1 Introduction
The upheaval in the early Middle English (ME) lexicon is a well-known phenomenon; Old English words disappear en masse, the numbers of French (and to a lesser extent Latin and Old Norse) loanwords increase rapidly, and home-grown Middle English word-formation processes are very productive (see e.g. Dekeyser & Pauwels Reference Dekeyser and Pauwels1990 for an overview). The commonly cited reasons for this are ‘need and prestige’ (Durkin Reference Durkin and Hickey2020: 172). Although prestige or need alone is not enough to bring about the adoption of French loanwords if large parts of the population never have the need or opportunity to come into contact with the language, Durkin (Reference Durkin and Hickey2020: 169) points out the difference between the ‘initial interlinguistic transfer’ and the ‘subsequent intralinguistic spread’. In the case of early Middle English, the first process would not have affected the large group of monolingual English speakers, but the second one certainly would have. Consequently, bilingual mediators would have played an important role in bringing the monolingual English populace into contact with French elements, since ‘words and structures need not necessarily be borrowed directly from the native speakers of the source language’ (Skaffari Reference Skaffari, Brinton and Bergs2017: 191). The role of the clergy as such bilingual mediators has recently been highlighted (Ingham Reference Ingham2018), as they were educated in French, or at least had to know certain religious terminology in French and were at the same time responsible for the instruction of monolingual English laypeople (Timofeeva Reference Timofeeva2020a).
While ‘need’ is a self-explanatory reason for borrowing, there are also many cases where it could not have been the main factor since Old English (OE) had perfectly serviceable words already, and ‘prestige’ must serve as an explanation. For example, Durkin (Reference Durkin2014: 238) describes the synonym trio frith, grith and peace:
Of these, frith barely survives beyond the early Middle English period, in one or two very restricted specialised uses. Grith remains quite frequent in Middle English (and occurs in some very limited contexts even later than this), although in later Middle English it occurs largely in verse texts in which its use was probably stylistically marked. (Later uses of both frith and grith are often in formulae paired with a more recent synonym, e.g. frith and grith or peace and grith.)
Of the three, frith is the OE word, grith was borrowed from Old Norse (ON) early on and already well attested in OE, so it seems that even though English had two available options to express the concept ‘peace’ (and therefore no ‘need’ to borrow from French), the French loanword was adopted and ultimately spread far enough to become part of the English lexicon. Following this development, the OE synonyms slowly fell out of use, surviving longest in poetry and binomial formulae.
A similar case is that of the trio envy, nith and onde: one French and two OE options, of which the French ultimately prevails and the others often appear as part of a binomial (Timofeeva Reference Timofeeva, Lorido, Prado-Alonso and Rodríguez-Puente2020b). Here, too, the use in poetry warrants notice, specifically the use in rhyming position: firstly, ‘[t]he perceptual salience of rhyme is a well-known fact, which, in our case, may have contributed to the lexical priming of onde and, in turn, its easier acquisition, wider diffusion, and, ultimately, longer life span’ (Reference Timofeeva, Lorido, Prado-Alonso and Rodríguez-Puente2020b: 63), and secondly, the French synonym envy far surpasses either OE synonym in terms of actual and potential rhyming partners (Timofeeva Reference Timofeeva, Lorido, Prado-Alonso and Rodríguez-Puente2020b: 63). This seems to fit neatly with the numbers we find in the Linguistic Atlas of Early Middle English (LAEME; Laing Reference Laing2013–), where envy is rhymed in 18/31 instances (58%), onde 24/71 (34%) and nith 10/53 (19%).
However, the replacement of Germanic with Romance vocabulary as in these examples is neither complete nor systematic. While prestige may serve as an explanation for why envy was borrowed, it does not explain why other, similar words in the same semantic domain were not also replaced by French borrowings. For many of the French loans in these domains, we find English near-synonyms which then go on to co-exist, and in some cases we can observe the continued survival of a word derived from Old English, despite supposed competitionFootnote 1 from a French loan, even in such domains that are supposedly heavily dominated by French. For example, among the cardinal sins, which can be put very firmly into the semantic domain of ‘religion’, we find surviving English-origin wrath and sloth, which had French competitors in the words erour and accyde, respectively (Käsmann Reference Käsmann1961: 296–300), but also French gluttony and envie, which eventually ousted their English-origin counterparts, gifernesse and niþ or onde, respectively (Käsmann Reference Käsmann1961: 293–5). What, if anything, gave one word an advantage over a contemporary (near-)synonym in cases like these?
As implied in the examples described above, it is possible that rhyme could be one such advantageous factor. In addition to these examples, there are others who mention the importance of rhyme in individual cases, e.g. Käsmann (Reference Käsmann1961: 99) on christianity, or Gardner (Reference Gardner2014) with regard to competing variants of -hood, but so far there has not been a quantitative study investigating this possibility. Using a corpus linguistic approach, this article aims to answer the question of which factors contribute to a better chance of survival for words in the early Middle English lexicon, with a particular interest in rhyming position. I will first elaborate on the effect of rhyme on mental processing, then describe my data collection method and the statistical model used for analysis, before finally discussing the results in general as well as a few specific examples.
