2. Terasawa Reference Terasawa1994: Data and Database Overview

2.1. Terasawa’s (1994) Metrical Constraints

Terasawa’s (1994) expanded empirical database of poetic NNs makes it possible to identify two types of NNs regarded as unsuitable for placement in verse (Terasawa Reference Terasawa1994:11–12, passim; 2011:7–76). The examples in 1 contain compound types avoided in the verse corpus.

Footnote ¹

Terasawa referred to the consistent avoidance of these patterns in OE verse as “metrical constraints”; in subsequent OE scholarship, they are commonly known as Terasawa’s Law (Hutcheson Reference Hutcheson1995, Russom Reference Russom and Jane Toswell1995, Suzuki Reference Suzuki1996, Fulk Reference Fulk2007, Goering Reference Goering2016a). We refer to the constraints exemplified by the data in 1 as Terasawa’s Constraints (TCs). Their modified version is provided below:

Terasawa’s presentation of NNs usage separates Beowulf from the rest of the poetic corpus, with further subdivisions of the material based on genre. The strikingly low occurrence of -LX compounds in the entire poetic corpus is initially puzzling; other compounds of that shape, while rare in verse, appear to be used freely in prose. His limited coverage of the patterns of NNs in the prose corpus (1994:60–62) overviews the distribution of the roughly synonymous words for ‘leader/chieftain/ commander’: here-toga (LX-LX) versus folc-toga (H-LX), here-wīsa, and here-tēma (LX-HL), showing a lopsidedly more frequent use of here-toga in prose: n=161 occurrences in prose versus n=3 elsewhere.Footnote ² The example of prosaic here-toga and poetic folc-toga, a brief comparison of the Prose Boethius versus the Meters of Boethius (1994:61–62), and a commentary on the preference for uninflected forms for the second element in verse, such as ende-dæg ‘last day’, here-paþ ‘main road’ versus inflected ende-dæge(s), here-paþe(s) common in prose and charters, leads him to the conclusion that TCs “are metrical or poetic rather than phonological” (Terasawa Reference Terasawa1994:62–63). We believe this to be a well-reasoned inference, yet it leaves open the question of how the verse data fit within a much more inclusive picture of observed versus expected patterns of compounding in the entire lexicon. To answer that question, we look at some philological information on the data included in our analysis.

2.2. Stem Variability in the NN Data

A key aspect of the analysis of the NNs distribution in the compounding data is the metrical treatment of sonorant-final stems subject to parasiting (anaptyxis/epenthesis) and syncope, in which the sonorant is either consonantal or syllabic. Thus, an originally monosyllabic word such as wundor ‘wonder’, Old Norse (ON) <undr> would have been subject to parasiting by the 7th century, but it can scan as mono- or disyllabic: Beo 920a versus Juliana 575b (Fulk 1992:66–91, Reference Fulk2007). The metrical treatment of such forms supports Fulk’s position that the stability of etymologically monosyllabic HR, orthographic <H(V)R> stems, realized as a single H in Beowulf and in other early verse, is significant for the dating of the poetic corpus. Presonorant epenthesis in both HR and the much rarer LR forms was early, variable, and gradual. In some items, such as ātor ‘poison’, ON <eitr>, meðel ‘assembly’, Gothic <maþl>, this can be tested in the late verse, where rarer monosyllabic forms must be related to formulaic diction, which makes disyllabic realization of such forms in speech highly probable.

Variability in speech extends to originally disyllabic LH stems, such as gomen ‘mirth’ and heofon ‘heaven’, which are considered targets for what Terasawa (Reference Terasawa1994:27, 41 and passim) identifies as “pseudo-epenthesis.” Terasawa (Reference Terasawa1994:12) isolates those variable stems, yet he ultimately keeps them as part of the argumentation for TCs on the assumption that the “opaque” linguistic situation would allow such disyllabic stems to be reanalyzed as monosyllables by analogy with etymological sonorant-final HR monosyllables, such as wundor ‘wonder’. For these originally disyllabic items, too, variability in speech is a reasonable but challenging to test hypothesis.Footnote ³

