Register

In the definition used by our Collaborative Research Centre, register is “intra-individual variation in language use, that is, recurrent and conventionalized linguistic patterns that speakers employ depending on the situational-functional context”.¹ The same speaker uses language differently in different situations. That variation can affect every linguistic level, from lexicon and morphosyntax to discourse structure, and it is shaped by purpose, participants and setting. Register is therefore usage-based rather than user-based: unlike a dialect or a sociolect, it is not a stable property of a group of speakers.

The idea did not have to reach Egyptology from linguistics. The discipline saw functional variation from its beginnings, even if it mostly cast it as written against spoken language and as a succession of stages. In 1837 Richard Lepsius distinguished a “dialecte sacré”, used for religious and scholarly purposes, from a “dialecte populaire” that served legal and private affairs. In 1889 Adolf Erman set the archaising idiom of literary texts apart from the prose of contemporary scientific compendia and rejected the idea that Papyrus Westcar reflected the vernacular of its time. In 1925 Kurt Sethe described Demotic and Coptic not simply as stages but as functionally distinct varieties, one written and tied to official and religious discourse, the other spoken. And in 1933 Erman distinguished an ordinary way of speaking from an elevated style used in poetry and in references to the king and the gods, noted that educated scribes could move between the two, and pointed out that in Medinet Habu minor figures speak more simply than the main narrative.²

Register studies have three main traditions, and they differ above all in which direction they work.

Systemic Functional Linguistics works from the top down. Its line runs from the anthropologist Bronisław Malinowski through J. R. Firth and the London School of Linguistics to Michael Halliday and his students, who built a theory in which context determines linguistic choice.³ Field (the activity or subject matter), tenor (the roles and relationship of the participants) and mode (the channel and organisation of the text) define a register before any linguistic analysis begins. It is a strong tool for interpretation, but not in itself empirical.

Close to the beginning of this line, though outside it, stands Alan Gardiner, author of the seminal Egyptian Grammar. Encouraged by Bertrand Russell, he developed a theory of language that put communicative purpose and speaker intention at the centre, an idea he had already worked out in his correspondence with Malinowski from 1917 onwards.⁴ His Theory of Speech and Language (1932) treats utterances not primarily as representations of thought but as actions directed at influencing a listener. It distinguishes grammatical sentence form from sentence function, and it introduces a “special sentence quality” for the communicative aim behind an utterance — what speech act theory would later call illocutionary force. Gardiner speaks of an “act of speech” a year before Karl Bühler’s Sprechakt, which is usually credited as the first.⁵ He himself thought this work “of far greater importance” than his grammar,⁶ but it found little response. Malinowski was institutionalised through his student Firth and the London School of Linguistics; Gardiner, financially independent and without teaching duties, worked outside academic structures, and Firth, while mentioning him in connection with the context of situation, found his writing hard to read. Yet the classification of utterance types, the interrelation of morphosyntactic form and pragmatic function, and the modelling of language as context-bound social action all anticipate core tenets of later register theory.⁷

As a theory of register, this is also the tradition Egyptology adopted first. Pascal Vernus used “register” in the linguistic sense as early as 1979, recasting Fritz Hintze’s distinction between discourse and narration as one between registers, and in the same volume introduced égyptien(s) de tradition as a spectrum of mimetic registers rather than a single stage of the language.⁸ Karl Jansen-Winkeln’s habilitation, finished in 1990, classifies text types in a way that closely mirrors field, tenor and mode.⁹ Around the same time Orly Goldwasser met British register theory during her doctoral studies; her reading of the coexistence of Middle and Late Egyptian forms in Papyrus Anastasi I as register variation remains the most cited discussion of register in Egyptology.¹⁰

Variationist sociolinguistics works from the bottom up. William Labov showed, on Martha’s Vineyard and in the New York department stores, that language varies with class, ethnicity and situation, and that such patterns are found in the data rather than imposed on it.¹¹ Its principle is that the frequencies of linguistic features reflect contextual pressures.

Multi-dimensional analysis carries that principle over to written and spoken corpora. Douglas Biber takes linguistic features that pattern together as the trace of an underlying communicative function, and treats register not as a category but as a position in a space of several dimensions at once.¹²

In Egyptology the call for this kind of analysis is explicit. Todd Gillen wrote in 2014 that registers should be defined “according to observable clusters of regularities”, and that this “requires the methods of machine-based corpus linguistics for statistical analysis and multidimensional comparison”.¹³ Stéphane Polis adopted every theoretical component of Biber’s framework — registers as continuous constructs, distinguished by relative frequencies, with co-occurrence as the decisive relation — and concluded that “the use of quantitative methods is necessary”.¹⁴ His study of the scribe Amennakhte in the same volume calls its core “a multidimensional analysis of Amennakhte’s linguistic registers”, but takes over the name rather than the procedure: there is no factor analysis, no modelling of feature covariation, no latent dimensions.¹⁵ As far as I know, the analysis below is the first one for Egyptian that actually runs it.

