown that the quality of our models and the range to which it is possible to make such a prediction strongly depends on choosing a suitable scale of classification: the model underlying Figure 2, which does not distinguish different subtypes of adjective phrases, might be useful to describe the overall shape of the frequency shift of non-locative AdjP inversion in a historical perspective. Yet, given that this shape was for many centuries determined by the dominant subtype containing “multal” adverbs, which has lost its dominance during the last few decades and is now on a par with the “comparing” subtype, it will not permit the prediction of the future development of this construction with acceptable precision. In contrast, the more fine-grained separate models underlying Figure 3 can capture the divergent developmental patterns of the AdjP subtypes, and allow us a look into the future of this ongoing linguistic change.

A similar example can be found in Mondorf (2009: ch. 10), who traces the historical development of the English comparative alternation. An analysis that does not distinguish between different adjective subtypes suggests that there is a general decline of the analytic more comparative in favour of the synthetic comparative using -er. However, this trend appears to be largely attributable to the development of only two types of adjective, namely monosyllabic adjectives and adjectives ending in -y, which together account for over half of the adjectives in Mondorf’s historical corpus. Other subtypes of adjectives do not follow this trend, or may show divergent developments. For example, the comparative of adjectives in -al (e.g. formal, special) is, and has been, almost exclusively formed by using more; and adjectives ending in -ere (e.g. severe, sincere) have developed a preference for the analytic comparative only after the beginning of the 20th century (Mondorf 2009: 126–158).

To sum up, what studies using historical corpora such as Mondorf (2009) and the present study illustrate is that a model that is based on an unsuitable generalization may fail to correctly predict the more nuanced developments. It appears to be necessary to find a useful degree of specificity. Once that the subtrends of an ongoing process of language change have been adequately identified, these smaller pieces can be put together to build a larger model, which may be able to predict the course of an ongoing language change at least to a satisfactory degree.

Notes

[1] All examples are taken from the COHA corpus (Davies 2010–) if not indicated otherwise.

[2] Of course, (2) could also be Even more interesting [than this] is that he thinks

[3] Non-locative inversion in principle also includes inversion following prepositional phrases when they refer to a property rather than a location. These cases are also limited to the verb be. An example of a non-locative PP-inversion is: Of greater concern to many New Englanders is the loss of the coastal passenger lines. (NEWS_CSMonitor_1948). Since PPs of that kind usually also contain an adjective, AdjP-inversion is the more prototypical representative of non-locative inversion.

[4] See a link to the CLAWS tagset in the Sources section.

[5] The category “paucal”, as Mittwoch, Huddleston & Collins (2002: 720–724) use it, has items of variable specificity (barely, but also somewhat), which made it more difficult to classify in terms of the subjectivity of the ranking involved. We decided against considering it as generally subjective.

Sources

CLAWS 7 tagset: http://ucrel.lancs.ac.uk/claws7tags.html

COCA = Davies, Mark. 2010–. The Corpus of Historical American English: 400 million words, 1810-2009. Available online at http://corpus.byu.edu/coha/

Datafile of the corpus and R file used in the analysis in this article.

References

Bækken, Bjorg. 2000. “Inversion in Early Modern English”. English Studies 81(5): 393–421.

Biber, Douglas, Stig Johansson, Geoffrey Leech, Susan Conrad & Edward Finegan. 1999. Longman Grammar of the English Language. Harlow: Pearson.

Birner, Betty. 1992. The Discourse Function of Inversion in English. Ph.D. dissertation, Northwestern University.

Birner, Betty. 1994. “Information status and word order: An analysis of English inversion”. Language 70(2): 233–259.

Bowman, Adrian W. & Adelchi Azzalini. 1997. Applied Smoothing Techniques for Data Analysis: The Kernel Approach with S-Plus Illustrations. Oxford: Clarendon.

Bresnan, Joan. 1994. “Locative inversion and the architecture of Universal Grammar”. Language 70(1): 72–131.

Brinton, Laurel & Dieter Stein. 1995. “Functional renewal”. Historical Linguistics 1993: Selected Papers from the 11th International Conference on Historical Linguistics, ed. by H. Andersen, 33–47. Amsterdam/Philadelphia: John Benjamins.

Chafe, Wallace L. 1992. “Immediacy and displacement in consciousness and language”. Cooperating with Written Texts: The Pragmatics and Comprehension of Written Texts, ed. by Dieter Stein, 231–255. Berlin: Mouton.

Chafe, Wallace L. 1994. Discourse, Consciousness, and Time. Chicago: The University of Chicago Press.

Chen, Ron. 2003. English Inversion: A Ground-before-Figure Construction. Berlin: Mouton de Gruyter.

Degand, Liesbeth & Ann-Marie Simon-Vandenbergen. 2011. “Grammaticalization and (inter)subjectification of discourse markers”. Linguistics 49(2): 287–294.

Detges, Ulrich & Richard Waltereit. 2011. “Turn-taking as a trigger for language change”. Rahmen des Sprechens. Beiträge zur Valenztheorie, Varietätenlinguistik, Kreolistik, Kognitiver und Historischer Semantik. Peter Koch zum 60. Geburtstag, ed. by Sarah Dessi Schmidt, Ulrich Detges, Paul Gévaudan, Wiltrud Mihatsch & Richard Waltereit, 175–190. Tübingen: Narr.

