A Constructionist and Corpus-Based Approach to Formulas in Old English Poetry

: This paper explores a constructionist and corpus-based approach to Old English formulaic language through an analysis of the “ maþelode system” of speech introductions. The analysis is performed on a section of the York-Helsinki Parsed Corpus of Old English Poetry , comprising the poems Beowulf , Battle of Brunanburh , and Exodus . The results show that most instances of the maþelode system belong to a well-attested construction continuum, structured by the widespread Old English (and ultimately Germanic) poetic devices of variation and kenning. This continuum ranges from more fixed repetitions that exclusively involve the verb maþelian to more schematic patterns that are also attested by other SPEECH verbs, by verbs of GIVING , as well as by a number of further verbs of various semantic types. The particularly high frequency of this pattern with SPEECH verbs and verbs of GIVING matches the prominent role, highlighted by previous studies, of both word-exchange and gift-exchange within Old English heroic ideology, and suggests that these formulaic patterns served the purpose to characterize the protagonists of SPEECH or GIVING events as heroic and/or lordly figures.


Introduction
Soon after Parry ([1928Parry ([ ] 1971) ) and Lord (1960) first demonstrated the oral-formulaic character of Homeric Greek poetry, scholars such as Magoun (1953) and Kiparsky (1976) drew attention to its relevance for other poetic traditions.More precisely, Magoun (1953) first argued that the poem Beowulf was the product of oral-traditional poetics and of composition-in-performance on the basis of a detailed analysis of the poem's first 25 lines, which attest an impressive number of phrases occurring in other Old English poems as well, closely matching Parry's ([1930Parry's ([ ] 1971, p. 272) , p. 272) original definition of the Homeric formula as, "a group of words which is regularly employed under the same metrical conditions to express a given essential idea".Starting from Magoun (1953), Old English scholars have also highlighted important differences between Homeric formulas and Old English diction, the latter being based on variation rather than economy, and on flexible and abstract patterns rather than fixed repetitions.This has led to less rigid definitions of the phenomenon of formulaic diction in Old English literature.
In recent years, the study of formulas in both Homeric and Germanic traditions has received new stimulus from innovative approaches stemming from contemporary linguistic theory.More precisely, research on English idiomatic expressions has led to the formulation of Construction Grammar (Fillmore et al. 1988;Goldberg 1995), a theoretical framework that takes constructions, i.e., "learned pairings of form with semantic or discourse function" (Goldberg 2006, p. 5), as the basic units of language.By avoiding a strict division between meters, allowed variability in the number of syllables or morae.As an example, Magoun (1953, p. 450) presents the expression on gear-dagum, which is part of a formulaic system, x-dagum, used to express the idea 'long ago'.Occurring alone or with one or two preceding unstressed words, it makes up a complete C-verse.Substituting gear 'of yore' with ar, eald, or orfyrn, the formula remains unchanged in meaning and meter, while at the same time meeting the exigencies of alliteration thanks to the variation in the first element of the compound (whereas the second element, dagum, is not variated).
Drawing on Magoun's work, Diamond (1959, p. 229) importantly observed that the notion of formula may change completely from one tradition to another because of the varying requirements of meter and syntax.Thus, Diamond was the first Old English scholar to acknowledge the importance of the dimension called "tradition-dependence" by Foley (1985, p. 68), a view that was supported by later scholars both within Homeric studies (Hainsworth 1964) and in Old English studies (Whallon 1965;Fry 1967a).Considering the variability that characterizes Old English formulas, Fry (1967a, p. 204) emphasized the importance of the formulaic system and suggested a new definition of formula as, "a group of words, one half-line in length, which shows evidence of being the direct product of a formulaic system"; similarly, Niles (1983, p. 126) suggested that, "a formula in Anglo-Saxon poetry may be considered a rhythmic-syntactic-semantic complex one half-line in length".