How a multi-dimensional analysis works

A linguistic feature, in this sense, is anything in a text that can be counted: a part of speech such as the noun, a grammatical form such as the imperative or the passive, a class of words such as the suffix pronouns of the second person. Its rate is how often it occurs per word of text. The input to the analysis is a table: one row per text, one column per feature, and in each cell the rate of that feature in that text. The analysis runs in four steps.

1. Correlations. For every pair of features the analysis measures how far they rise and fall together across texts. This correlation runs from +1 (always together) through 0 (unrelated) to −1 (the one rises as the other falls).
2. Latent variables. Factor analysis then looks for a small number of hidden variables, the factors, that would explain this web of correlations. It first extracts the combination of features that accounts for most of what the features have in common — their shared variance — then the next combination independent of the first, and so on.
3. Rotation. The first solution is mathematically optimal but hard to read, because most features are moderately tied to most factors. Rotation reorients the solution, without changing how much it explains, until each feature is tied strongly to as few factors as possible.
4. Output. Loadings, between −1 and +1, say how strongly each feature is tied to each factor; a negative loading means the feature goes with the opposite end of the scale. Scores say how much of each factor’s pattern is present in each text.

Each factor is a dimension: a continuous scale, not a yes-or-no category, on which every text has a position. The interpretation rests on a single assumption — features that keep occurring together probably serve a shared communicative function. If questions, first-person pronouns and short clauses cluster, and the texts in which they cluster read as personal and interactive, the dimension is read as something like “involved versus informational”. The label follows the pattern. What factor analysis cannot do is name the factors or say what they mean; that remains the researcher’s job.

Short texts

A multi-dimensional analysis presupposes a corpus that is representative, sufficiently large and balanced. No corpus of an ancient language meets that. The record is fragmentary, its distributions are skewed, and above all its texts are short. Earlier Egyptian before Amarna may serve as an example: in the TLA it comprises 12,975 texts. The longest has 4,088 words, the mean is 40.7, the median 9.

A rate computed over nine words is either zero or very high: one occurrence of a feature makes it frequent, none makes it absent. Long texts are rare: before Amarna, 1,128 texts reach 100 words, 263 reach 300, and only 142 reach 500.

How long does a text have to be for its rates to mean something? One way to find out is to take long texts, cut them to a given length, and then cut them in half. If the two halves give the same picture, the length suffices; if they do not, what one sees is largely chance. How closely the halves agree is the split-half reliability: 1 means they agree completely, 0 not at all, and from about 0.8 upwards a measurement counts as reliable. The figure shows this for 89 long texts from the Eleventh to the end of the Eighteenth Dynasty, cut to lengths from 40 to 640 words.

The line reaches the reliable zone at around 160 words. The two dotted lines mark what this means in practice. At 51 words, the typical length of a dated text from this period, it lies far below that zone. At 500 words, the minimum length of the texts used in this study, it is well inside it.

More texts do not make up for short ones. This can be tested by simulating the whole analysis with texts of realistic length and comparing the dimensions that come out with the true ones. The measure for that comparison is the congruence of two dimensions: 1 means they are identical, values from about 0.85 upwards count as similar, lower values as different. With texts of realistic length the congruence stayed at about 0.5, whether the simulation used 89 texts or 800. With 500 words per text, 89 texts already reached 0.81.

For Earlier Egyptian alone, there are too few long texts. A factor analysis needs about five texts for every feature it measures. The analysis presented below works with 43 features and so needs at least some 215 texts; before Amarna, only 142 texts reach 500 words.

There is a way around this in principle: pooling short texts into longer samples. A simulation that cuts long texts into fragments of realistic length and recombines them shows that samples of some 200 to 300 words reproduce the dimensions almost exactly, with a congruence of about 0.97, provided the fragments pooled together come from the same genre and period. Pooled at random, the samples still yield dimensions, but the differences between registers disappear: at 500 words about 60 per cent of the differences between samples are lost. Pooling within genre and period therefore needs reliable labels for both. Egyptology has no agreed classification of its texts by genre, and the notion of genre is itself contested.¹⁶ The analysis below therefore does not pool, but uses texts that are long enough on their own.

A workaround corpus of 244 texts

The corpus used here therefore draws on all pre-Demotic stages of the language: 244 texts of at least 500 words each, selected by hand and arranged in six groups. Of texts attested in several copies only one was kept, the longest, and funerary texts were thinned out — the Book of the Dead, for instance, enters as fifteen spells, three of each type — so that they would not swamp rarer groups such as the letters. It is a diachronic corpus as a deliberate workaround, and in size it is comparable to the corpora of multi-dimensional analyses of living languages — Biber and Hared’s analysis of Somali, for instance, worked with 158 written and 121 spoken texts.¹⁷ What mixing the stages does to the dimensions is taken up further below.