2 Rhyme and poetic language
Words in rhyming position are very salient, as evidenced by, for example, the long-lasting and continuous use of rhyme as a mnemonic device. Awareness of rhyme is also associated with successful language acquisition; rhyme judgement tasks are used in studies of child language acquisition to gauge their phonological awareness (e.g. Carroll & Snowling Reference Carroll and Snowling2001), and children's ability to recognise and judge rhyme ‘appears to be essentially adult-like by the age of 7’ (Coch & Gullick Reference Coch, Gullick, Kappenman and Luck2011: 497 and references therein). Recent studies on the effects of rhyme on language and cognition show that, overall, rhyme lessens the cognitive load (which manifests as, e.g., shorter response times in experiments) in spoken, written and cross-modal language settings, as well as in experiments with non-words (Cutler, Van Ooijen & Norris Reference Cutler, Van Ooijen and Norris1999; Slowiaczek et al. Reference Slowiaczek, McQueen, Soltano and Lynch2000; Rapp & Samuel Reference Rapp and Samuel2002; Coch et al. Reference Coch, Grossi, Skendzel and Neville2005; Mitra & Coch Reference Mitra and Coch2018). It should therefore be possible to investigate rhyme in medieval texts that we can only access in written form, despite our uncertainty regarding their mode of delivery and reception, and the degree to which the audience was familiar with the sound and structure of new loanwords.
As an aside, it should also be mentioned here that there are similar effects that can be observed in the case of word-initial overlap, i.e., that it facilitates recall, memorisation and attention (Slowiaczek et al. Reference Slowiaczek, McQueen, Soltano and Lynch2000: 533; Lea et al. Reference Lea, Rapp, Elfenbein, Mitchel and Romine2008: 713; Egan et al. Reference Egan, Cristino, Payne, Thierry and Jones2020). In poetry, this is present in the form of alliteration, which was of course a very important feature of Old English poetry. And while there are fewer alliterating poems in the (early) ME period than in the OE period, there still are a few, some of which even combine both alliteration and end-rhyme. For this study, however, I will only be focusing on end-rhyme.
Another aspect of poetry that is relevant in this context is its specialised vocabulary. It is well known that the Old English lexicon has a large number of words that appear only once or very rarely, and only in poetry (Scragg Reference Scragg, Godden and Lapidge2013: 60). In Middle English, too, the corpus of texts associated with the Alliterative Revival ‘possesses a rich inventory of poetic lexis that is genetically derived from the specialised diction of Old English poetry’ (Pascual Reference Pascual2017: 250). In general, if such words are part of a poetic register and/or nonce formations created to fill a particular gap in the metre or rhyme scheme, we would not expect them to gain traction and become part of the common lexicon. Indeed, due to their limited use, these words are more likely to appear very short-lived in a study like this.Footnote 2
Nevertheless, given the overall cognitive benefits of rhyme for language processing, I assume that this could have had an effect in the ME period, too. More specifically, I assume that being used in rhyme would increase a word's overall salience, which would make it easier to remember and access, and thus increase its chances of survival.
3 Data
The corpus used for this study is the electronic version of the Linguistic Atlas of Early Middle English, 1150–1325 (LAEME), which contains about 650,000 words. In addition to syntactic and morphological information, the tagging in LAEME also includes information on whether a word appears in rhyming position. Approximately 52 per cent of all words in the corpus appear in verse texts, and 48 per cent in prose texts (the few alliterating text(s) were counted as verse for this purpose; this concerns 2,428 words in text #172 and 364 words in #136).Footnote 3
Since the focus is on nouns in rhyming position, I collected my data based on nominal suffixes: -ship, -ness, -hood, and -dom, the most frequent both in terms of types and tokens (Gardner Reference Gardner2014: 71), for the Old and Middle English subsets (i.e. words that are present in the ME lexicon which were already attested in OE, and words which consist of Germanic elements but are for the first time attested in ME, respectively), and -ity, -ery and -ment for the French subset. The lexelsFootnote 4 in LAEME follow the modern-day spelling for words that still exist, but in those cases where a word (both of French and English origin) did not survive until today, a non-standardised spelling is used for the lexel, e.g. vilte. In this case, the word goes back to Latin vilitas (Französisches Etymologisches Wörterbuch (FEW), s.v. vilitas, ‘lowness of price or value’), but the lexel's suffix is not spelled <-(i)ty>, which is the usual spelling in Present-day English for words derived from Latin words ending in -(i)tas (Durkin Reference Durkin2014: 239) such as dignity < dignitas. In order to be able to include such cases, as well as to have a subset that is comparable in size to the OE and ME ones, I expanded my criteria to include French nouns appearing in the LAEME lexel list that end in -i, -y or -e.