The syllabic count of the sonorant-final stems can be ambiguous in the verse; the whole issue has attracted much attention and conflicting opinions (see Lehmann Reference Lehmann and Allan1968; Russom Reference Russom1987; Fulk Reference Fulk1992, Reference Fulk2007; Terasawa Reference Terasawa1994; Hutcheson Reference Hutcheson1995; Suzuki Reference Suzuki1996; Fulk et al. Reference Fulk, Bjork and Niles2008 inter alia). The debatable issues revolve around what counts as metrical, since ortho-graphically, these items are mostly written as <-VR>, with the vowel letter either underdotted or not, depending on the editorial decision. Outside verse, the evidence for variable syllabicity of the sonorants in OE LVR stems such as befer ‘beaver’, hræf(e)n ‘raven’ versus heofon ‘heaven’, of(e)n ‘oven’ is only testable in relation to lengthening of the originally short stressed vowel in later English, which is crucially dependent on the nature of the second syllable onset (Minkova & Lefkowitz Reference Minkova and Lefkowitz2020). If the confirmation of TCs rests on verse, the syllabic shape and count of sonorant-final stems must be treated as a metrical issue, regardless of how such sequences were realized in regular speech or in oral performance.

Terasawa (Reference Terasawa1994) also syncopates the middle syllable in a potentially and orthographically trisyllabic second member of the compound—XLX—as in æppel-fealụwe ‘apple-fallow’ (Beo 2165a) or hrīmgicẹlum ‘by rime/hoar-icicles’ (Seafarer 17a). This is another paraphonological variable, and we followed Terasawa in excluding compounds with a trisyllabic second member from our consideration to avoid this complication.

Separating Beowulf, a composition identifiable as “early” on philo-logical grounds, that is, 8th century (Fulk Reference Fulk1992, Fulk et al. Reference Fulk, Bjork and Niles2008), from the rest of the corpus is justifiable. Beowulf is exceptional in many ways. Fulk et al. (Reference Fulk, Bjork and Niles2008:cxii) count 903 “distinct substantive compounds, of which 518 are peculiar to Beo… [their] deployment evinces considerable inventiveness and poetic skill.” Remarkably, one third of the whole Beowulf vocabulary is compounds. This is part of the overall “considerable statistical difference between Beowulf and presumably later poems” (Fulk Reference Fulk1992:79). Terasawa (Reference Terasawa1994) identifies only n=13 HX-LX and n=9 LX-LX compounds in Beowulf, all of them considered doubtful on the basis of epenthesis, syncope, or alternative readings, except for mere-faran ‘sea-farer.gen.sg’ (Beo 502a). The creativity with compounding and the manifest sensitivity to the importance of weight to the meter suggest that the choices are at least partly specific to the Beowulf poet/scribe.

For the remaining approximately 60 texts, Terasawa (Reference Terasawa1994) gives only individual lists and numbers, no totals, classifying the first elements of NNs into HX-, LX-, and H-.Footnote ⁴ His Appendix 3 (1994:119–123) only lists NNs that do not conform to the HX-LX or LX-LX pattern, again excluding items with potential epenthesis, syncope, alternative readings. However, there is a separate list of NNs with “pseudo-epenthesis”—the abovementioned items, which are originally disyllabic, such as gomen ‘mirth’ and heofon ‘heaven’.

Our count shows n=27 HX-LX items (n=20 with pseudo-epenthesis of the first member, that is, HX=H), and n=80 LX-LX items (n=70 with pseudo-epenthesis), including 39 instances of heofon-cyning ‘king of heaven’. The unexpected density of this lexically specific compound is fully in line with Fulk’s (2007) reference to such usages as instances of formulaic diction.Footnote ⁵ Occasionally, the OE scops create their own unique compounds; they are sampled in 3, where the a and b in the texts refer to the position of the verse in the alliterative line, and not to one of the Sieversian alphabetical types A–E.Footnote ⁶

The rare and unevenly distributed exceptions to TCs, the open questions about the syllabicity of sonorant-final stems, and the existence of specifically poetic -LX NNs invite further examination of the rationale behind the very limited use of such compounds in verse. TCs link up directly to a more general question: Do TCs constitute evidence that poetic compound formation is a systematic transfer of weight-based patterns of compounding from the general OE grammar to the poetic grammar? We now move to a previously unexplored area of inquiry: How does the avoidance of -LX NNs in verse relate quantitatively to the available lexicon?