During the 1960s, a significant development in the study of formulas involved an effort to categorize them based on their syntax rather than their semantic meaning.In the realm of Old English scholarship, O'Neil (1960) and Gattiker (1962) proposed that the syntactic structure formed the foundation of Old English poetic composition.They suggested that poets had the freedom to interchange words within the specified syntactic and metrical patterns.This theory was later applied by Cassidy (1965) and Green (1971), and in the context of Homeric studies, a similar approach was proposed by Russo (1963Russo ( , 1966)), who set out to catalogue syntactic-metrical schemas that underlie the surface realization of formulas.Although most scholars recognized that structural patterns too must have played a role in the poet's compositional technique, many did not agree that they should all be equally called formulas.
Other studies focused on the fundamental concepts conveyed by formulas rather than their formal expression.Nagler's (1967) study on formulas in Homeric Greek also influenced studies on Old English formulas (Olsen 1986, p. 567).Nagler contended that formulas should be viewed as cognitive templates with semantic content, with their surface form being merely an incidental manifestation.Similarly, Nagy (1974Nagy ( , 1976) ) and Watkins (1976) emphasized the fundamental ideas conveyed by formulas, arguing that the latter affect the meter and give rise to metrical patterns, rather than the other way around.
Despite the differences, all the studies mentioned above view formularity as a distinctive feature of poetry.The first attempt to equate formulas with bound expressions found in ordinary language is attributed to Kiparsky (1976), who also reconciled syntactic and semantic aspects of formulas.He distinguished between flexible idiomatic expressions (1)a, which are syntactically well behaved and have compositional semantics, and fixed idiomatic expressions, characterized by syntactic idiosyncrasies and non-compositional meaning (2)a.The former correspond to formulaic systems in Homeric poetry, such as (1)b, whereas the latter can be equated to straight Homeric formulas, such as (2)b: (1) a.While he recognized the true essence of the formula in "the abstract bond" between, e.g., the Ancient Greek noun álgos 'pain' and the verbal root path-'suffer ' (1976, p. 86) in Homeric poetry (1)b, Kiparsky argued that formulas consist of syntactic constituents dominated by a single node, thus recognizing the importance of the syntactic layer into the underlying structure of a formula.From his perspective, the primary distinction between fixed and flexible formulas lies in their memorization processes.Fixed formulas are memorized superficially, making them unable to undergo syntactic alterations.Conversely, flexible formulas are memorized at a deeper level, allowing for potential syntactic changes.Despite categorizing these types distinctly, Kiparsky (1976, p. 113) claimed that this is a gradient phenomenon, with fixed formulas and flexible formulas at the two poles of the continuum, and all kinds of gradations of flexibility in between.However, within the framework of Generative Grammar, he faced challenges in devising a model that could adequately account for both types of formulas.In the context of Old English studies, Fry (1981, p. 172) also noted that the most commonly accepted theory in the late 1970s assumed a generative model of memorized patterns rather than memorized phrases, which allowed enormous artistic freedom to the poetic tradition.
