It was with delighted anticipation (and some trepidation about my own competence in data science) that I began Axel Bohmann, Julia Müller, Mirka Honkanen and Miriam Neuhausen’s recent book Linguistic Data Science and the English Passive: Modeling Diachronic Developments and Regional Variation. Luckily, this excellent volume turned out to be so well written that my trepidation was unfounded, and I can heartily recommend it to anyone who is interested in the development of the passive voice in English, including those of us who still code our corpus data manually. It clearly explains how computational methods can be used to analyze enormous amounts of data and shows how automation can increase reliability and replicability – provided there are thinking human linguists at the helm. This book does not, in fact, shed much new light on the development of the English passive, as the authors themselves recognize (p. 216), but instead it mainly corroborates earlier, smaller-scale studies of the passive. Linguists should find this tremendously comforting as we enter an age of large language models and machine learning.
The aim of the authors, as described in the first chapter, is to develop and implement automated methods for retrieving and analyzing staggering numbers of be - and get-passives. Analyses of millions of tokens from the Corpus of Historical American English (COHA) and Global Web-based English (GloWbE) largely support previous claims about the diachronic development of the passive and the synchronic distribution of the passive across international varieties of English.
Linguistic Data Science and the English Passive is so reader-friendly that it is not necessary to go into the book equipped with either a thorough understanding of machine learning or a deep familiarity with the English passive. I find the overview of previous research on the English passive in chapter 2 to be so thorough and clear that I would highly recommend it as a good starting place for students and researchers who are new to the topic (‘The English passive voice: Be and get’, pp. 7–30). The authors begin with a comprehensive definition of what the passive voice is, followed by a presentation of the English be-passive (e.g. she was blamed) and its less-frequent cousin, the get-passive (e.g. she got blamed). Although be- and get-passives are presented as competing variants in the so-called ‘passive alternation’, the authors devote a great deal of time to discussing the constraints on the get-passive that have been proposed in the literature (and which indicate that get is not, in fact, a serious competitor to be, at least not at present). I particularly appreciate that the authors not only describe the proposed constraints on the get-passive, but also outline the methods and coding strategies that scholars have used to arrive at them.
Chapter 3, titled ‘Centrality and the passive gradient’ (pp. 31–46), details the authors’ efforts to automate the retrieval of true, central passives (those where the reading of the participle is purely verbal) from large datasets. The authors give a nuanced overview of previous methods and demonstrate that it is, in fact, possible to automate identification of central passives with satisfactory (if not perfect) results. I must admit that I had an easier time following the excellent discussion of the passive gradient, and a more difficult time following the description of the automated approach (especially table 4, which shows the aptly named ‘confusion matrix’ for the centrality classification). But this is assuredly due more to my own background than the authors’ competence.
The automated methods used in the studies are a main contribution of the book, and they are described in meticulous, replicable detail in chapters 4–6. Every consideration behind the classifications and automation thereof is explained and problematized in a way that will delight even the most pedantic reader. The nuance in the method discussions is particularly impressive. Subject animacy is (as often elsewhere) used as a measure of subject responsibility, which get-passives are supposed to confer upon their subject (i.e. saying she got fired is supposed to convey that she may bear some responsibility for her own firing as compared to the more neutral she was fired). It was a pleasure to read the careful consideration in chapter 4 of how fuzzy a category animacy truly is and how difficult its coding is to automate, to the point where the coding had to be carried out manually for the most frequent noun types (‘Operationalizing subject responsibility’, pp. 47–58). I do wonder about the decision to code for types and not tokens, but with data points in the millions it is an understandable sacrifice. The discussion in chapter 5 of how to operationalize coding for the supposedly adversative, or at least non-neutral, nature of get-passives compared to be-passives is conducted in similarly impressive detail (‘Operationalizing adversativity and non-neutrality’, pp. 59–72). Here, the authors’ reasonable conclusion that ‘sentiment analysis is a highly subjective endeavor with many ambiguous instances’ (p. 71) almost seems to beg the question of whether it is worth doing at all, whether manually or with a computer. But their results, when we arrive at them later in the book, do seem to tally with what we would expect from previous research.
In ‘Distributional verb semantics’ (chapter 6, pp. 73–94), the authors give a similarly detailed account of how they used prediction-based methods to find clusters of participles used with each auxiliary in their full data set, which also includes semi-passives like be disappointed (and probably some pseudopassives, though many of these were excluded, p. 39), and of how they subsequently assigned semantic categories to the groups. This particular research question is data-driven, and the semantic participle groupings often seem counterintuitive from a human standpoint (as in the ‘aggression and industry’ cluster, p. 84). I am also hesitant about the decision to form the clusters from the full dataset rather than the central-passive dataset and then use the clusters to draw conclusions about passives. My opinion after examining the participle clusters is that the authors might have been better served trying to categorize them by situation type/aktionsart (especially telicity) rather than semantics.
