A Theory of Linguistic Individuality for Authorship Analysis, by Andrea Nini, appeared in 2023 as an instalment in the series Elements in Forensic Linguistics published by Cambridge University Press. Coming in at just short of ninety pages of main text (which I read in the e-book version), the book presents a rather focused contribution, that is pleasantly concise and adopts a writing style that should be accessible to specialists and non-expert readers alike. Nini, currently a senior lecturer at the University of Manchester, has structured his text in five chapters with a logical build up that is generally adequately paced and convincing. Nini is no newcomer in the field of forensic authorship analyses – for some years now, my students in Antwerp have had to replicate his inspiring analysis of the corpus of ‘Jack the Ripper’ letters, which we featured as a fascinating real-world homework exercise in our Humanities Data Analysis textbook. His aims in this book, however, are much more ambitious and certainly transcend that of a single case study.
The introductory chapter (pp. 1–20) is perhaps also the most substantial one. Here, Nini frames his endeavor against the wider background of computational authorship analysis, especially approaches to this task which are based on language usage. Whereas this task is, of course, ‘fundamentally linguistic’ in nature (p. 2), Nini sets out from the correct observation that there has been a growing disconnect in recent decades between the scholarly practice in stylometry – the computational study of writing style – and linguistics. He offers a survey of prior scholarship in this area that highlights the major milestones in the discipline’s history, including the pioneering work by Mosteller & Wallace (Reference Mosteller and Wallace1963) on the infamous Federalist Papers. Scholars nowadays agree that function words and character n-grams are among the most efficient textual properties to model authorship and Nini traces how these (heavily lexically oriented) style markers rose to dominance. Much attention is paid to Burrows’ (Reference Burrows2002) influential Delta metric (to calculate the stylistic distance between texts), as well as subsequent interpretations and variations of the metric.
Whereas stylometry has invited contributions from a variety of fields (forensics, information retrieval, machine learning, but also philology, which is my own home discipline), Nini makes the somewhat radical observation that authorship analysis has, in effect, mostly become a subfield of computer science; linguistic contributions to the field, on the other hand, have indeed become exceedingly sparse. This development is mainly worrying, he argues, because it happened at the expense of explainability (p. 8) – the field has been under-theorized, if not de-theorized altogether. A majority of studies in this area have focused solely on performance chasing and appear to be ‘atheoretical in their conception of language’ (p. 9), according to Nini. While a number of isolated studies have presented valuable work that dug deeper, the author correctly identifies a major gap in the existing literature. Nini sets out to remedy this lacuna. Even if the stylometric community by now has developed a methodological practice that seems to produce satisfactory results in many circumstances, he makes a strong case that our joint understanding of why this would work in the first place is far less advanced, to say the least.
Nini argues that the solution lies in (re)connecting with usage-based linguistics, and more specifically cognitive linguistics. From within this framework, he advances the critique that stylometry has gravitated too much towards low-level, lexical features which disregard the insight that lexis, in reality, cannot be properly separated from grammar, but also the fact that evidence has mounted that language is processed (and to some extent also stored) in larger chunks. These chunks, however, are not covered in traditional grammars and do not straightforwardly correspond to conventional feature extraction techniques in stylometry. Nini discusses how such chunks can form flexible hierarchies of units through the processes of entrenchment and conventionalization. This hierarchy must be understood (and modeled) as a network of interdependent nodes.
Nini goes on to connect these insights to linguistic individuality: he convincingly argues that commonalities across language users have received much more attention than individuality in linguistics. Interestingly, intra-user variability is often disregarded as noise in cognitive studies too. He discusses the work by scholars such as McMenamin (Reference McMenamin2002) and Coulthard (Reference Coulthard2004) on combinatorics in language to engage with the notion of idiolectal, viz. individual style. Here (and elsewhere), the author argues that writing style is not a stable or invariant characteristic of an individual: rather, it emerges from an individual’s interaction with a specific context. He calls attention to a recent line of empirical, corpus-driven research in linguistic individuality that shows how strong this aspect of language usage really is. Nini notes, however, that prior work cannot sufficiently explain precisely why an individual idiolect emerges – although the author points to language acquisition, this aspect of the discussion is developed in less detail than I would have hoped.