The groups are a heuristic for sampling and description. The factor analysis never sees them: it works on the feature table alone, and relabelling or dissolving the groups changes no loading.

Why 500 words? Since reliability keeps rising with length, the threshold should be as high as the factor analysis allows. With 43 features at roughly five texts per feature, at least some 215 texts are needed. At 600 words only 191 remain, at 800 only 136. Five hundred is the strictest cut the analysis can sustain. Whether the threshold itself shapes the dimensions is taken up below, under the checks.

Features

Biber’s original inventory for English comprised 67 features in 16 categories, identified through part-of-speech patterns, dependency relations and lexical matching. The Egyptian inventory was built functionally: starting from what multi-dimensional analyses of ten living languages measure, and asking which Egyptian forms carry those functions. The result is 70 features. Not all of them enter the factor analysis. A feature that shares little of its variation with the others — its communality, the share it has in common with the factors, is below 0.20 — contributes noise rather than pattern and is dropped. That leaves 43. Three features are kept on theoretical grounds although they fall below the threshold: lexical diversity, the stative, and stance verbs. The table lists all 70 features; the 43 that enter the factor analysis are set in bold.

One absence stands out. Multi-dimensional analyses usually draw heavily on clause-level syntax: types of subordinate clause, coordination, dependency hierarchies. All of that is out of reach here. The TLA is lemmatised and annotated for morphology, but it supplies no syntactic trees or dependency relations, and extracting syntactic features would require either manual annotation on a scale that is not affordable or a syntactic parser for Egyptian, which does not yet exist. What the inventory captures is the morphological and lexical surface of functional variation, not its syntactic depth.

Five dimensions

How many dimensions should the analysis extract? Factor analysis can produce as many factors as there are features, but each further factor explains less of what the features have in common. How much a factor explains is its eigenvalue. Plotted in order, the eigenvalues form a scree plot, named after the rubble at the foot of a slope: where the steep descent turns into a flat run, further factors add little. A second check, parallel analysis, repeats the calculation on random data of the same size. A factor is worth keeping only if it explains more than the corresponding factor of random data.

Here the curve bends twice: sharply after the second factor, and once more after the fifth, after which it runs flat. By the scree plot alone, three factors would be defensible, and so would five. Parallel analysis sets the upper limit: after the seventh factor the curve falls to the level of random data. The two checks therefore mark a range, from three to seven, not a number. Within that range, Biber’s procedure is to compare the solutions and to choose the one whose factors can all be interpreted plausibly, erring towards more factors rather than fewer;¹⁹ here that is five. That the fourth and fifth factors lie between the two bends shows again in the next test: they are the two dimensions that do not hold up under every check.

Whether a dimension holds up is tested by running the whole analysis again with other settings and on other corpora, and comparing each dimension with its counterpart, using the congruence introduced above (known as Tucker’s φ). By a common convention, φ ≥ 0.95 counts as practically identical and φ ≥ 0.85 as similar. The four alternatives are a lower threshold of 0.15 for dropping features, no dropping of features at all, only texts of at least 600 words, and 110 additional texts of 400 to 499 words.

Dimensions 1 to 3 come out virtually identical every time. Dimensions 4 and 5 change shape whenever the corpus changes and are therefore labelled tentative. The two right-hand columns also answer the question whether the 500-word threshold makes the dimensions: raising it (191 texts) and lowering it (354 texts) both reproduce the stable core.

Each of the five dimensions is shown below in two panels. The upper panel lists the features that load on the dimension, with their loadings: bars to the right belong to one end of the scale, bars to the left to the other. The lower panel shows where the texts of each group fall on the scale, as a box plot. The box covers the middle half of the texts of a group, the line inside it marks the median, the whiskers reach to the remaining texts, and dots mark single texts far outside the rest.

The names of the dimensions are interpretations, not results of the calculation. Before a dimension was named, it was checked whether something other than a shared communicative function could explain why its features occur together — for instance the subject matter of the texts, or two features that are counted from partly the same forms.

Where texts and parts of texts fall

A first check is whether the groups fall where situational knowledge predicts. They do. Of all groups, the letters have the highest average on Dimension 1; they are also positive on Directive Address, the pragmatic core of letter writing, and negative on Elaborated Reference — letters act rather than describe. Medical texts are their mirror image: the lowest values on Dimension 2 in the corpus, recipe lists instead of connected prose; high on Generic Reference, with unnamed patients and counted substances; negative on Dimension 1, nominal and classificatory rather than verbal and sequential.