3. NNs in the OE Lexicon: Corpus and Simulation Details

The avoidance of -LX NNs in the verse, a prosodic frame required for metrical resolution (LX=H), has direct implications for the recon-struction of stress placement in OE, since resolution also requires stress on the light syllable. Traditional accounts (Sievers Reference Sievers1885a,b, 1893; Campbell Reference Campbell1959; Bliss Reference Bliss1967; Suphi Reference Suphi1988; Hogg Reference Hogg and Hogg1992; Lass Reference Lass1994; Hutcheson Reference Hutcheson1995; Page Reference Page2006; Minkova Reference Minkova2008; Fulk Reference Fulk2014; Russom Reference Russom2017) describe OE stress as syllabic and morphologically-based. That position is not shared by all researchers: The view that resolution in the meter is evidence for a weight-based account of stress is also widespread. McCully & Hogg (Reference McCully and Hogg1990) use resolution specifically in arguing for a weight-sensitive secondary stress; their account of primary stress is morphological. Dresher & Lahiri (1991) use metrical resolution as support for positing a moraic account of stress, similarly to Fikkert et al. (Reference Fikkert, Dresher, Lahiri, van Kemenade and Los2006) and Tanaka (Reference Tanaka1991, Reference Tanaka1992). The status of resolution as a phono-logical component of OE is discussed skeptically in Minkova & Stockwell Reference Minkova and Stockwell1994 and Stockwell & Minkova Reference Stockwell and Minkova1997.

Such divergent views prompt a critical examination of the association between TCs, resolution, and stress, as there are in truth many reasons why one might find TCs plausible and their existence sustainable. After all, it is striking that poets would refrain from using specific NNs when such compounds are characteristic of the vocabulary as a whole: Are they truly avoiding these structures for metrical reasons? Are they avoiding -LX elements in subordinate ictic positions only for reasons of weight or perhaps because of some other factors?

One can argue that TCs are directly linked to the weight-to-stress constraint in the general grammar, if the compounds found in the lexicon do not have the same distribution of shapes predicted by a grammar that randomly pairs first elements with second elements. We hypothesized, however, that TCs are not applicable to just the verse but to the lexicon as a whole. Thus, we asked whether these structures are avoided in the verse specifically or if there is a general avoidance of these compounds in the lexicon. To test this hypothesis, we simulated one million Old English compound lexica using Python 3 and compared the simulation results to the attested lexicon.Footnote ⁷

We gathered every headword labeled “noun” in DOE that contained a hyphen and was thus marked as a compound. We then removed a variety of edge cases. In programming, an edge case typically involves input values that require special handling. Here the special handling was removal. The nouns were removed if they:

1. (i) were marked as hypothetical or of doubtful meaning, marked “?” (n=102);

2. (ii) had at least one Latinate member, such as Aprelis-mōnaþ (n=12);

3. (iii) had at least one member with the prefixes ge- (n=362), be- (n=6), or for- (n=69);

4. (iv) had at least one member with the suffixes -end(e) (n=111), -nes (n=270), or -ing/-ung (n=281);

5. (v) were a triple compound (n=61).

With the triple-membered compounds, we further included their two-membered compound member if the word was decomposable as [member₁ + [member₂ + member₃]] and member₂ began with a letter after I. This was because the DOE does not include headwords beginning with L and onwards, so this was an opportunity to include NNs that were otherwise unavailable (n=12). Additionally, we conducted a crosscheck of these inferred two-membered NNs with inferred NNs that begin with A–I; these NNs were found to be already in the database, thus supporting the validity of our addition.

Some entries were excluded for multiple reasons, resulting in a final dataset of 4,076 NNs with only mono- and di-syllabic members, starting from an initial corpus of 5,338. These were then coded for the phono-logical shape of their members (H, HH, LL, etc.). For additional philo-logical details on the corpus, see the comments in section 2 and the preface to the supplementary material. For the corpus itself, see the supplementary material (doi.org/10.17605/OSF.IO/J9FY7).

For each simulated lexicon, we generated 4,076 words—the size of the relevant corpus. We simulated the outcome of a lexicon that combined members entirely randomly as approximated by using the relative frequency of appearances in our corpus as the probabilities of a particular shape being chosen. For example, within our simulation, the probability of the first compound member having the phonological shape H was the probability of a noun with the phonological shape H occurring as the first compound member throughout the corpus, and not the probability of an H-shaped noun occupying any position within a compound.

