In this stimulating book, Hilpert combines the theoretical approach of Construction Grammar with sophisticated corpus methodology to investigate the processes of language change. A range of different corpora and statistical methods are used in three case studies in the history of English, each addressing a different level of language structure: allomorphy, derivational morphology and syntax. Through this range of studies, the author goes beyond the syntax focus of most Construction Grammar research to develop a broader perspective on constructions and the diachronic processes in which they participate.
The introductory chapter provides theoretical background. Hilpert develops the concept of ‘constructional change’, defining it as change which selectively affects a single construction (a conventionalized pairing of form and meaning), altering its form, function, frequency or distribution in the linguistic community. He distinguishes it on the one hand from more pervasive, systemic processes of language change (such as a general change from OV to VO word order), and on the other from grammaticalization. Hilpert's definition is notably broad in the types of change encompassed. As well as changes in form and meaning (or function), he includes changes in frequency or in distribution across speakers or genres. With regard to frequency, Hilpert advocates corpus research that goes beyond a focus on simple text frequency to investigate other issues such as type frequency (a measure of the productivity of a construction) and the relative frequencies of functional and structural variants of a construction.
Chapter 2 focuses on data and methodology. It briefly surveys the corpora used in the empirical studies, before discussing in detail the quantitative corpus methods used and their application to the study of constructional change. A valuable feature here is the use of corpus linguistic studies to illustrate each method. These studies draw on a range of different corpora and cover a variety of different grammatical topics. This substantial, data-rich methodological chapter requires some effort from less quantitatively minded readers, but the clear exposition makes it a good tutorial.
The first important method introduced in the chapter is variability-based neighbour clustering (VNC), proposed by Gries & Hilpert (Reference Gries and Hilpert2008) as a way to identify the relevant periods of development in diachronic corpus data. Typically, diachronic corpora are divided into time periods which are arbitrary (e.g. fifty-year intervals) or reflect general stages of language development (e.g. Early Modern English) – but the stages of development of a particular construction need not match these periods. VNC is a type of clustering analysis which groups together adjacent periods where the data show more similar patterns. With this method, the analyst can use the data for a particular construction to motivate a division into relevant periods, rather than relying on a subjective analysis of a plotted trend of change. Further chapter sections discuss three different types of multivariate analysis (analysis involving multiple variables). Multivariate analyses are particularly useful for Hilpert's purposes, as they allow him to investigate the multiple dimensions along which constructional change can occur. Each of these methods will be described later in the review, in relation to the case study where it is used.
Chapters 3 to 5 present the three empirical case studies. These chapters contain substantial discussion of related research and issues of linguistic analysis, and sometimes present several different data analyses. This review will focus on the central multivariate analysis of each chapter, to highlight the potential of these methods and the interpretive and theoretical issues they raise.
Chapter 3 presents the first case study, on allomorphy, investigating the gradual replacement of the possessive determiners mine and thine by my and thy respectively. Previous corpus studies have shown the importance of phonological conditioning in the loss of the final /n/, as this loss occurred first in preconsonantal contexts (i.e. before a word beginning with a consonant). However, they have also shown that multiple factors, both social and structural, were relevant to the choice between the ‘n-variants’ and ‘n-less variants’ over time, making a multivariate analysis appropriate.
Hilpert examines changes in the relative frequency of n-less variants from 1150 to 1718, drawing data from the Penn–Helsinki Parsed Corpus of Middle English (Kroch & Taylor Reference Kroch and Taylor2000) and the Penn–Helsinki Parsed Corpus of Early Modern English (Kroch et al. Reference Kroch, Santorini and Delfs2004), which include various prose genres. He covers a longer time period than previous studies, including preconsonantal examples. He uses information about the year of text production to carry out a VNC analysis which periodizes the trends in a data-driven way. The analysis divides the data into seven periods of widely varying length (two being single-year clusters). The percentage of n-less variants shows an early sharp increase from about 5 per cent in period 1 (1150–1200) to about 47 per cent in period 2 (1225). After a modest dip in period 3, the increase of n-less variants continues, reaching about 96 per cent in period 7 (1516–1718). The VNC periods are then used to code for the factor of time in the multivariate analysis.
The multivariate method used is binary logistic regression (or more specifically, a recent refinement of that method called generalized linear mixed-effect modelling, which allows the inclusion of ‘random’ factors). Binary logistic regression is already a standard technique in the analysis of linguistic data (usually synchronic). Its purpose is to investigate the factors which influence the choice between two linguistic variants. It identifies both main effects of individual factors (or variables) and interaction effects between factors. In diachronic studies, any significant interaction between time and another factor is of particular interest in showing that the effect of that factor is not stable over time.