For each word, I also collected information about its first and last dates of attestation from the Historical Thesaurus of English (HTE). Where the HTE and LAEME disagreed, I took the earlier date of the two for the first attestation and the later date of the two for the last attestation. In LAEME, dates are given in the following format: C13b1 stands for the first quarter (1) in the second half (b) of the thirteenth century (C13). In cases where the LAEME dating was used, I chose the most extreme possible date for the given range. For example, the HTE dates scendness only as ‘OE’, meaning that it is only used in Old English texts, but LAEME has three attestations, the latest of which is dated to the first quarter of the fourteenth century (in text #286). In this case, I used 1325 as the date of last attestation.Footnote 5
Since LAEME also provides a localisation of a text's language wherever possible, I also included information about a word's geographical spread in my dataset. The localisation is given both as a placename as well as a six-digit National Grid reference (Laing Reference Laing2007) which essentially functions as a set of x- and y-coordinates whose origin point is located to the southwest of Cornwall. For example, steadfastness occurs in text #150, which is localised to West Norfolk with the grid reference 579 307. The first set of three numbers serve as x-coordinates and the second set as y-coordinates. When this information was given for instances in at least two different locations, I used these coordinates to determine the geographical spread of a word. To quantify this, I calculated the diagonal of a rectangle that encloses the most distant x- and y-coordinates. The advantage of using the diagonal over the area is that this more accurately represents the wide distribution of words whose coordinates on one axis happen to be very far apart while those on the other axis happen to be very close. Consider, for example, the geographical distribution of ointment and steadfastness (figures 1 and 2Footnote 6): ointment is attested both in the North and in the southern Midlands (marked by red X's), and steadfastness is attested in the East and West Midlands. The longer sides of the two rectangles (in green) are roughly the same length (255 km and 213 km, respectively). However, since the shorter side of the rectangle encompassing the two texts which contain ointment is only half the length of the shorter side of the rectangle encompassing the texts containing steadfastness (26 km and 52 km, respectively), its area is also only about half of the other (6,630 km2 and 11,076 km2, respectively). Using these numbers would give the misleading impression that steadfastness has twice the geographical spread of ointment. If we calculate the diagonal (in blue), however, the two are much more similar (256 km and 219 km, respectively). The log value of these diagonals was used for the analyses.
Figure 1. Geographical spread of ointment
Figure 2. Geographical spread of steadfastness
Finally, I recorded the absolute number of occurrences of each word in LAEME, as well as the number of times each word occurred in rhyming position. For the analysis, I used the log value of the absolute frequencies, as well as the rhyme-to-total ratio. One decision regarding the data collection that I had to take at some point was whether to include or discard duplicates of the kind where excerpts from the same texts from different manuscripts and/or dialect areas were included in LAEME. For example, swicolhood appears four times in LAEME, and is rhymed in all four cases, but only half of these are strictly speaking distinct because they are from the same text passage of The Owl and the Nightingale, which appears twice in LAEME (texts #2 and #1100). This distorts the rhyme ratio somewhat, but I felt it was necessary to keep these for the geographical spread as well as the absolute frequency counts.
The final dataset contained 118 words in the French subset, 164 in the ME subset and 193 words in the OE subset.
4 Method
In order to test whether rhyme had an effect on a word's chances of survival, I used two different approaches in R: survival analysis (SA) and random forests (RF) in combination with conditional inference trees. Both have been used for linguistic analysis in recent years (see, e.g., Tagliamonte & Baayen Reference Tagliamonte and Harald Baayen2012; Hundt, Rautionaho & Strobl Reference Hundt, Rautionaho and Strobl2020; Röthlisberger Reference Röthlisberger, Zehentner, Röthlisberger and Collemanforthcoming (on random forests); Ota & Green Reference Ota and Green2013; Smolík Reference Smolík2014; Carlson, Sonderegger & Bane Reference Carlson, Sonderegger and Bane2014; Burke, Morita-Mullaney & Singh Reference Burke, Morita-Mullaney and Singh2016; Van de Velde & Keersmaekers Reference Van de Velde and Keersmaekers2020 (on survival analysis)). Both approaches have, to my knowledge, only rarely been applied to historical data, especially historical English data.
In the random forests method, random subsets of the entire dataset are analysed to find out which combination of variables (here: suffix, geographical spread, frequency, rhyme ratio) leads to which outcome (here: survival or obsolescence). The analysis of one such subset can be visualised in tree form (see e.g. figure 17), and the total number of trees generated from a large number of subsets is called a ‘forest’. Observations in the dataset that were not used in this initial stage are then used to test how accurately the forest classifies the data: the results of the individual trees are aggregated and the predicted outcome is compared to the actual, recorded outcome. The random data sample in table 1 shall serve as an example to illustrate the process. First, it seems that the suffix -ness is a good predictor for survival because none of the words that are obsolete end in -ness, whereas 4 out of 6 of the surviving words do. Furthermore, in the same random sample, a high rhyme ratio seems to go hand in hand with survival, too, since the obsolete words have ratios of 16.6 per cent or lower, whereas half of the surviving words have a ratio higher than that. Similarly, the surviving words tend to have higher values for the variable Frequency: it is 1.8 or lower for the obsolete words, whereas all but one of the surviving words (mirkness) have a higher log frequency. Finally, only one out of three of the obsolete words in the sample (cwēadship) has a Spread value of equal-or-greater than 5.4, whereas the same is true for 5 out of the 6 surviving words. This leads to the preliminary assumption that words have a higher chance of surviving if they: (a) end in -ness; (b) have a rhyme ratio of over 16.6 per cent; (c) have a log frequency of over 1.8; and (d) have a geographical spread of over 5.4. This process is then repeated many times with different, random subsets, and the results are aggregated in order to determine the combination of criteria that are overall most effective for survival.