In the second chapter, the book turns to a more ‘formal grounding’ for a theory of linguistic individuality (pp. 21–31). Using mathematical notation, based on set theory and graph theory, the author aims to develop a concrete model to put his views to the test. Here the goal is to develop a novel method to approximate an individual’s grammar from example writings, i.e. a linguistic fingerprint. Since such a fingerprint can only be extracted from a limited set of writing samples, such a model can only be an approximation of an individual’s true grammar. Here, Nini provides an interesting discussion of Herdan-Heap’s Law on vocabulary growth, which captures the diminishing returns of sampling more tokens on type diversity. If the same regularity applied to an individual’s grammar, at some point, sampling more writing from an author should stop yielding quintessential characteristics, which is encouraging from a modelling perspective.
Next, in the third chapter (pp. 32–69), Nini starts putting his ideas to the test in formal experiments using two English-language corpora: the English section of the refcor corpus (75 novels by 25 historic authors – 3 novels each) and the contemporary c50 corpus, in which 50 authors are each represented by 100 texts, again in English. Both these corpora have been the subject of prior studies in the field that relied on more conventional methods, including Delta-inspired distance metrics. The point that Nini is trying to drive home seems to be the following: if linguistic individuality emerges from the idiosyncratic repertoire of units, a binary vector representation of the appearance of these units in texts should suffice to capture that repertoire (whereas traditional work typically works with more complex scalar vectors that capture usage frequency). Consequently, a binary coefficient (or distance metric) should suffice to attribute a text of unknown provenance to its correct author. To approximate an author’s grammar and represent it in an individual profile, Nini resorts to tracing n-grams of various sorts, including skipgrams of tokens, part of speech tags, etc.
In the ensuing sections, a variety of experiments are reported in which coefficients on binary feature vectors are compared to more conventional approaches, such as Burrows’ Delta and variations. In many cases, variants of Delta are outperformed by Nini’s binary approaches (as evaluated in leave-one-out validation schemes). Interestingly, the binary coefficient metrics typically perform relatively better if they also consider the common absence of units in two vectors, as opposed to only looking at the common presence of units. The more useful feature types in this analysis appear to be somewhat longer than the shorter, isolated function words that are dominant in the field: word 2-grams or character 9-grams, for instance, yield superior results, but even longer n-grams appear to work surprisingly well. This does raise the uneasy question to what extent such units pick up on rare, topical information, which stylometrists normally prefer to avoid, to maximize the scalability of their representations. Nini argues that the longer a recurrent n-gram is, the less likely it is that topic would be responsible for its reoccurrence (p. 61). I must admit I did not find this argument entirely convincing, because so many of the examples listed (e.g. table 9) seem very specific to the specific (e.g. internet-mediated) text varieties in these corpora.
I did not spot any methodological flaws in the experimental sections, which are reported with sufficient care and detail. I do wonder whether the empirical chapters can put enough weight in the scholarly scale to provide support for the ambitious theory developed in the introductory chapters. The experimental sections are, in the end, based on just two corpora, in just one language. That is perhaps a pity: even if no study of human authorship can be exhaustive of course, large, trustworthy, multilingual corpora are nowadays readily available for authorship research. Plenty of worthwhile theoretical ideas developed in the introduction are eventually not put to empirical scrutiny. The idea of a network – a grammar as a graph – is interesting but not operationalized in the end, for instance. The cross-genre case, where attribution problems are presented across different text varieties, or even simpler, cross-topic cases are hardly considered. Such problems are nevertheless worth considering, given the theoretical assumption that style emerges from an individual’s interaction with a specific context – genre or topic, but also intended audience, seems a case in point. Would binary coefficients still work in such cases? And why (not)? Also, the claim that an author’s individual grammar would be subject to temporal change would be valuable to study from this newly suggested framework. Those critical questions do not invalidate the author’s views, but serve to underline that even more research will be necessary in the future, especially when it comes to the open challenges in the field. Adversarial attacks, for instance, are a fascinating new research topic: which verification methods will prove to be robust in the face of malignant forgeries? Wouldn’t longer units be very prone to deliberate imitation by epigones? These questions grow even more prominent with the recent rise of Large Language Models that are increasingly used for paraphrasing.
In conclusion, Nini advances valuable ideas in a transparent language that will be accessible to non-linguists as well. The corpora, as well as Nini’s code, are publicly available, which adds to the replicability, transparency and the reliability of the work. Ultimately, the book might not immediately impact how stylometry is carried out on a practical or methodological level – but that probably wasn’t its intention either: first and foremost, the book deserves praise for its impressive and bold attempt to re-establish the connection between linguistics and the (heavily under-theorized) discipline of stylometry. A Theory of Linguistic Individuality for Authorship Analysis is recommended reading – even required reading – for anyone active in the field of stylometry and authorship studies, especially those inquisitive scholars who are keen on developing a deeper understanding as to why what we do seems to work.