Recently, Bozzone (2014Bozzone ( , 2024)), Frog (2014aFrog ( , 2014bFrog ( , 2014c)), and Pagán Cánovas and Antović (2016) have further developed the idea of Lord (1960, pp. 22, 36) that formulaic language is acquired in the same way as natural language, and thus functions similarly.They proposed examining the compositional techniques of Homeric (Bozzone 2014(Bozzone , 2024;;Pagán Cánovas and Antović 2016), South Slavic (Pagán Cánovas and Antović 2016), and Old Norse (Frog 2014a(Frog , 2014b(Frog , 2014c) ) poetic traditions using Construction Grammar.The foundational premise of the latter theoretical framework (Fillmore et al. 1988;Goldberg 1995) is that "constructions", defined as recurring pairings of specific forms with specific functions, serve as the fundamental units of language.Any linguistic expression that is not entirely compositional and requires memorization falls within the notion of construction.This includes idiomatic expressions as well as syntactic structures, since the latter convey meaning, albeit abstract, even prior to being "filled" with words.Construction Grammar rejects a strict division between lexicon and syntax and represents all grammatical knowledge consistently along the syntax-lexicon continuum (Goldberg 1995, pp. 6-7).Consequently, constructions exhibit a gradient along two dimensions: from substantive to schematic (e.g., the idiom kick the bucket vs. the transitive construction) and from atomic to complex (e.g., a single word vs. a syntactic construction).Constructions form a structured inventory of a speaker's knowledge of the conventions of their language (Langacker 1987, pp. 63-76).This inventory is represented in terms of a taxonomic network of constructions, where taxonomic relations describe relationships of schematicity between two constructions (Croft and Cruse 2004, pp. 262-63).Take, for instance, (3), where the substantive idiom is an instance of the more schematic idiom The X-er, the Y-er: (3) a. [The X-er, the Y-er] | b. [The bigger they grow, the harder they fall.]Bozzone, Frog, and Pagán Cánovas and Antović offer a new perspective by recognizing that, akin to idiomatic expressions, formulas are recurring pairings of specific forms with specific functions, thereby qualifying as constructions.As shown by Bozzone (2014, pp. 40-41), this understanding of formulas as constructions encompasses various degrees of flexibility, ranging from schematic (4)a to partially substantive (4)b, to substantive (4)c, 2 and accounts for the fact that formulas, similar to other constructions, constitute a structured inventory.Finally, defining formulas as constructions acknowledges that they are stored in the poet's mind and are characterized not only by their formal attributes but also by their semantic, syntactic, and discursive functions (Bozzone 2014, pp. 42-67;cf. now Bozzone 2024, pp. 52-54).By conducting a series of case studies, Bozzone illustrated how adopting a constructionist perspective in analyzing Homeric formulaic language can open fresh avenues of inquiry and significantly enrich our linguistic, philological, and literary comprehension of oral-formulaic texts.Additionally, she proposed extending this approach to the study of other poetic traditions, particularly those less rich in substantive formulas, and characterized instead by networks of lexically flexible constructions (Bozzone 2014, pp. 39, 225).
Since 2014, several attempts have been made to apply the constructionist approach to the study of formulaic language in other Indo-European poetic traditions.In a series of studies, Frog (2014aFrog ( , 2014bFrog ( , 2014c) ) introduced an approach to Old Norse kennings as constructions.Similar to constructions of everyday language, kennings found in skaldic poetry form a continuum from schematicity to fixity.According to Frog, on the schematic end of the continuum, a kenning can be represented as the equation: NP 1 + NP 2 = NP 3 (cf.also below, Section 3), whereas on the fixed end of the continuum, we find crystalized expressions, which are reflected in the corpus as recurrent word combinations, such as geira dynr 'din of spears' for BATTLE (Frog 2014a, pp. 121, 123-24) and bryn-þing 'armor-assembly' for BATTLE (Frog 2014c, p. 293).In between, we have more or less conventionalized pairings of base-words and determiners, which tend to be representatives of conventional semantic categories; for instance, a BATTLE could be WEATHER OF WEAPONS, NOISE OF WEAPONS, NOISE OF ARMOR, etc.In the realization of such semantic associations, lexical variation is enabled by a rich vocabulary of semantically equivalent poetic terms.Thanks to this system of synonyms, in the semantic formula STORM OF SWORDS, the first element can be filled with any WEATHER-word, such as él 'snow-shower', drífa 'a fall of snow, sleet', hregg 'rainstorm', regn 'rain', etc.In the same way, the second element can be filled with other words for SWORD, such as hjo ˛rr 'sword', or even the proper name of a famous sword, such as Laufi, the name of the sword of the hero Bo ˛ðvarr bjarki (Frog 2014a, p. 106).The continuum described by Frog can be represented as follows: (5) a. NP 2 -Gen + NP 1 =NP 3 b.WEAPONS-gen + NOISE =BATTLE c.Eiríkr í dyn geira 'Eiríkr, in the din of spears (i.e., in battle)' (Hallar-Steinn Rekstefja 22.6) Following Bozzone's, Frog's, and Pagán Cánovas and Antović's proposals, Ginevra (2023, pp. 239-44) combined a constructionist approach with a historical-comparative perspective for the analysis of formulaic patterns attested in mythological texts in several ancient Indo-European languages, mainly focusing on Old Norse and Sanskrit material, but more briefly discussing texts in Ancient Greek and Old English (namely, Exeter Riddles, Boethius, and Solomon and Saturn) as well.