Table 1. Random sample of ME subset, censored at 1500
SAs are often used in medical studies (see examples in Kleinbaum & Klein Reference Kleinbaum and Klein2012: 4–5), for example to test the effectiveness of a medication over a placebo, or to compare the survival rates of two groups (e.g. men and women). Over the course of such a study, the time up to a particular event, such as the death of a patient, is recorded. The SA is conducted in order to determine whether certain variables (such as type of medication, age, gender, etc.) have a significant effect. In other words, it is used to test whether certain variables influence the patient's chances of survival, i.e. how much time passes between the start of the study and the event (i.e. death). If the observation period ends before the event occurs, the data are said to be censored at this point in time. Using table 1 as an example again, cweadship is attested about twenty years, from 1205 to 1225. This falls within the period of observation (which in this example is, for the purpose of illustration, 1150–1500), and the event (‘death’ or obsolescence) occurs after twenty years. For a word like manhood, on the other hand, which is attested from c. 1225 to beyond the period of observation, the survival time is only measured up to 1500 (i.e. 275 years), at which point the event (obsolescence) has not yet occurred, and so it is censored.
The survival probability calculated in this manner is visualised with Kaplan–Meier (KM) survival curves such as the one in figure 3, with time (in years) on the x-axis and survival probability on the y-axis. This figure shows the survival probability according to different levels of geographical spread (low, medium, high). For calculation purposes, absolute survival times are considered, that is, survival probability is calculated based on the length of attestation, regardless of when the different words first occurred during the observation period. Whenever the population or group size decreases because the event (i.e. death or obsolescence) occurred, the probability of a word in this group surviving up to this point in time decreases, and there is a step down in the ‘curve’. If a word is merely censored, i.e. if the length of its attestation is cut short not because it died out but because the time of observation ended, this is marked by a small vertical line in the curve. In figure 3, there are several drops in the middle curve (representing the medium category, in green), around the 100-year mark. This indicates that there are a number of words that fell out of use approximately 100 years after they were first recorded (regardless of the specific year in which they were first recorded). In the same survival curve, there are small vertical lines that cross it, e.g. around the 50-year mark or just after the 100-year mark. This means that there are words in the medium geographical spread category which were attested during only 50 and slightly more than 100 years, respectively, during the 200-year observation period, but they do not decrease the survival probability because they did not become obsolete during this time.
Figure 3. Kaplan–Meier curves for the geographical spread in the ME subset, censoring date 1350
With this method, only one (or very few, depending on the number of levels) categorical variable can be considered at a time. A more complex method, the Cox Proportional Hazards model, can be used to analyse several (non-categorical) variables at the same time. However, as the name implies, the model assumes that the hazard between groups is proportional, which is not the case for my dataset, as can be seen in figures 4–15, where the curves often cross each other and do not maintain equal distance from each other. This, among other things, led to issues with model fit, which is why the current study only uses KM curves.
Figure 4. Kaplan–Meier curves for the different rhyme ratio categories of the French subset, censoring date 1500
Figure 5. Kaplan–Meier curves for the different rhyme ratio categories of the ME subset, censoring date 1350
Figure 6. Kaplan–Meier curves for the different rhyme ratio categories of the OE subset, censoring date 1350
Figure 7. Kaplan–Meier curves for the different suffixes of the OE subset, censoring date 1350
Figure 8. Kaplan–Meier curves for the different suffixes of the ME subset, censoring date 1350
Figure 9. Kaplan–Meier curves for the different suffixes of the French subset, censoring date 1500
Figure 10. Kaplan–Meier curves for the different frequency categories of the OE subset, censoring date 1350
Figure 11. Kaplan–Meier curves for the different frequency categories of the ME subset, censoring date 1350
Figure 12. Kaplan–Meier curves for the different frequency categories of the French subset, censoring date 1500
Figure 13. Kaplan–Meier curves for the geographical spread in the ME subset, censoring date 1350
Figure 14. Kaplan–Meier curves for geographical spread in the OE subset, censoring date 1350
Figure 15. Kaplan–Meier curves for geographical spread in the French subset, censoring date 1500
Adapting the terminology of SA to the data of the present study, words are, metaphorically speaking, the ‘patients’, the event of interest is a word's ‘death’, for which I take the year of its last attestation, and the survival time is measured as the difference between its last and first attestation. The main variable of interest – the type of medication, if you will – is rhyme, which is measured as a word's appearance in rhyming position in proportion to the total frequency. Additional variables that were included are the frequency, the suffix and the geographical spread. Finally, the censoring date is 1350, shortly after the timespan covered by LAEME. However, because of how the data are distributed, the survival curves for the French subsets were not very enlightening; only seven out of 118 words had disappeared by 1350, and most of the French loanwords in my dataset entered the language after 1225, where we find a big spike in first attestations, with another big spike around 1280. This means that there is less time for observation than in the OE and ME set (which contain words that are already attested earlier). I therefore moved the censoring date for the French subset to 1500.