Group means, however, also hide structure. The medical group is positive on Dimension 5 on average, but that average conceals a split.

The diagnostic treatises, with their if-then structure of case and verdict, reach up to +22; the pure recipe collections sit at the negative end. “Medical text” as a category covers two quite different ways of organising discourse that happen to share a subject. The same holds across categories: among the texts with the highest scores on Dimension 5 are, besides medical diagnostics, spells of the Book of the Dead, the Dream Book, one of the Heqanakht letters and the Teaching of Ptahhotep. Casuistic framing is a resource available to several domains. Directive address behaves the same way — ritual invocations and letters share it. Linguistic configurations track communicative tasks, not domains.

Separately by language stage

For a comparison by language stage, the texts were not selected by hand but drawn from the TLA automatically: Old and Middle Egyptian by date (Third to Eighth Dynasty; Eleventh Dynasty to before Amarna), Late Egyptian by the TLA’s language tag. That yields 328, 261 and 117 documents of at least 150 words each, many of them shorter than the 500 words of the main corpus: the median length is about 220 words for Old, 300 for Middle and 400 for Late Egyptian. Run separately for each stage, and in a joint analysis of all three, the analysis gives a more sober picture. The check here is a different one: the documents of a stage are divided at random into two halves, the analysis is run on each half, and its dimensions are compared with those of the whole stage. Repeated many times over, only one to three dimensions per stage keep their shape; the others change from one division to the next. The three stages share no common set of dimensions: between stages the congruence reaches at most 0.74, well below the level at which dimensions count as similar. Only 22 features vary at all in all three stages, because much of the new grammar — articles, the progressive, loanwords — only exists from Late Egyptian onwards. What recurs is a basic contrast in each stage between nominal, monumental or enumerative texts and involved, verbal ones.

As a diachronic signal, one narrow strand survives every check: the shift from synthetic to analytic, with the status constructus declining and the indirect genitive, the article, demonstratives and possessive articles rising. It holds within a single group of texts as well — the decorum texts, the only group usable across all three stages. Beyond that, the separation of language stage and register has a principled limit, not only a statistical one: change runs through registers, the register categories themselves are not stable over fifteen hundred years, and what survives from each stage depends on its registers, because stone outlasts papyrus.

What is gained

The obvious objection is that all this was known. Every philologist knows that Papyrus Ebers contains both diagnostics and recipes. In that generality, yes. What was not known is which forms covary, that the diagnostic sections share their configuration with wisdom literature rather than with the recipes of the same papyrus, and that this emerges from a model that never saw the categories. Where the method reproduces what we knew, that is validation, and it is what makes the deviations worth taking seriously.

Knowing that texts differ is also not knowing on how many independent axes, by how much and with what internal spread. A text scored on the dimensions can be placed, whatever its genre label, including fragments and unprovenanced pieces. And Egyptian becomes comparable with the other languages analysed this way, to which it adds a time depth none of them has.

A second objection is that comparing the dimensions against situational expectations brings back the subjectivity that statistics claims to remove. It does. The claim is not that the interpretation is objective, but that the pattern is: anyone may read the same loadings differently, but they have to read the same loadings. Quantification does not remove subjectivity; it makes it inspectable, documentable and revisable. The relevant contrast is not objective against subjective, but checkable against uncheckable subjectivity.

Tagging, and the missing syntax

With a fully lemmatised corpus, Egyptian is comparatively well placed for this kind of analysis. Feature tagging in multi-dimensional analysis was always at least partly automatic, and Biber himself considered a coverage of “90 per cent or better” — after manual editing — good enough.²⁰ For English it is now fully automatic, and fast. One estimate takes Common Crawl, a freely available archive of text collected from the web, at around 1.1 trillion tokens (roughly: words). A tagger based on a neural network, run on eight graphics processors in parallel, would need about 13.6 days to tag it; the established Multidimensional Analysis Tagger, a conventional program running on ordinary processors with the same degree of parallelism, about 756 days, more than two years.²¹

For Egyptian, neural models for lemmatisation and part-of-speech tagging exist too.²² The table gives the share of words, in per cent, for which their answer is correct, separately by language stage and by the form in which the text is fed in: as transliteration or as a sequence of hieroglyphs. Detailed part of speech (XPOS) is a simplified version of the TLA’s own word classes, which, for instance, tell divine, royal and personal names apart and separate titles and epithets from other nouns. Basic part of speech (UPOS) is a coarse class such as noun or verb from a scheme shared across languages. Lemma is the dictionary entry a word belongs to, down to which of several homonymous entries it is.