Hilpert finds that various factors influence the choice between n-variants and n-less variants, including phonological context, frequent collocations, priming and gender. The effect of some factors varies with time. This is true, for instance, of phonological context (whether the next word begins with a consonant, /h/, or a vowel). While a following consonant always increases the probability of an n-less variant, the effect is strongest in period 5 (1388–1450) and declines thereafter; the difference beween an /h/ and a vowel does not always have an effect, but is highly significant in periods 6 and 7 (with the /h/ contexts ahead of the vocalic ones in the loss of /n/).
No significant effect is found for grammatical person of the determiner (first versus second). Mine/my and thine/thy show very similar patterns of change over time. Hilpert therefore concludes that we can view this as a single constructional change, with speakers forming a generalization over these two members of the paradigm of possessive determiners. He argues that decisions over levels of generalization should be driven by the data: had the two showed different patterns of change, he would have analysed them as two different constructions. However, while the data certainly show evidence of a generalization by speakers across the two items, it seems strange to capture this generalization by treating them as a single construction, even at a more abstract level. The parallel patterns of change are presumably due to the close analogy on the levels of both form and meaning, and perhaps this kind of generalization should be captured in a different way.
In chapter 4, Hilpert presents his case study on derivational morphology, using data from the OED to investigate the historical development of the nominalizing derivational suffix -ment – or, rather, of the ‘V-ment construction’. This constructional schema has an abstract stem slot typically filled by a verb (or occasionally by a noun or adjective), and typically expresses the meaning of an action or the result, means or place associated with an action (exemplified by adjustment, assortment, refreshment, and abutment ‘junction’, respectively). The construction originated in French words containing the suffix which were borrowed into Middle English. By the fourteenth century, we start to see combinations of -ment with native English stems, indicating that speakers had formed a productive schema. Later, however, the productivity was lost, even though Present-Day English vocabulary still contains a very high frequency of types (different words containing the suffix).
Hilpert uses multivariate analysis to go beyond previous OED-based studies of this suffix, developing a finer-grained account which traces the history of different subschemas of the V-ment construction. He includes five linguistic variables in the analysis: etymological source (borrowed or derived), stem type (verb, adjective, or noun), semantic transitivity of the stem (transitive or intransitive), semantic type (activity, result, means, or place), and branching (binary, left, or right). The branching variable concerns morphological structure: the binary type involves simple stem + suffix (e.g. puzzle + -ment), the left-branching type involves prefixed stem + suffix (e.g. bedevil + -ment), while the right-branching type involves prefix + suffixed stem (e.g. non- + attachment). The derivation of this last type does not require use of a productive V-ment schema. The assignment of values to these linguistic variables is not always straightforward, and Hilpert includes discussion of criteria and difficult instances. The time period is also included as a variable; five development periods are identified in a VNC analysis based on productivity data for the construction.
The multivariate method applied is an exploratory one called Hierarchical Configural Frequency Analysis (HCFA). This is an extension of the chi-squared test that can be used to identify distinctive ‘types’ (or subschemas) of a construction, in the form of particular configurations of characteristics (values of the variables) that occur more often than expected. Diachronically, it identifies patterns typical of a particular period. In the V-ment study, the HCFA method allows Hilpert to describe the history of the construction in terms of ‘a sequence of typical formation patterns’ (p. 148). For instance, the earliest two periods, spanning 1250 to 1399, show only borrowed types, such as the configuration ‘verbal, means, transitive, binary, borrowed’ (e.g. payment, supplement). Only in period 3, 1400–1649, do derived types start to appear (though instances of derived forms occur earlier). The most frequent type in this period is ‘verbal, action, transitive, left-branching, derived’ (e.g. embitterment, renewment). This pattern accounts for almost a quarter of the forms in the whole database, so can be regarded as the prototype of the construction. It occurs as a type again in period 4, 1650–1899. Right-branching derived types (formed by prefixation) appear late, with one in period 4 and two in period 5, 1900–2000. These are the only types for the last period, suggesting that the suffix was no longer productive by the start of the twentieth century.
It is important to note that a type identified by HCFA need not be very frequent – it simply needs to occur more often than expected by chance at a particular time. For instance, dreariment and jolliment are two of six forms belonging to an infrequent derived adjectival type identified for period 3, apparently representing a passing fad. By my calculations, only about a third of the forms in the database are accounted for by the ten types identified in the HCFA analysis. Thus, as Hilpert acknowledges, there is also considerable variation. Many forms depart from the types, or subschemas, in some way. The results therefore need to be interpreted in association with other data, such as the single-variable trends that Hilpert plots for each of the five linguistic variables.
In summary, Hilpert shows that the HCFA method can be informative in identifying different subschemas of a construction which rise and fall in productivity over time. However, his study also shows the complexities of interpreting the results. It can be misleading to focus only on the significant configurational ‘types’ identified by the method, as this overlooks the variation in the remaining data. The study raises interesting questions about how variation around subschemas can best be analysed and described within a constructional approach.