Numeric variables were used for the RF, but since categorical variables were needed for the SA, I transformed them as follows: there are four frequency categories, low (1 or 2 occurrences), medium-low (3–10 occurrences), medium-high (11–31 occurrences) and high (more than 31 occurrences). The ‘low’ category is intended for words that occur only once or twice in the corpus, and the ‘high’ category for words that occur exceptionally often. A cut-off point of 31 was chosen because in both the French and the ME subset, there is a gap after this, followed by only 4 words in each subset that have a much higher absolute frequency. The geographical spread was divided into three levels: low, medium and high. The ‘low’ category includes words with a spread value of up to 4.7, which in general amounts to occurrences that are contained within one dialect area.Footnote 8 The ‘medium’ category includes words with a spread value of up to 5.6, which is roughly equivalent to two large dialect areas. Any words with a spread value above 5.6 are in the ‘high’ category. For the rhyme ratio variable, I split the data into five categories: less than 20%, 20–40%, 40–60%, 60–80% over 80%.
The statistical tests were done in R using the survival (Therneau Reference Therneau2022) and survminer (Kassambara et al. Reference Kassambara, Kosinski and Biecek2021) packages for SA. For the trees and forests, I used the party (Hothorn et al. Reference Hothorn, Buehlmann, Dudoit, Molinaro and Van Der Laan2006; Strobl et al. Reference Strobl, Boulesteix, Zeileis and Hothorn2007, Strobl et al. Reference Strobl, Boulesteix, Kneib, Augustin and Zeileis2008), partykit (Hothorn & Zeileis Reference Hothorn and Zeileis2015) and pdp (Greenwell Reference Greenwell2017) packages. The RF analysis was done using conditional variable importance and the initial value of ntree was set to 500 and mtry to 2, which performed quite well for the French subset (C-index=0.923, accuracy=0.88, out-of-bag (OOB) C-index=0.79 and OOB accuracy=0.88), but not for the OE or the ME subset. Increasing ntree had little to no effect, but increasing mtry to 3 led to slightly better scores. However, the OOB C-index is still below the desired minimum of 0.7 for both ME (C-index=0.81, accuracy=0.74, OOB C-index=0.69, OOB accuracy=0.71) and for OE (C-index=0.8, accuracy=0.72, OOB C-index=0.67, OOB accuracy=0.64). There is no difference in variable importance between mtry=2 and mtry=3 for the OE subset, and only a slight increase in the importance of the Suffix variable for the ME subset with mtry=3. The trees were made using the default settings (minimum p-value=.05, minsplit=20, minbucket=7). The accuracy of the trees is somewhat low, as are the C-indices: 0.69 (both accuracy and C-index) for OE, 0.68 accuracy and C-index 0.66 for ME, and 0.91 accuracy and C-index 0.67 for French. Reducing the minimum sizes for nodes (minsplit=10 and minbucket=5) somewhat improved the performance for the ME data, but not OE or French. Since the OOB C-indices for the RF and the C-indices for the trees were quite low, especially for the OE and ME subsets, the results described in the following section should be seen as exploratory.
5 Results
5.1 Survival analysis
The result of the survival analysis is visualised by the Kaplan–Meier curves in figures 4–15. The curves for the rhyme ratio variable (figures 4–6) show that there is no indication that a higher amount of rhyme leads to a higher survival probability. In all three subsets, the survival probability for the >80 per cent category falls quite steeply and/or is one of the lowest at the end of the observation period. However, rhyme nevertheless does seem to have an influence on survival, albeit only in the French and Middle English subsets – for OE, the difference between the five ratio groups is not statistically significant. In the French subset (figure 4), words with a rhyme ratio between 20 and 60 per cent have the highest survival probability, whereas those with 60–80 per cent and more than 80 per cent rhyme ultimately have the lowest survival probability. The group of words with the lowest rhyme ratio (under 20 per cent) is somewhere in the middle all throughout the observation period. None of the five ratio groups’ survival probability drops below 50 per cent during the observation period. For ME (figure 5), the picture is slightly different: here, too, the 20–40 per cent group has the highest survival probability initially, but ends up around the 75 per cent mark together with the third and fourth ratio group (40–60 per cent and 60–80 per cent). In contrast, the groups at either extreme of the rhyme–ratio spectrum fare the worst: the group with less than 20 per cent rhyme has a very low survival probability from the very beginning and both it and the group with more than 80 per cent rhyme drop below 50 per cent survival probability approximately seventy-five years after first occurrence.