In chapter 7, titled ‘The structure of the proposed models’ (pp. 111–18), the reader is offered a well-written, comprehensive overview of what is known about be- and get-passives in the varieties of English included in GloWbE. I would consider this chapter, like chapter 2, to be a good starting place for students and other readers who are interested in the English passive. In fact, this chapter can be read almost entirely independently of the rest of the book, as there are few explicit ties to the studies aside from hypotheses for some varieties.
After a brief but important account of the models in chapter 8, we get to the results and analyses. It makes sense that the reader is kept on tenterhooks about the results for the better part of the book, since its main contribution is arguably the automated methods that the authors developed and (to some degree) used the passive alternation as a test case for. But anyone interested in the passive voice will be delighted to finally dig into the findings.
Chapter 9 presents the results of the diachronic studies of the proposed constraints on the get-passive compared to the be-passive (‘Diachronic developments in the English passive alternation’, pp. 119–56). There are not a lot of spoilers here, and earlier, smaller studies on the topic are supported. The authors find, for example, that be-passives have been decreasing and get-passives increasing, but that get-passives are still very infrequent compared to be-passives. Their findings are also consistent with earlier studies that show that the get-passive has been more common in informal genres, but that this constraint has been weakening over time in COHA. Get-passives are found to occur more often with animate subjects, which aligns with the constraint on subject responsibility, though the association does weaken over time, just as previous research has indicated. The get-passive is shown to be used in semantically non-neutral contexts more than the be-passive is, which tallies with earlier research as well (although not earlier research that claims that the get-passive is apt to be adversative rather than non-neutral, as some studies have suggested). Other parameters are also examined in this chapter, and I heartily recommend reading them – but the findings are almost entirely in line with previous work on the get-passive, which should be reassuring both to scholars who have used smaller corpora and manual coding (and who were able to come up with good results nonetheless) and to people who want to use automated analyses of huge datasets (since the findings seem reasonable).
The one place where the diachronic study of the COHA data feels a little less convincing, as the authors themselves write in their concluding discussion, is in the semantic analysis of clusters of main verbs with each passive. This exploratory investigation was, as I understand it, an attempt to put the computer in the linguistic driver’s seat and do something data-driven, but the results are somewhat messy and a bit difficult to meaningfully interpret. I find myself grateful to the authors for their critical note on the limited usefulness of this endeavor and enjoy their reflection that their analyses worked best when ‘the automated modeling of predictor variables was directly informed by a linguistically grounded understanding of the variable in question’ (p. 228), rather than working backwards from the computer-generated categories.
The synchronic comparison of the passives across different varieties of English, presented in chapter 10, also largely corroborates earlier findings from smaller-scale studies (‘Regional variation in the English passive alternation’, pp. 157–212). The constraints on the get-passive that are supported by the diachronic COHA data are visible here, although there are slight differences between varieties. The authors’ reflection on the results from both COHA and GloWbE is that focusing on central passives is ‘worth the effort’ (p. 212), which also tallies with previous work. It is in this group of true passives, where the only possible interpretation of the participle is verbal, that we find the clearest diachronic development (p. 154) and the best comparisons between the auxiliaries (since it is in this context that they are most clearly in alternation).
As with most well-written academic books, the final discussion chapter (chapter 11, pp. 213–32) can almost be read as a standalone text and will certainly provide a reasonable starting point for many readers before they dive into the nitty gritty of the methods and models presented in the body chapters. The authors use the final chapter to reflect on their findings and on the success of their methods, and they have succeeded in contributing to the field on both counts. While their findings do not revolutionize our understanding of the passive alternation in English, they are very comforting. We find that, while new, computational methods of studying language are exciting, our previous efforts were not in vain. In fact, the new methods succeed the best when they are employed to investigate earlier claims from smaller-scale studies.
In their final chapter, the authors also address the unease many linguists may feel about using large-language models to analyze linguistic data. I appreciate this discussion of the shortcomings of machine learning and the care that the authors put into making sure their methods were clear and replicable. I am also encouraged that they highlight the important role of the human theorist and interpreter. One point that I think is missing from their discussion is the challenge that compilers of giant, web-based corpora will soon face. One of the main contributions of this book is that the authors’ methods enabled analysis of data on a gargantuan scale. But if we want to continue studying the development of the passive voice in English, it will soon be difficult to obtain large amounts of present-day data that is guaranteed to have been written by humans rather than bots. The great advantage of corpus data, i.e. that it is real language made by real people for real communicative purposes, will be lost. If we want to keep up with language development in Present-Day English, we may need to go back to compiling small, careful, representative collections of texts and looking at them with human eyes.