Diewald, Gabriele. 2011. “Pragmaticalization (defined) as grammaticalization of discourse functions”. Linguistics 49(2): 365–390.

Dorgeloh, Heidrun. 1997. Inversion in Modern English: Form and Function. Amsterdam/Philadelphia: John Benjamins.

Dorgeloh, Heidrun. 2006. “Inversion in descriptive and narrative discourse: A text-typological account following functional principles”. Cahiers de Recherche 9: 101–114.

Drubig, Hans Bernhard. 1988. “On the discourse function of subject verb inversion”. Essays on the English Language and Applied Linguistics on the Occasion of Gerhard Nickel’s 60th Birthday, ed. by Josef Klegraf & Dietrich Nehls, 83–95. Heidelberg: Julius Groos.

Gries, Stefan Th. & Martin Hilpert. 2010. “Modeling diachronic change in the third person singular: A multifactorial, verb- and author-specific exploratory approach”. English Language and Linguistics 14(3): 293–320.

Heine, Bernd. 2013. “On discourse markers: Grammaticalization, pragmaticalization, or something else?” Linguistics 51(6): 1205–1247.

Hopper, Paul J. & Elizabeth Closs Traugott. 2003. Grammaticalization. Cambridge: CUP.

Huddleston, Rodney & Geoffrey Pullum. 2002. The Cambridge Grammar of the English Language. Cambridge: CUP.

Kiesling, Scott F. 2011. Linguistic Variation and Change. Edinburgh: Edinburgh University Press.

Kreyer, Rolf. 2006. Inversion in Modern Written English: Syntactic Complexity, Information Status and the Creative Writer. Tübingen: Narr.

Kroch, Anthony. 1989. “Reflexes of grammar in patterns of language change”. Language Variation and Change 1(3): 199–244.

Labov, William. 1994. Principles of Linguistic Change: Internal Factors. Oxford: Blackwell.

Leech, Geoffrey, Marianne Hundt, Christian Mair & Nicholas Smith. 2009. Change in Contemporary English: A Grammatical Study. Cambridge: Cambridge University Press.

Los, Bettelou. 2009. “The consequences of the loss of verb-second in English: Information structure and syntax in interaction”. English Language and Linguistics 13(1): 97–125.

Maguire, Warren & April McMahon, eds. 2011. Analysing Variation in English. Cambridge: Cambridge University Press.

Mittwoch, Anita, Rodney D. Huddleston & Peter Collins. 2002. “The clause: Adjuncts”. The Cambridge Grammar of the English Language, ed. by Rodney D. Huddleston & Geoffrey K. Pullum, 663–684. Cambridge: Cambridge University Press.

Mondorf, Britta. 2009. More Support for More-Support. Amsterdam/Philadelphia: John Benjamins.

Nevalainen, Terttu. 1997. “Recycling inversion: The case of initial adverbs and negators in Early Modern English”. Studia Anglica Posnaniensia: An International Review of English Studies 31: 203–214.

Prado-Alonso, Carlos. 2011. “Text structuring in written English: The role of inversion”. English Studies 92(4): 449–463.

Prado-Alonso, Carlos & Juan Carlos Acuña Fariña. 2010. “A comprehensive account of full-verb inversion in English”. Folia Linguistica 44(2): 509–554.

Stockwell, Robert. 1984. “On the history of the verb-second rule in English”. Historical Syntax, ed. by Jacek Fisiak, 575–592. Berlin: Mouton.

Szmrecsanyi, Benedikt. 2013. “Diachronic probabilistic grammar”. English Language and Linguistics 19(3): 41–68.

Tagliamonte, Sali A. & Alexandra D’Arcy. 2009. “Peaks beyond phonology: Adolescence, incrementation, and language change”. Language 85(1): 58–108.

Traugott, Elizabeth Closs. 1997. “The role of the development of discourse markers in a theory of grammaticalization”. Paper presented at ICHL XII, Manchester 1995. http://www.wata.cc/forums/uploaded/136_1165014660.pdf (PDF)

Traugott, Elizabeth Closs. 2010. “(Inter)subjectivity and (inter)subjectification: A reassessment”. Subjectification, Intersubjectification and Grammaticalization, ed. by Kristin Davidse, Lieven Vandelanotte & Hubert Cuyckens, 29–71. Berlin: Mouton.

Traugott, Elizabeth Closs. 2012. “Intersubjectification and clause periphery”. English Text Construction 5(1): 7–28.

Ward, Gregory, Betty Birner & Rodney D. Huddleston. 2002. “Information packaging”. The Cambridge Grammar of the English Language, ed. by Rodney D. Huddleston & Geoffrey K. Pullum, 1363–1449. Cambridge: Cambridge University Press.

Warner, Anthony. 2007. “Parameters of variation between verb-subject and subject-verb order in late Middle English”. English Language and Linguistics 11(1): 81–111.

Wolk, Christoph, Joan Bresnan, Anette Rosenbach & Benedikt Szmrecsanyi. 2013. “Dative and genitive variability in Late Modern English: Exploring cross/constructional variation and change”. Diachronica 30(3): 382–419.