Some of the curves can be explained by an imbalance in the data (see table 2). For example, in all three subsets, there is a large number of words that appear only once, which means that these automatically fall into the under-20 per cent or the over-80 per cent group, because their rhyme ratio can only be 0 or 100 per cent. Words that appear only twice also tend to fall on one or the other end of the scale. Together, words that appear only once or twice in the corpus make up 42 per cent (81/193) of the OE subset, 61 per cent (100/164) of the ME subset, and 48 per cent (57/118) of the French subset. The proportions of those that also fall into the under-20 per cent or over-80 per cent group are similar: 41 per cent (79/19) of the OE subset, 60 per cent (99/164) of the ME subset, and 43 per cent (51/118) of the French subset. In the case of the OE subset, these instances tend to be the last attestations of a word, and so the under-20 per cent and over-80 per cent survival rates are automatically a lot lower than the others. In the French subset, the under-20 per cent group contains more words with higher frequencies, so the picture looks somewhat different here. Another example is the low survival probability of the 60–80 per cent group in the French subset: this group contains only six words, half of which happen to survive. Similarly, the sudden and complete drop in the curve of the same group in the OE subset (figure 6) can be explained by its containing only one word (lyþerness).
Table 2. Absolute frequencies per language subset for the five ratio groups. Percentages given in the rightmost column are the percentages of the totals of the respective subsets.
For the suffix variable (figures 7–9), the OE and ME subsets show slightly different behaviour. In the OE subset (figure 7), there is almost no difference between the four suffix groups in the first fifty to sixty years, after which they split in two: both -ship and -ness drop off, with -ness falling below the 50 per cent survival point after around 125 years (i.e. around the year 1275–1300),Footnote 9 whereas more than 75 per cent of words ending in -dom/-hood are still in use by the end of the observation period (1350). In the ME subset (figure 8), all four suffix groups show a steep drop very early on, with more than 50 per cent of -ship words disappearing within fifteen to twenty years of their first attestation. The -hood group, which had a high survival probability in the OE subset, also falls below 50 per cent around seventy-five years after the first occurrence, while -dom and -ness have relatively similar survival curves that stay above 50 per cent until the end of the observation period. In the French subset (figure 9), almost all suffixes maintain a very high survival probability. The notable exception here are words ending in -ery (table 3), whose survival probability decreases quite quickly around 175–200 years after first attestation, after which it stays just above the 50 per cent mark.
Table 3. Words in the dataset ending in -ery, sorted by spread. Note that those words still in use at time of censoring, i.e. 1500, have that date as their ‘last attestation’, but are marked as ‘not obsolete’ at this point.
The patterns for frequency (figures 10–12) and spread (figures 13–15) are quite similar, which is to be expected because they correlate (Pearson's R = 0.8 for OE and 0.76 for ME, p<0.05 for both). The overall tendency is very clear in both subsets: words with higher frequency and geographical spread have a higher survival probability, words with low frequency and geographical spread have a lower survival probability. In the OE subset, we also see a similar pattern to that of the suffix graph: there is not much difference between the different levels of frequency/spread in the first fifty or sixty years, after which they clearly separate. There seems to be no difference between the low-medium and medium-high-frequency words. The low-frequency/spread words drop below the 50 per cent mark after about 100 years (i.e. around 1250). In the ME subset (figure 11), the survival curve for the low-medium-frequency words is very similar to that of the low frequency words, and both of them drop below 50 per cent survival probability after 100 and 75 years, respectively. In the French subset (figure 15), the words with lower geographical spread have a significantly lower survival probability than those with medium or high spread (p=0.0017), but it is still relatively high compared to OE and ME. For frequency (figure 12), there are only small differences which are not statistically significant, but it is again the group with the lowest frequency (one or two occurrences) whose survival probability decreases the most.
5.2 Random forests and conditional inference trees
The results of the RF analysis allow us to rank the variables according to their importance and show which variables are important and which ones are not. To complement the variable importance plots, conditional inference trees were used to visualise interactions between the variables. In some cases, partial dependence plots are used to more closely evaluate individual variables. Each of the three language subsets and their respective forests and trees are discussed in order, starting with OE and ending with French.
In the OE subset, Spread and Suffix are important predictors, while Ratio and Frequency are not (figure 16). The tree (figure 17) did not show any interactions between Spread and Suffix, and instead the importance of Spread seems to outweigh everything else. It did, however, split up the spread variable into three groups, which are the three bar-plot leaf nodes; dark grey represents the proportion of words that survived, light grey those that did not. The first split (node 1) is made between those with a spread of 5.67 and above, which have almost a 100 per cent survival rate, and those below. This group is split again (node 2) into those with a Spread value between 4.47 and 5.67, and those below 4.47. The former have a survival rate of about 50 per cent, the latter around 30 per cent. These are, coincidentally, roughly the same as the cut-off points chosen for the categorical spread variable. Suffix does not show up in the tree, even though is an important predictor according to the RF's variable importance plot – albeit much less important than Spread. A partial dependence plot for Suffix (figure 18) shows essentially the same as the KM curves: -ness is most likely to become obsolescent (coded as 1), followed closely by -ship, whereas -hood and -dom are more likely to survive (coded as 0).