What is missing is syntax. Lemmas and word classes can now be assigned by machine with the accuracy shown above; for syntax no such model for Egyptian exists yet, and one can only be trained on texts whose syntax has already been annotated by hand. Wolfgang Schenkel set out to create such material early on. In 1968 he demonstrated M.A.A.T. (Maschinelle Analyse Altägyptischer Texte), a system he had built to analyse spells of the Coffin Texts lexically, morphologically and syntactically, and called its principle “integrated” capture: one machine-readable corpus, annotated once on every level, for every kind of study.²³ In 1985 he added what such a corpus was for: the linguistic character of a text was to be judged not by collecting individual Late Egyptianisms and classicisms but by “quantitative estimates, if not statistics”.²⁴ A multi-dimensional analysis is a statistic of this kind. His Coffin Text data, around 400,000 fully syntactically annotated tokens, are now being integrated into the TLA. Two more recent projects work towards the same end on a smaller scale: Carlos Gracia Zamacona’s database of the Coffin Texts, with more than 27,000 records, most of them a simple sentence each, with fields for syntactic analysis,²⁵ and the Egyptian-UJaen Treebank of Roberto A. Díaz Hernández, in its current release some 34,000 tokens from the Pyramid Texts.²⁶ Material of this kind could serve to train models for automatic syntactic annotation — and with that, a complete multi-dimensional analysis would become possible.

The limits that the corpus imposes — few text types, little overlap between the stages, short units — cannot be calculated away. The limit set by the missing syntax can be lifted.

244 texts from the TLA raw data, at least 500 words each, all pre-Demotic stages. 70 features extracted in Python, 43 in the factor analysis; five factors. Robustness checked by alternative feature selection, alternative corpora and split-half comparison. Code and feature definitions are not published; I am glad to share them on request.

Why a model and not a count

The obvious objection to any table of proportions in Egyptology is that the things being counted are not alike. Four problems in particular, all of which a hierarchical model is built to handle and a cross-table is not.

Everything is unbalanced. zꜣ “son” appears as head 1,005 times, ꞽwf “flesh” 642 times, nb “lord” 562; the median head is attested seven times, and the twenty commonest heads carry 42 per cent of all attestations. Some texts contribute two attestations, others two hundred. A raw proportion lets the big contributors decide, and “the corpus says” quietly means what the half-dozen largest sub-corpora say.

Everything is confounded with everything. A genre looks archaising until one notices it happens to use certain words; a lexeme looks conservative until one notices it occurs mainly in early texts. Counting one factor at a time cannot separate these. A model weighs them simultaneously and reports what is left of each when the others are held constant — which is why, for example, the animacy effect below is nearly invisible in the raw figures and clear once frequency is accounted for.

Rare things must not be over-read. The most useful property of a hierarchical model is one Egyptologists will recognise as an old methodological worry made precise. A lexeme attested twice, both times with the indirect genitive, does not get a rate of 100 per cent; it is pulled towards the corpus average, and how far depends on how much evidence it has. Lexemes with many attestations keep their own value, lexemes with few borrow strength from the rest. Nothing has to be thrown away for being rare, and nothing rare gets to shout.

Dates are spans, not years. The time axis in the model is a smooth curve fitted to the midpoints of dating ranges, which is a compromise, but a stated one.

The result is a hierarchical Bayesian logistic regression: separate levels for the head lemma, the dependent lemma and their combination, alongside text, sentence, sub-corpus and genre, plus a curve over time and a set of ordinary predictors. Fitted in Stan; the implementation is not the subject here.

Reading the numbers below

Two conventions, because the rest of the post uses them throughout.

Coefficients are in log-odds. They are additive on a scale that turns into percentages only once you know where you started. A coefficient of +0.33 means: if comparable cases sat at 40 per cent indirect, this feature moves them to about 48. The same +0.33 applied to cases already at 80 per cent would only move them to 85 — the scale compresses at the ends, which is exactly what you want when the thing being modelled is a proportion. Negative means “towards the direct genitive”, positive means “towards the indirect one”. As a rough guide: ±0.2 is small, ±1 is large, ±2 is overwhelming.

Intervals mean what they appear to mean. The brackets are 90 per cent credible intervals, and this is where the Bayesian framework earns its keep for philological work: “−0.99 [−1.65; −0.35]” says there is a 90 per cent probability that the value lies in that range, given model and data. That is the statement everyone wants to make and which a classical confidence interval does not license. When an interval covers zero, the data do not tell us the direction — which, as will become clear, happens more often here than one would like.