The code used for our simulation is available in the supplementary material. However, some slight exposition is desirable for our imple-mentation of frequency matching and our choice for the size of our simulation. To implement frequency matching, we first tallied up all the nouns with the same phonological shape, for each shape, for both positions. Table 1 shows the results.

Table 1. NNs in the DOE.

Functions were defined that would select a shape with faithfulness to the distribution of the corpus of 4,076 potential inputs to NNs. Thus, for the first member, we defined a function mem_1_type() that generates a number between 1 and 4,076, inclusive. If the number is ≤ 2,526, the function output is “H”; if 2,526 < number ≤ 3,420, the output is “HX”; for all other numbers, the output is “LX”.Footnote ⁸ We defined an analogous function for the second member. In this way, we did not simulate words but rather word shapes, as the simulated words were immaterial.

The outputs of these two functions were concatenated 4,076 times, and the counts of the resultant compound shapes were tracked. This entire process was repeated 1 million times. Each simulated lexicon was kept distinct, allowing us to count the number of lexica in which there were exactly n of any given compound shape.

Why did we choose to simulate 1 million lexica? In statistics, according to the central limit theorem, distribution of the outputs of a random process, repeated again and again, tends toward a normal distribution as the number of repetitions approaches infinity. For our purposes, the higher the number of repetitions, the more confident we can be that we have truly captured a representative sample of random lexica, including the statistically unlikely ones. As we suspected that the extant OE compound lexicon was an unlikely one, we wanted to emphasize this point. Why not give the most accurate model feasible?Footnote ⁹

4. Statistical Analysis

The 1 million simulated lexica were imported into R. We show below probability density histograms. We can interpret the simulated lexica with respect to the OE corpus by considering that each simulated lexicon is the result of 4,076 frequency-matching simulations of the creation of a compound. Each lexicon has some number N of compounds of a given shape; the normalized frequency distribution of those NNs is the probability density distribution that a frequency-matching simulation will contain N compounds of a given shape. For example, 5,478 of our simulated lexica had 460 H-HX compounds. Thus, in our simulations, the probability of there being 460 H-HX compounds is 5,478/1,000,000, or 0.5478%.

Furthermore, comparing the observed number O of compounds of a given shape to the density distribution tells one how likely it is that O was the result of a frequency-matching grammar—approximately 0.5% of frequency-matching grammars are expected to have exactly 460 H-HX compounds. Given the central limit theorem, 95% of all simulated NNs fall within ±1.96 standard deviations (SD) from the mean of N. Thus, O is not likely to have been the result of frequency-matching if O is >1.96 or <-1.96 SD from the mean of N. If O is >1.96 SD from the mean of N, it is likely that some factor in the grammar of OE favored the formation of compounds of this shape. If O is <-1.96 SD from the mean of N, some factor likely disfavored the formation of compounds of this shape. The mean number of simulated H-HX compounds was 492 and the SD was 21, but the observed number of H-HX compounds was 460. The observed count is -1.57 SD from the mean, so we do not have reason to believe any factors favored or disfavored H-HX compounds.

In each histogram, we have marked the observed (true, attested) number of compounds of the relevant shape with a solid red line. We have also included a dotted curve corresponding to the normal distribution that best fits the data and a dotted line corresponding to its center.

4.1. Evidence for TC1 and TC2

For convenience, we repeat the modified version of TCs in 4 below.

TC1 states that compounds of the shape HX-LX are to be avoided. Thus, we ask if the number of attested HX-LX compounds is lesser than the expected number of such in a frequency-matching grammar. In figure 1, the observed count is -5.28 SD away from the mean of the simulated counts, providing evidence for TC1.

Figure 1. Probability density histogram of simulated HX-LX counts.

As shown in figure 1, the number of HX-LX compounds in the OE corpus is 105. However, given our corpus, we expect purely frequency-matching grammars not subject to any phonological constraints to contain an average of 172 HX-LX compounds with a SD of 13 compounds.Footnote ¹⁰ Accordingly, approximately 0% of frequency-matching grammars have 105 or fewer HX-LX grammars. This result thus shows that some factor disfavored HX-LX compounds, providing evidence that TC1 is observed across all of OE, not just in verse.

TC2 states that compounds of the shape LX-LX are to be avoided. Thus, we ask the same question as above, but for this shape. In figure 2, the observed count is -3.95 SD away from the mean of the simulated counts, which provides evidence for TC2.