Another potential advancement highlighted by Bozzone (2014, p. 142) is the automation of the extraction process for constructional networks (see also Bozzone 2024, pp. 33-44, on the automatic extraction of collocational measures).In three recent papers, Biagetti (2023), Brigada Villa et al. (2023), and Biagetti (2024) adopted a quantitative approach to the study of formularity in the Rigveda, the most ancient extant text of Sanskrit literature, by exploiting morphosyntactically annotated texts from the Vedic TreeBank and semantic information from the Sanskrit WordNet.In particular, with a case study on the famous formula connected with the essential idea KILL A DRAGON (Watkins 1995), Biagetti (2023) proposed a methodology to detect lexically filled and more abstract instantiations of a formula.Brigada Villa et al. (2023) analyzed formulas including SPEECH verbs in the Rigveda, whereas Biagetti (2024) performed a co-varying collexeme analysis (Stefanowitsch and Gries 2005) of Rigvedic similes to detect recurrent, formulaic associations of parameter and standard of comparison.In a similar vein, Brigada Villa et al. (forthcoming) used the Ancient Greek Dependency Treebank enhanced with metrical annotation and semantic information from the Ancient Greek WordNet to extract formulas of COMMUNICATION and KILLING occurring in the Iliad under specific syntactic and metrical conditions.

Data Extraction 2.2.1. The York-Helsinki Parsed Corpus of Old English Poetry
The York-Helsinki Parsed Corpus of Old English Poetry (henceforth, YCOEP) is a 71,490word syntactically annotated corpus, following the constituency format of the larger Penn Parsed Corpora of Historical English (Kroch 2020).It contains a subset of Old English poetry, divided into 14 text-files, some of which contain a single text (e.g., cobeowul.psdcontains only Beowulf ), whereas others contain samples of different texts, e.g., conorthu.psdcontains Caedmon's Hymn, Bede's Death Song, The Leiden Riddle, The Dream of the Rood, and The Ruthwell Cross. 3 However, in the latter case, passages belonging to the different texts are not differentiated in their ID number, making the attribution of sentences to the right text not straightforward.Moreover, not every text has been added to the corpus in its entirety, and only some samples are present (as is the case of Christ).For this reason, we chose to concentrate on .psdfiles containing one single text in its entirety, namely Beowulf, The Battle of Brunanburh, and Exodus.
Despite the presence of syntactic annotation in the YCOEP, it could not be used to extract data.The constructions needed could not be simply retrieved looking for -PRN tags, i.e., marking appositive or parenthetical phrases (which are essential to the formulas analyzed in the present study): even though -PRN tags are linked to the noun they refer to through (XP *ICH*-n) tags, i.e., non wh-trace (scrambling and extraposition), the latter is not necessarily dependent on the verb, which is the main focus of this investigation, as verbs are the pivot of the constructions analyzed in Section 3.This is the main reason why we resorted to a CoNLL-U conversion of the YCOEP treebank, as explained in Section 2.2.2.

Conversion and Extraction
While for detecting and analyzing formulaic expressions in the Homeric poems and the Rigveda the researchers were able to leverage dependency treebanks for Ancient Greek and Vedic Sanskrit, a dependency treebank for Old English is still missing.Thus, before extracting the patterns of potential formulaic expressions, we had to convert the YCOEP constituency treebank into CoNLL-U format, an adaptation of the CoNLL-X format (Buchholz and Marsi 2006).To do so, we adopted the same methodology described by Brigada Villa and Giarda (2023), isolating the relevant tokens from the tags and parenthesis used to annotate the treebank, and converting each tag into a combination of parts-of-speech and morphological features, compliant with the guidelines of Universal Dependencies (UD; De Marneffe et al. 2021).The advantages of having the annotation structured in this format are manifold.Firstly, the CoNLL-U format facilitates a more user-friendly qualitative analysis by placing the text of the sentences above the annotation, as opposed to the original format, which does not report the entire sentence separated from the annotation.Furthermore, the adopted annotation scheme enables the retrieval of syntactic patterns that might underlie more schematic formulas.