The head noun is what matters

Before the individual claims, one result that none of them states. Of everything the lexicon contributes to the choice, the combination of the two nouns accounts for roughly 57 per cent, the head for 36, and the dependent for only 7. The head is worth five times the dependent — Jansen-Winkeln’s thesis, confirmed. But the largest share belongs to neither noun separately. Which two words stand together matters more than either one does alone, and no property of either noun predicts it.

zꜣ “son” shows what that means in practice. Werning finds all nine kinship uses in the Book of Caverns direct and explains them as an old lexicalised compound, with Coptic ⲥⲓⲧⲉ “basilisk” from zꜣ tꜣ as the residue. Across the whole corpus it is a strong preference rather than a categorical one: 22 per cent of its 1,005 attestations are indirect, and they run through every period.

The number rule holds — for two formulas

Jansen-Winkeln’s clearest lexical observation concerns monosyllabic masculines, especially body parts: in the Pyramid Texts they take the indirect genitive in the singular and the direct one in the dual and plural. He does not confine the pattern to that corpus — he cites ḥm nï “the majesty of” and finds traces still in Late Middle Egyptian — and his list of lexemes ends with “and others”.

Half of this is right. Raw, the singular stands indirect in 40.6 per cent of cases, the plural in 40.6 — no difference at all — and the dual in 16.2. In the model, with everything else held constant, the dual sits at −0.99 [−1.65; −0.35] and the plural at +0.07 [−0.29; +0.41]. So the pull towards the direct genitive is substantial in the dual and absent in the plural: a refinement rather than a refutation, since paired body parts in the dual are the core of his examples.

But it is less even than that. Of 419 dual attestations, 252 belong to three head lemmas, and two of those are single fixed expressions: ꜥꜣ-wï= p:t “the two door leaves of the sky”, 99 attestations, and zḫn-wï= p:t, 89 attestations, which is one lemma pair. Both are direct almost without exception. Remove the three commonest dual heads and the dual stands at 35.9 per cent against the singular’s 40.6 — five points instead of twenty-four.

The model has a level for the lemma pair precisely so that formulas like these can be absorbed there rather than masquerading as grammar, and it half does: ꜥꜣ + p:t sits at −0.93, but with an interval of [−2.07; +0.20] that covers zero. With 99 attestations of one expression and almost no counter-evidence from the same pair in the singular, the model cannot fully decide whether what it sees is a property of the expression, of the lexeme, or of the dual. All three explanations fit.

The honest formulation is that the rule holds for certain lexicalised dual expressions rather than for the dual as a grammatical category. That is not a small distinction: it moves the phenomenon from morphology into the lexicon, which is where Jansen-Winkeln located the whole question in the first place.

Material and measure: strong, but not a rule

He also holds that expressions of material, length and content always take the indirect genitive; Werning finds the same for the genitivus materiae, which appears only indirectly on his semantic map.

The claim is not originally Jansen-Winkeln’s. He reports it from Erman and Osing, who separate two genitives — belonging and dependence on the one hand, qualifying specification (material, measure, content, kind) on the other — and he adds a morphological account: the second type goes back to a reversed, passive nisba nï, as in mdw nï z “the man’s word” against z nï mdw “a man of words” (§§9–11). On his view the second type should strictly speaking not be called a genitive at all.

The preference is strong and the exceptionlessness is not there. Material words as dependent stand indirect in 68.8 per cent of cases and measures in 72.2, against 40.0 elsewhere — but 29 material and 10 measure attestations stand in the direct genitive.

Whether the preference belongs to the class or to its eight members cannot be decided here: the class effect sits at +0.50 [−0.19; +1.19] and demonstrates nothing. What carries the class is three words: ṯḥn:t “faience, glass” at 94 per cent of 17 attestations, ḥmt “copper, ore” at 83 per cent of 24, and mḥ “cubit” at 73 per cent of 33. The second-largest material lexeme, bd “natron”, stands at 45 per cent of 22, barely above the corpus at large. “Material words” as a category carry nothing.

Werning’s animacy artefact does not appear here

Werning’s animacy finding is not really a claim about animacy. It is a claim about an artefact: the reversed relation he measures is, he argues, a by-product of cases like nb= ḏꜣḏꜣ “owner of the head”, where the possessum grammatically owns the possessor. Remove those and the relation nearly disappears. His raw figures put the reversal plainly: against Kammerzell’s expectation and against the iconicity principle, inalienable possession stands indirect in 35 per cent of his cases and alienable in only 15.

Checking this needs an animacy scale, which the TLA does not have — its dependent categories (divine name, toponym, title, royal name, personal name, common noun) are not one, because the common noun mixes people with things. So 200 dependent lexemes covering 88 per cent of attestations were coded by hand, on a sheet that deliberately did not show how those lexemes behave.

The test is a comparison of two runs, one with his class in the model and one without, since his claim is about distortion rather than about a value. The animacy coefficient moves by 0.002 — two thousandths of a log-odds unit, which is nothing. His class itself sits at −0.29 [−0.92; +0.35] and carries nothing either.