Figure 16. Variable importance in the OE subset
Figure 17. Conditional inference tree for the OE subset, censoring date 1350
Figure 18. Partial dependence plot for Suffix in the OE subset. The y-axis (‘yhat’) represents the predicted survival probability.
For the ME subset (figure 19), too, Spread is the most important variable, followed by Suffix, while Frequency is clearly unimportant. Ratio is very close to the red line, i.e. very close to being unimportant, and a partial dependence plot (figure 20) shows that words with higher rhyme ratios tend to survive more than those with lower rhyme ratios, but the difference is very small. Ratio also does not appear in the tree (figure 21), which only includes Spread and Suffix. The first split is between those words with a Spread higher or lower than 4.362 (node 1). Those at or below this level have a roughly 30 per cent survival rate (node 2; dark grey represents surviving words, light grey obsolete words). Among the group with a higher geographical spread, the tree splits further (node 3) on the Suffix variable: those ending in -dom or -ship (node 4) have a similarly low survival rate to those words with lower geographical spread, but this node contains only eight words and is therefore only applicable to a very small part of the dataset. Words ending in -hood or -ness (node 5), on the other hand, have a survival rate of just over 80 per cent.
Figure 19. Variable importance in the ME subset
Figure 20. Partial dependence plot for Ratio in the ME subset. The y-axis (‘yhat’) represents the predicted survival probability.
Figure 21. Conditional inference tree for the ME subset, censoring date 1350
The RF analysis of the French subset shows that both Suffix and Spread are important predictors for survival (figure 22), but unlike the OE and ME subsets, Suffix is more important than Spread. According to the tree (figure 23), those with a medium to high geographical spread (split at node 1) have a 100 per cent survival rate (node 5), but among those that fall below a certain threshold (3.748), there is another split on Suffix (node 2): those ending in -ery (node 3) fare decidedly worse, with a survival rate of roughly 30 per cent, than those ending in other suffixes (node 4), with over 80 per cent. This matches the survival curves for the French suffixes, which also indicated that -ery had a significantly lower survival probability than the other suffixes. Since the tree suggests that this may also be dependent on the geographical spread, it is worth having a look at the -ery group, which is given in table 3. Of the 12 words ending in -ery, 5 are obsolete before the end of the observation period (1500). Unfortunately, all of these are attested only between one and three times in LAEME, which automatically makes it very difficult, if not impossible, to gauge the extent of the geographical spread.
Figure 22. Variable importance in the French subset
Figure 23. Conditional inference tree for the French subset, censoring date 1500
Overall, the results of the RF analyses and SAs complement each other. In all three subsets, geographical spread and suffix were the most important and most robust predictors for survival, usually in combination with each other. The other variables are not as important according to the RF's variable importance plot, but there nevertheless seem to be (for the most part) significant differences between the different levels of these variables, too.
6 Discussion
Finally, I would like to have a closer look at some of the cases that fall outside the expected pattern, such as words that survive despite low geographical spread, or words that fall out of use despite high geographical spread. The OE words that survive up to or beyond 1350 (table 4) despite an apparent disadvantage seem to be mostly deadjectival nouns, and their adjective bases are for the most part still in use today. This is the case, for example, with soreness and sore, blindness and blind, or leanness and lean. Among the obsolete words, on the other hand, we find for example anrædnes, arfæstnes, diegelness or hluttornes, which are also derived from adjectives, but their adjective bases have either already fallen out of use by the beginning of the ME period (hlutor, arfæst), or do so soon after (anræd c.1230, diegol c.1275, HTE s.v. anræd, diegol). A few words in this group, bishophood, earldom, lorddom and lordship, are not based on adjectives but on terms of rank or office that still exist today. One factor that is often mentioned in connection with lexical loss is that a term is more likely to fall out of use if the concept or object it denotes is no longer relevant or in use (e.g. Rundblad Reference Rundblad2000: 40; Dworkin Reference Dworkin1989: 335), and so the continued relevance and existence of these ranks and titles most likely contributes to their survival to this day. The trend of deadjectival nouns surviving longer alongside their adjectival bases agrees with other observations made by Rundblad (Reference Rundblad2000: 39–40), who points to the importance of transparency and the connection between the base and the derivation; that and having both the derivative and the base available in the lexicon at the same time seems to facilitate frequent use and thus survival (Rundblad Reference Rundblad1998: 171; Reference Rundblad2000: 40). Similarly, Dworkin (Reference Dworkin1989: 336) observes in a study on Old Spanish that ‘derivationally transparent lexical items depend heavily on the head of the word family to which they belong … In the case of many adjectival abstracts, the loss of the primitive (almost) inevitably led to the demise of the corresponding substantive(s).’
Table 4. Words in the OE subset that survive up to or past 1350, where spread < 4.47. Words that have no last attestation date are still in use today.