Figure 2. Probability density histogram of simulated LX-LX counts.

Figure 2 shows that our simulations predict an average of 127 LX-LX compounds with a SD of 11 compounds. We observe 83 LX-LX compounds, and approximately 0% of our simulations have 127 or fewer such compounds. TC2 as a factor in all of OE is thus also supported.

4.2. H-LX as Repair for HX-LX

Terasawa (Reference Terasawa1994:11–12) suggests that compounds of the shape HX-LX undergo repair (through the choice of paraphoological variables due to, for example, truncation, the use of pre-epenthetic forms, and syncope; see sections 2 and 3) to become H-LX. Though he does not formulate this repair requirement as part of TCs, the issue is clearly germane. If H-LX is potentially the outcome of a repair procedure, the question arises: Given that we have found evidence that TC1 and TC2 both generally hold across OE, is H-LX a generally favored shape or is it a result of repair in the verse? Figure 3 shows that the former is the case. The observed count in figure 3 is 5.38 SD away from the mean of the simulated counts, showing that H-LX compounds are quite over-represented in the OE compound lexicon when compared to a frequency-matching grammar. Figure 3 shows that our simulations predict an average of 488 H-LX compounds with a SD of 21 compounds, but we observe 600. Approximately 100% of our simulations have 600 or fewer such compounds, showing that H-LX compounds are, in fact, statistically overrepresented across the OE lexicon.

Figure 3. Probability density histogram of simulated H-LX counts.

It could be suggested that the reason that H-LX is overrepresented has something to do with the shape of the individual members. If this were so, H-LX and LX-H would be expected to behave similarly. However, our simulations show that the attested LX-H count (424) is only 1.2 SD greater than the mean, meaning that LX-H is not statistically over-represented; we do not have cause to believe that H-LX and LX-H behave similarly.

4.3 A General Constraint Against XX-LX Compounds

Since our simulations provide evidence for both TC1 and TC2, can they be collapsed into a single general prohibition against all compounds with a disyllabic first member and an LX second? As figure 4 reveals, the answer is a resounding yes. In figure 4, the observed count is 6.7 SD away from the mean of the simulated counts, showing that XX-LX compounds are greatly underrepresented in the OE compound lexicon when compared to a frequency-matching grammar.

Figure 4. Probability density histogram of simulated XX-LX counts.

Our simulations predict an average of 300 XX-LX compounds with a SD of 17, but the observed count is only 188. This outcome is expected approximately 0% of the time in grammars with no generalized TCs.

Note too that the observed count is -6.7 SD from the simulated mean, a more extreme value than that for HX-LX (-5.28) and LX-LX (-3.95). This suggests that it is more parsimonious to combine TC1 and TC2, as the effect of a combined prohibition is overall greater than two separate ones: -5.28 SD means that there is less than a 1 in 100 million chance that the observed data are from a frequency-matching grammar; -6.7 SD means less than 1 in 100 billion.Footnote ¹¹

One might go one step further and ask if the dispreferred phono-logical shape is LX instead of -LX. To answer this question, we looked at the difference between HX-LX compounds, which are dispreferred, and LX-HX compounds. The idea was that if LX-HX were also dispreferred, the LX shape might indeed be the crucial point. As it turns out, LX-HX is within the predicted range. Our simulations predict a mean of 128 such compounds with a SD of 11. The observed number is 149, which is 1.89 SD from the simulated mean. However, though 149 is within the predicted range (recall that the threshold number is 1.96), this should be taken with a grain of salt as the observed count is still greater than that found in 97% of the simulations. Regardless, it appears that LX- is different from -LX, as LX-HX is marginally preferred, while HX-LX is severely dispreferred. Recall also (section 4) that H-LX and LX-H do not behave similarly.

Based on all of these results, it is abundantly clear that the OE compound lexicon itself observes TCs. In fact, we have found that a generalized constraint against XX-LX forms explains the data even better than the separate components of TCs. We must conclude that TCs account for the general pattern of compound formation in the language, while the creation and selection of compounds in the verse is additionally circumscribed by specific poetic and formulaic diction, as well as incompatibility with acceptable metrical templates.