Animacy in this corpus points the ordinary way rather than the reversed one: +0.33 [+0.06; +0.60] for human and divine possessors among common nouns, which is about eight percentage points. And it is a good illustration of why one models rather than counts, because raw it is nearly invisible: 41.0 per cent indirect for animate possessors against 39.5 for inanimate. A point and a half.

What hides it is frequency. Animate common nouns are much more frequent words, and frequency pulls towards the direct genitive; the two effects nearly cancel. Compare like with like and the difference appears — in the middle frequency band, animate possessors stand indirect in 49.7 per cent of cases against 30.2 for inanimate ones. Twenty points. A cross-table would have reported “no animacy effect” and been wrong, and a cross-table split by frequency would have run out of data.

One caveat on the negative result. Because every head lemma has its own effect in the model, nb, ḥqꜣ and nzw are already accounted for individually before Werning’s class is switched on. What is tested here is whether the class distorts anything beyond that. It does not — but that is not the same as showing his correction is unnecessary in a simple cross-table, where it may well be needed.

Archaising belongs to compositions, not to genres

Werning reads the 80 per cent direct genitives of the Book of Caverns as deliberate archaising. The contrast that makes the reading plausible is with his contemporaries: across the New Kingdom the corpus sits at 51.7 per cent indirect, so a text at 20 per cent is far outside the range of what was usual when it was composed.

What the Book of Caverns gets wrong is not only the proportion. Jansen-Winkeln stresses that in Old and Middle Egyptian the direct genitive is not a compound but a free word group: the dependent can be expanded by a suffix, a further genitive, an adjective or a relative phrase, and the sequence can be broken (§4). Werning finds in the Book of Caverns 14 clear and 23 disputed cases of directly nested genitives, a pattern all but unattested in genuine Old and Middle Egyptian, and concludes that its authors no longer understood the direct genitive as a phonologically marked compound but as bare juxtaposition. The rule he reconstructs for their mental translation — delete the article, inflect the connector or delete it — contains no phonology at all.

That is a prediction one can check: religious compositions of the New Kingdom should stand markedly more direct than contemporaries once date, lexicon and grammar are held constant.

They do, and strongly. Religious and funerary texts of the New Kingdom sit at −1.73 [−2.14; −1.33] against everything else of their time, with the Book of the Dead at −1.31.

The spread is nearly fifty points, and none of it is chronological — this is all the same five centuries. Technical prose sits at the indirect end: medical texts 69.4 per cent, magical texts 58.4. Everyday documents sit at the direct end, letters lowest of all at 20.8. The Book of the Dead, at 42.3 per cent, is nearly ten points below literary texts of the same period. That is register: the same language, and writers choosing differently according to what they were writing.

The prediction fails for Werning’s own material, though, and for a reason worth knowing. The netherworld books of the New Kingdom sit at +0.10 [−0.37; +0.59] — on the average, nothing archaising about them. And the Book of Caverns itself is in the TLA with six text pieces, five dated to the 30th Dynasty and one to Psammetichus I, because the TLA dates by witness and the surviving copies come from TT 33 and later. Only two of its attestations reach this study at all. His text is effectively not in this corpus.

What that leaves is sharper than the original question. Archaising is not a property of religious texts in general but of particular compositions: the Book of the Dead holds on to the direct genitive, the netherworld books do not — same period, same lexicon, same wider category of text.

The largest effect has no explanation

One thing the model turns up is larger than anything discussed so far, and the corpus cannot settle what lies behind it.

Because the number rule is lexical rather than general, each head lemma gets its own slope for “not singular” — its own answer to the question of what the plural does to it. The spread of those slopes is 2.11 in log-odds, the largest single quantity anywhere in the analysis: larger than any group level, larger than date, larger than any fixed effect. And it runs in both directions. ꞽb= nṯr “the god’s heart” is indirect in 70 per cent of its attestations; ꞽb-w= nṯr-w “the gods’ hearts” in none of its fifteen plural attestations. ꜥ “arm” is indirect in 72 per cent of singulars and 2 per cent of duals. ḥꜣb “festival” goes the opposite way: 22 per cent in the singular, 83 in the plural.

The first thing to rule out is that this is an artefact of the shrinkage described earlier — that the model is inventing variation among lexemes it knows nothing about. It is not: 412 of 625 head lemmas have no non-singular attestation at all, and their slopes have a spread of 0.02 against 1.47 for the rest. The model pulls the uninformed ones to zero exactly as it should, and the 2.11 comes from the roughly 35 lemmas that have real evidence.

The cleanest demonstration that the effect is linguistic dispenses with the model entirely. Of the 42 lemma pairs attested at least five times in both numbers, the average difference in proportion between singular and non-singular is 24 percentage points, and eleven differ by more than 30. Same head, same dependent, different number, different construction.