In the ME subset, words that have disappeared despite being attested in a wide geographical area and ending in -hood or -ness are lordhood, wodhood, holihood on the one hand, and skillwiseness, untrueness, dreadness on the other. For those ending in -hood, there is usually a synonym ending in -ship, -dom or -ness in use at the same time. In addition to lordhood, there are lorddom and lordship (attested since OE). In the case of holihood, too, halidom and holiness are attested since OE, and next to wodhood, wodness is attested since OE. While a co-existing base word and its derivative may reinforce each other in terms of regularity and frequency of use (Rundblad Reference Rundblad2000: 39–40), the same is apparently not true for the existence of parallel derivatives like lordship/-hood/-ness and holihood/-dom/-ness. Lindsay & Aronoff (Reference Lindsay, Aronoff, Hathout, Montermini and Tseng2013: 135) point out that ‘in general, languages do not tolerate true synonymy’, and these synonyms seem to be synonymous enough to be sometimes interchangeable and thus in competition with each other (Gardner Reference Gardner2014: 29–30). Among the words ending in -ness, it is notable that one of the available synonyms is always the simplex from which it is derived or to which it is related: skill for skillwiseness, untruth for untrueness and dread for dreadness.
The next category in the ME subset that stands out is the group of words that survive despite a smaller geographical spread. The majority of these, too, ends in -hood or -ness and their bases generally survive(d) alongside the derivatives. In the ME period, -ness is ‘the most frequent suffix across the board chronologically and geographically’ (Dalton-Puffer Reference Dalton-Puffer2011: 128). Since geographical spread was the most consistently significant predictor of survival in the SA, the general wide geographical distribution of -ness may thus be the reason why -ness is one of the factors that is associated with high survival rates in the tree (figure 21) as well as the SA (figure 8).
The one factor that stood out in the French subset is the significantly higher disadvantage of words ending in -ery in combination with low spread. Indeed, half of the words ending in -ery have become obsolete by 1500, but most of these, as mentioned above, only appear very few times in the corpus, most likely because they only appear towards the end of the period covered by LAEME, which often makes it impossible to determine the geographical spread. I suspect that the cause lies in the small size of the dataset and the unfortunate absence of other infrequent words ending in -ery. Gadde (Reference Gadde1910: 27–31) lists several derivations in -ery, mostly from native English bases, from the early fourteenth century that are not in LAEME. Probably because his scope is larger than that of the present study or that of Gardner (Reference Gardner2014), he also finds a much larger number of words in -ery, and a relatively high proportion of derivatives formed within English: ‘Out of more than 600 new-formations with -ery (-ry) in my word-list more than 250 are from native roots’ (Gadde Reference Gadde1910: 13). This suggests that the heyday of -ery only began at the beginning of the fourteenth century and that my corpus was not suited for an SA in this case.
There are few other ways in which the make-up of my dataset may have skewed the results. Due to the nature of the suffixes chosen for purely pragmatic reasons, i.e. to make data extraction easier, the words included in my dataset denote for the most part concepts that are abstract, and thus rather general (see, e.g., Trips Reference Trips2009: 188, who also equates abstractness with generalness). Sylvester et al. (Reference Sylvester, Tiddeman and Ingham2020) looked at different levels of specificity in a handful of semantic domains and found the more general levels to have a greater number of synonyms (cf. graph 2). In other words: the more specific a sense, the more likely it is that there are only a few (one to three) co-hyponyms, or synonyms at the same level of specificity, and the more general a sense, the more likely it is that there are a lot of synonyms available. The more abstract and general the meaning of words in -ity, -ness, -ship etc., the more likely they are to have a large number of synonyms, and thus more competition. And while these suffixes may have acquired rather specific meanings/semantic niches in Present-day English, they probably used to be more general at the time of borrowing (cf. Gardner Reference Gardner2014: 191 on -ity and -ness; Lloyd Reference Lloyd2011: 35–9 on -ment).
7 Conclusion
The original aim of this article was to investigate whether frequent usage in rhyme potentially increased the survival chances of medieval English nouns. Since the focus was on rhyme, the data were chosen based on suffixes, that is, -ship, -hood, -dom, -ness for Old English words and new Middle English formations, and -ity/-e, -ment, -ery/-i for French loans. However, among the variables that were included, geographical spread and suffix stood out as most significant across these three subsets. For the individual subsets, there were some additional significant factors: -ness and -hood in the ME subset leading to higher survival, and -ery in the French subset leading to lower survival. The effect of the suffixes in the ME and French subset require more study since they may be significant only due to the limitations of the dataset and the corpus from which the data were extracted (in the case of -ery, which only appears towards the end of the timespan covered by the corpus) or they may be secondhand effects of another factor (overall wide geographical distribution in the case of -ness). The C-indices for the trees and forests in this study were also rather low, so the significance of these factors will need to be re-examined with models that perform better.
In addition to addressing these issues, future research would benefit from including additional factors which were not taken into account in the present study but stood out as potentially important in the discussion section, such as the parallel survival of the derivatives’ bases and the number and specificity of available synonyms. Overall, the influence of rhyme on survival is likely to be restricted to individual words, as in the cases of christianity (Käsmann Reference Käsmann1961: 99) or onde (Timofeeva Reference Timofeeva, Lorido, Prado-Alonso and Rodríguez-Puente2020b), but does not seem to be generally measurable in the Middle English lexicon, unlike the general positive effect of geographical spread.
Open access