5. Discussion

The statistical outcomes regarding OE NNs show that the linguistic basis of poetic compound-formation is not different from compounding in the ambient language. The modest pool of XX-LX compounds that the scops could draw on preconditions their limited use in poetry, where the restrictions are further enhanced by metrical provisos. In metrical studies, the matching of the language’s vocabulary to its meter(s) is covered by the principle of Fit. However, Fit is “not an inviolable constraint that a meter must satisfy… For historical, cultural, or other reasons languages may have meters that fall short of naturally accommodating all their words” (Hanson & Kiparsky Reference Hanson and Kiparsky1996:294). Compounds of the -LH shape are equally available to prose writers and poets, yet the OE scops are constrained further by the learned, inherited metrical templates that they replicate. In the most widely accepted taxonomy of metrical types in OE verse, XX-LX compounds are hard to fit. The restrictions are detailed in Terasawa (Reference Terasawa2011:76; emphases ours):

Additionally, HX-LX compounds are problematic in Type B, where they are practically unattested (Hutcheson Reference Hutcheson1995:130–131, 218). The taxonomy of verse types is, of course, based on generalizations from observed metrical patterns; this self-confirming loop does not explain the patterns themselves. This leaves the limited availability of XX-LX NNs in the language as the initial gating mechanism leading to scarcity of types accommodating such compounds in the meter.

The statements in 5 make it clear that the avoidance of XX-LX NNs is directly linked to metrical resolution. Like alliteration, resolution presents a case of selection and placement of available linguistic material, but both alliteration and resolution are verse-design con-ventions that follow preidentified metrical templates. Similar to the way that alliteration is systematically avoided in the final ictic syllable of the b-verse, XX-LX compounds can only fit particular verse types. We agree with Terasawa (Reference Terasawa1994) and Fulk (Reference Fulk2007) that the clustering of some lexical LX items in the poetry, such as heofon-cyning ‘king of heaven’ or compounds, such as bealu ‘evil, calamity’ (see note 5) is a matter of poetic stylistic choices and formulaic diction. However, as adumbrated in section 3, accounts of stress placement in OE using resolution in verse as evidence persist in the literature. Dresher & Lahiri (1991, 2005) even borrow the metrical term resolution in their account of OE stress based on the moraic equivalence between H and LL strings.Footnote ¹² If, as Terasawa (Reference Terasawa1994) proposed and our analysis of the lexicon confirmed, the severely restricted placement of XX-LX NNs in verse results from specifically metrical/poetic preferences added to already active lexical filters, this is yet another barrier to coopting resolution as evidence for weight-based stress placement in OE. The initial syllables of the first member in both LX- and HX- NN compounds must have primary stress and fill an ictus. The initial syllables of the second member of both -LX and -HX NN compounds have secondary stress and can also be ictic.

All moraic accounts acknowledge, directly or indirectly, that OE primary stress falls on the first syllable of the stem regardless of its weight, albeit cautiously (Goering Reference Goering2016b), and with the caveat that verse evidence for stress is a “low diagnostic… suggestive, but not conclusive” (Bermúdez-Otero Reference Bermúdez-Otero2018).Footnote ¹³ Our examination of the data on NNs suggests that the link between TCs and the modeling of OE stress should be abandoned.

The statistical analysis in section 4 shows that XX-LX NNs are dispreferred in the language, which prompts two queries. First, can one identify the rationale behind the restricted formation of XX-LX NNs? Second, is there evidence that bears on the weight-to-prominence preference in the prosodic system?

As far as the first query is concerned, the dispreference for XX-LX compounds in the language—recall from section 4.3 that LX-HX is marginally preferred, while HX-LX is severely dispreferred—awaits its theoretical modeling. Though this dispreference is confirmed, there is no clear phonological reason for its existence, especially given that compounds with an LX- first member are not dispreferred. One might conjecture that primary stress is easily hosted by a syllable of any weight, but a heavier syllable is preferred for the weaker secondary stress, perhaps a compensatory reinforcement of the perceptibility of secondary stress in binominals. Such a conjecture would be in line with a well-attested difference in the treatment of identical LX nouns as phrasal heads versus as second members of NNs (Terasawa Reference Terasawa1994:10–11), which potentially mirrors the different treatment of compound stress and phrasal stress. To illustrate, in the poetic corpus, the disyllabic LX noun gryre ‘terror’, as in JDay II 8a: þurh winda gryre ‘through the winds’ terror’, Beo 384a: wið Grendles gryre ‘against Grendel’s terror’, is used as an independent ictic and alliterating item 22 times and eight times as the first member in gryre-HX compounds, but there are only two textually and editorially problematic instances of gryre as the second member in XX- compounds (Sat. 431b, 454b).