The obvious explanation is alienability. The lemmas pulling towards the direct genitive in the plural are, on the whole, body parts and soul parts — ꜥ, ꞽb, qs “bone”, bꜣ, dp — and those pulling the other way are alienable things and events. It would be a striking result, because Werning finds no alienability split as a main effect: it would exist only in interaction with number.

It does not survive testing. Coding 213 head lemmas by hand and adding alienability as a predictor on the slope gives −0.75 [−1.33; −0.15] across all coded lemmas. But thirty-five of those lemmas are the ones that suggested the hypothesis in the first place. Restricted to the other 178, it is −0.47 [−1.14; +0.20]: the right direction, about 87 per cent of the probability mass below zero, and an interval that includes nothing happening. And even in the generous version it accounts for about 8 per cent of the variation it is meant to explain; in the honest version, 2.

A sharper test would need more duals and plurals than exist. There are 1,622 in 16,812 attestations, 9.6 per cent, spread over 213 lemmas.

There is one candidate explanation older than alienability, and it comes from Jansen-Winkeln himself. He takes the complementarity of the monosyllables — indirect in the singular, direct in the dual and plural — to have morphological and phonological rather than semantic causes, and in a footnote he offers a mechanism: a monosyllable necessarily has a short stressed vowel in the singular, so there is no long-to-short reduction for the status constructus to make, whereas the plural and the feminine have one, and in the dual a reduction would at least be easier (§8 with n. 78). He marks this as a conjecture. The lemmas with the steepest negative slopes here are largely his monosyllables. What stands in the way of testing it is that every head lemma would have to be classified by syllable structure and by its behaviour in the status constructus, and the TLA annotation supplies neither. The explanation that fits best is the one this corpus cannot be asked about.

So the largest thing this analysis measures is real, is lexical, and is not explained here.

A sentence from 2000

Having described how lexicon and morphology steer the choice, Jansen-Winkeln writes (§8):

Diese Regeln im einzelnen (und für die verschiedenen Epochen) zu entschlüsseln, würde > umfangreiche Statistiken erfordern. > > To decode these rules in detail, and for the various periods, would require extensive > statistics.

The sentence is twenty-six years old. He asks for extensive statistics. What they need has come into being since: a digital, lemmatised corpus, and the computing power for a Bayesian model that gives each lemma an effect of its own. This post works through one, for the part of the material where both constructions were genuinely available. The commonest result is that the question cannot be decided here. That, too, is an answer.

16,812 attestations from TLA v20, all periods. Hierarchical Bayesian logistic regression in Stan: crossed levels for head lemma, dependent lemma and lemma pair, alongside text, sentence, sub-corpus and genre, a cubic spline over date, and predictors on the lexeme levels; non-centred parameterisation throughout, four chains, no divergences. Code, coding sheets and full coefficient tables are not published; I am glad to share them on request. This is a side project and the findings are provisional.

Reading the plot

The y-axis is the sum of the weights, in tokens per year. The pale line is the raw sum, the bold line the same data smoothed with a Gaussian kernel of 25 years; the distance between them shows how much of the shape comes from the data and how much from the smoothing. The bands in the background are the conventional periods.

In absolute mode the curve shows the lemma’s own weight. Since closely dated texts give off more weight per year than loosely dated ones, and the corpus is unevenly dense across the centuries, it rises and falls with the corpus as much as with the word. In relative mode the same weight is divided by the weight of all lemmatised tokens of the same year.

What is counted

Every token carrying a lemma ID is counted; broken and unidentified words contribute nothing. Alternative readings of a passage are stored as separate files in the TLA and would be counted twice, so only the canonical reading is used. Where a text carries several dating statements, its weight is divided among them.

Two exclusions: texts whose dating carries no year interval in the thesaurus drop out, and lemmata with fewer than three tokens in the corpus are not exported at all.

Comparing two forms

I built this not for the single curve but for the comparison of two. Where a Middle Egyptian form and its Late Egyptian counterpart compete, plotting both in one frame shows whether the one recedes as the other advances, whether they stand side by side for centuries, and where they cross — if they do.

This needs the relative mode (Figure 2). In absolute terms both curves rise and fall with the corpus, and two forms attested in the same texts will look alike however different their histories; only as shares of the same year’s evidence can they be read against each other.

The tool

Attestation density in the TLA takes one or more lemma numbers, separated by commas, and draws them in one frame. Demotic lemmata carry a leading d; since the TLA keeps two lemma lists, linked entries can be merged to follow a word across the change of script.

Built from the TLA raw data, corpus issue v20. The underlying index holds, per lemma, the dating intervals of the texts it occurs in with a token count, aggregated and without text identifiers.