Chapter 5 presents the case study on syntax, focusing on ‘concessive parentheticals’ (hereafter CPs) headed by although, though, while and if. Hilpert includes in this category non-finite or verbless clauses with concessive meaning, such as:
(1) But this belief, while steadfastly maintained in public, is becoming increasingly hollow.
(2) Although rare, family violence did occur.
He also allows for CPs which are syntactically embedded, as in (3), where the CP forms part of a structure modifying the head of a VP:
(3) The average has been steadily, although slowly, increasing with time.
The more compelling of Hilpert's two data studies of CPs addresses the issue of their development over time: whether we see a process of ‘levelling’ towards a ‘macro-construction’ (a more abstract, general constructional schema), or a process of dissimilation between related lower-level ‘mini-constructions’ which gradually become more individualized. To investigate this, Hilpert compares data from the last fifteen decades of the Corpus of Historical American English (COHA; Davies Reference Davies2010), which includes various written genres. He analyses a database of over 36,000 CPs in terms of twelve construction types based on the intersection of two structural dimensions. The first dimension concerns the syntactic form of the material following the conjunction, and has six values (e.g. AdjP, NP, past-participial clause). The second dimension concerns whether or not the CP is syntactically embedded, as in (3) above. Each example is also coded for conjunction (although, though, while or if) and decade.
For this study Hilpert uses an exploratory multivariate method called metric multidimensional scaling (MDS). This method identifies similarities between entities, typically reducing data to two main dimensions of variance, and transforms the differences into spatial distances which can be shown on a plot. In the present study of CPs, the MDS analysis compares the four conjunctions in terms of the frequencies of the twelve construction types in which they occur. Each conjunction is plotted as a ‘bubble’ on a two-dimensional plot or ‘MDS map’. The size of the bubble indicates the overall frequency of occurrence of the conjunction in CPs, and the distances between the four bubbles show how differently the conjunctions pattern in their distributions across the construction types.
An MDS map is obtained for each of the fifteen decades, and a comparison of these maps provides a rather effective way to visualize changes over time, though interpretation requires scrutiny of the numerical data for the twelve construction types. (Only the 1860s data are provided; it would have been helpful to provide the full data in an appendix.) The overall configuration remains basically similar over the decades. Although and though are close together at the start, and become more similar in usage over time. If is at first fairly near to though, but gradually moves further away, mainly because of a marked increase in occurrence over time with embedded structures, especially adjectives (e.g. an honest, if late, apology). While differs at the start by showing a high frequency of (non-embedded) -ing-participial clauses, and further diverges by increasing this preference. Thus, Hilpert concludes, we see assimilation between though and although over time, but dissimilation between although/though, if, and while. He argues that this combination of assimilation and dissimilation cannot be included under grammaticalization, as the latter is taken to involve development toward greater structural generality (e.g. through expansion of the class of elements with which an item can enter into construction).
Hilpert interprets these results as evidence against a single abstract CP construction. Instead he proposes a construction family with three different schemas (or ‘meso-constructions’), each schema having lower-level subschemas (frequent subtypes). It seems to me that this study raises several questions about the analysis of constructions which deserve further attention. First, how can we determine which dimensions of structure or function should be used to identify levels of constructional generalization? In the present study, for instance, linear position and type of concessive meaning are further dimensions of constructional variation discussed in this chapter as being relevant to CPs, but are not included in the MDS analysis. (This may be because of apparent interrelationships that Hilpert notes between the dimensions of position, meaning, and structural complexity (pp. 177–8). This complicates matters, but still leaves open the question of which dimension(s) should be considered criterial.) Second, to what extent can we use the frequencies of different subpatterns as criteria for differentiating constructions? For instance, if two forms occur in exactly the same range of patterns, but with somewhat differing distributions of frequency across these patterns, does this justify the identification of two different constructions? How different would the frequencies need to be for us to do so? Following a usage-based approach (e.g. Bybee Reference Bybee2010), Hilpert attributes to speakers probabilistic knowledge of distributions, but how fine-grained should we take this knowledge to be? There is a danger of ‘reifying’ frequency information gained from corpus distributions as a part of constructional representations. A similar issue arises if a social rather than a linguistic variable is considered, such as gender in the first case study. At a certain time, females may be using a higher proportion of a certain variant than males, but speakers may or may not be aware of this to the extent that they cognitively represent that variant as ‘associated with females’.
In the final chapter, Hilpert takes stock. He comments that the diachronic application of Construction Grammar is in its early days, and expresses the hope that his work will stimulate meaningful discussion of the concept of constructional change. This will surely be the case. Hilpert's wide-ranging empirical studies raise important theoretical and analytical issues concerning constructions and language change. They also showcase the use of various quantitative methods which allow the corpus linguist to exploit frequency data in more sophisticated ways. The book will provide thought-provoking reading for scholars and advanced students concerned with language variation and change, constructional approaches to language, and corpus methodology.