In any case, the hypothesis that weight considerations drive avoidance of a light stressed syllable in the second member of NNs faces the serious challenge of the very robust attestation of H-LX compounds both in the general lexicon and in the poetry (see note 4 and section 4 above). Note also that positing prosodic equivalence between H and LX, that is, resolution, is incompatible with the linear arrangement of H-LX versus *LX-LX and LX-HX versus *HX-LX elements in nominal compounding. Are these preferences incidental in the lexicon? Attributing the NN distributional facts to weight is, in our view, an unpromising direction of inquiry. Alternative explanations, however, surely exist.

An argument in favor of disassociating syllable weight from stress placement as well as from compound and phrasal prominence in OE does not negate the importance of weight in other parts of the phonology. OE is close to having a minimal lexical word requirement, with a caveat on pronouns’ postlexical vowel reduction. Weight definitely interacts with vocalic changes throughout the period, and its importance increases in post-OE English. Focus on prosodic weight is revealing in the accounts of segmental phonological changes, such as the much-debated High Vowel Deletion in pre- to early OE (Hogg Reference Hogg and Lahiri2000, Goering Reference Goering2016b), as well as a series of quantitative vowel changes (Bermúdez-Otero Reference Bermúdez-Otero1996, Fikkert et al. Reference Fikkert, Dresher, Lahiri, van Kemenade and Los2006, Minkova Reference Minkova2013, Goering Reference Goering2016b, Minkova and Lefkowitz Reference Minkova and Lefkowitz2020).

Further, if one decouples stress from weight, and if potentially resolvable XX-LX items in the verse mirror the lexicon, can one dispense with resolution as a metrical device? The question has been asked repeatedly, and the positions advocating either complete rejection of resolution or unconditional across-the-board application of resolution are insightfully discussed and critiqued in Suzuki Reference Suzuki1996:262–275 and contextualized within Germanic in Suzuki Reference Suzuki2014. His own position “that resolution provides a principled account for an array of missing syllable sequences as unmetrical” is the dominant view in OE metrical studies, and we are fully in agreement with it. Within the parameters of metricality established for either the Sievers/Bliss system, or the four-position principle, or the word-foot theory, resolution with all its vagaries belongs in the account of OE meter. Acknowledging the importance of metrical resolution leads to the question of how this metrical device came to be applied, albeit variably, to the OE verse tradition. One line of inquiry points to the textual history of the inherited verse corpus. Most of the surviving, usually single, copies of the poems were produced after the 10th-century monastic reforms. This suggests another, admittedly tentative, discussion of resolution within the educational and cultural milieu behind the preparation of the poetic codices.

Anglo-Latin and OE verses were composed and recorded by the same people, and their familiarity with the metrical patterns and requirements, whether through direct training in versecraft or through “reading and absorption” (Steen Reference Steen2008), is beyond doubt (see also Anderson Reference Anderson and Lanham2002, Ruff Reference Ruff2005, Thornbury Reference Thornbury2014). It is also significant that word division is onset-maximal in all widely used textbooks for the study of Latin, making light syllables easily identifiable. The original “doctrine” of <V-CV> adopted in scribal education referred to the written word: In scribendo, in scriptura are phrases used repeatedly by the grammarians in reference to this “rule”. Under rigorous instruction, replicating the Latin prescriptive pattern was not difficult. Word division at the right edge of the manuscripts shows an unprecedented 99.9% pedantic adherence to <V. CV> (Wetzel Reference Wetzel1981:117–118). The privileged position of Latin orthography and metrics in the curriculum would most likely impart authority and prestige to these matters. These additional considerations of the scribes’ educational and cultural background lend support to our proposed separation of metrical resolution and linguistic stress. Hanson & Kiparsky (Reference Hanson and Kiparsky1996:294) write on choices for settings of the metrical structure parameters:

These brief references to the cultural setting of manuscript production are intended to record our hypothesis that the already scant pool of XX-LX compounds in the language and their even narrower use in verse highlights the possibility of considering resolution an adopted and conventionalized metrical constraint.Footnote ¹⁴

6. Summary and Concluding Remarks

Avoidance of XX-LX NNs in OE verse, also known as Terasawa’s Law, has been a central point of reference in the reconstruction of OE meter and prosody. Terasawa himself (1994:62–63) interprets the paucity of second-member -LX compounds as being “less useful from the poets’ point of view due to incompatibility with the Sieversian metrical types.” However, the Sieversian taxonomy is descriptive, and its useful empirical adequacy with respect to the constraints on compound selection has not been statistically examined outside the poetic lexicon. Our study sought to establish if and how the poetic NN lexicon differs from the available pool of compounds recorded in the DOE. The statistical analysis shows that the low frequency of XX-LX compounds in verse is predictable if one looks at the entire lexicon.

While all the attested poetic quadrisyllabic NN compound choices are quantitatively within the expected range, the observed high rate of avoidance specifically of XX-LX compounds is cumulatively enhanced in the verse. The cumulativity is both metrically triggered due to incompatibility with pre-set frames, and it can be a matter of traditional formulaic poetic style, as concluded by Terasawa (Reference Terasawa1994; see also Fulk Reference Fulk1992, Russom Reference Russom and Jane Toswell1995, Hutcheson Reference Hutcheson1995, Stockwell & Minkova Reference Stockwell and Minkova1997). We believe that this combination of factors sheds more light on the findings. An approach linking the general lexicon to specific metrical and stylistic requirements is consistent with the relaxation of TCs in some of the OE religious verse texts. It is likewise significant that nonobservance of *XX-LX is a feature of some compositions in Old High German, Old Saxon, and Old Norse, which deviate from the common Germanic alliterative tradition (Terasawa Reference Terasawa1994:76–78). The observed differences in the strictness of avoidance of XX-LX NNs within the poetic corpus and across similar, but not identical, Germanic metrical traditions sub-stantiate further the inclusion of formulaic and genre-specific lexical choices in the account of compound formation and selection in OE.

The main takeaway from the research presented here is that poetic compound formation and compounding in the language share a linguistic basis, which opens up other targets of inquiry into the meter-prosody interface. The stricter observance of the *XX-LX constraint in the poetry is directly linked to the established taxonomy of verse types. That taxonomy includes resolution (LX=H) as a central criterion, often reified as prosodic evidence, though its applicability to suprasegmental pheno-mena is unmotivated and untestable outside the meter; hence the importance of resolution to the whole issue of the principles of stress assignment. Exposure to resolution through instruction in Latin versi-fication and onset-maximal syllabification as a learned scribal practice make the weight-based variability in the size of metrical lifts readily understood and applied by the poet scribes. Weight-mapping in the meter thus continues to function independently of stress, as it did in Old Norse (Ryan Reference Ryan2017). Such weight-mapping is not necessary for prosodic stress, though it can be relevant to segmental changes. If compound formation, stress assignment, and resolution are discrete but overlapping processes, this entails that the system can accommodate independently ranked constraints for stress and other phonological processes.

Detailed argumentation on the problematic nature of resolution—both as a learned metrical convention and in terms of its pertinence to the reconstruction of stress in OE—is beyond the scope of this paper, but the analysis of the lexicon presented in section 4 justifies a conclusion that compound formation and resolution are discrete processes. The metrical accommodation made available by resolution in some verse types and positions does not depend on, nor is it diagnostic of prosodic stress. More specifically, the metrical nature of resolution and its avoidance in specific positions and under secondary stress is confirmed by the unrestricted distribution of LX in syntactic phrases, where the promi-nence contour of compounds is not a factor.

Our discussion did not offer a hypothesis on the reasons for the asymmetry between LX-HX versus HX-LX compounds in the lexicon. The different rates of compounding for quadrisyllabic compounds XX-XX and trisyllabic X-LX and LX-X compounds also remain uncharted, and it may be the case that the observed discrepancy has a semantic and/or pragmatic component. The corpus can be expanded by including adjectival and adverbial compounds. Possible further directions of inquiry include data collection and analysis of actual attestation of NNs in prose and in verse, and comparison of these data to the dictionary head-word data we have collected and analyzed, but the task of manually doing so is daunting. However, as computational methods become more and more sophisticated, it becomes ever more possible to test hypotheses that previously seemed to require inordinate effort, as our paper and Neidorf et al. Reference Neidorf, Krieger, Yakubek, Chaudhuri and Dexter2019 show. Perhaps such an endeavor will be more tractable in the near future.