Unfortunately, exploiting syntax was not possible at this stage: the conversion of the YCOEP only involved the parts-of-speech tags and the morphological features.There have been attempts to add the syntax to Old English syntactic resources formatted as the YCOEP, using a multilingual parser (Brigada Villa and Giarda 2023) or a rule-based approach (Brigada Villa and Giarda 2024), but the results of the automatic parsing were not satisfying enough to be employed in our analysis and the rule-based approach involved only the syntactic root.It is worth mentioning that these attempts were made on Old English prose, and the poetic nature of our texts would have had a negative impact on the models used on the prose.Hopefully, in the future, a conversion into the CoNLL-U format adhering to the UD guidelines, adapted for texts in Old English, as suggested by Martín Arista (2022), will be available to the community of scholars, and studies on Old English formularity will benefit from this.
We present now the queries we designed to extract the potential formulas from the sections of the YCOEP that make up our corpus.As shown in Figure 1, the only fields available to be queried were the ID of a token, the word form, the UD part-of-speech tag, the part-of-speech tag employed in the original YCOEP, and the morphological features.lish prose, and the poetic nature of our texts would have had a negative impact on the models used on the prose.Hopefully, in the future, a conversion into the CoNLL-U format adhering to the UD guidelines, adapted for texts in Old English, as suggested by Martín Arista (2022), will be available to the community of scholars, and studies on Old English formularity will benefit from this.
We present now the queries we designed to extract the potential formulas from the sections of the YCOEP that make up our corpus.As shown in Figure 1, the only fields available to be queried were the ID of a token, the word form, the UD part-of-speech tag, the part-of-speech tag employed in the original YCOEP, and the morphological features.We decided to structure our query on the basis of the main pattern attested by the majority of instances of the formulaic "maþelode system", as explained below in Section 3. Since we lacked syntactic annotation, we decided to focus on sequences of tokens: our approach started from the extraction of the patterns at whose center was a verb, looking at the five tokens that preceded and followed this central token.We filtered the results, adding some restrictions on the elements that had to appear before and after the verb for the pattern to be considered valid.To be considered for the analysis, a pattern had to include: • a verb (tagged as VERB) or an auxiliary (tagged as AUX), • an element X, being a noun or a pronoun (i.e., having the UD part-of-speech tag NOUN, PROPN, or PRON), preceding the verb in any position, We decided to structure our query on the basis of the main pattern attested by the majority of instances of the formulaic "maþelode system", as explained below in Section 3. Since we lacked syntactic annotation, we decided to focus on sequences of tokens: our approach started from the extraction of the patterns at whose center was a verb, looking at the five tokens that preceded and followed this central token.We filtered the results, adding some restrictions on the elements that had to appear before and after the verb for the pattern to be considered valid.To be considered for the analysis, a pattern had to include: • a verb (tagged as VERB) or an auxiliary (tagged as AUX), • an element X, being a noun or a pronoun (i.e., having the UD part-of-speech tag NOUN, PROPN, or PRON), preceding the verb in any position, • an element Y in the same case of the element X and being a noun or a proper noun (i.e., having the UD part-of-speech tag NOUN or PROPN), following the verb in any position, • an element Z, modifier of Y, i.e., a noun in the genitive case, following the verb and contiguous to Y.
In addition to that, in order to facilitate the manual analysis of the extracted patterns, we also extracted the order of the elements X, Y, and Z, and the order of their parts-ofspeech, as they appeared in the patterns.This extraction yielded 336 passages, which were manually scrutinized, resulting in the identification of 108 valid occurrences of the pattern.

Results
We tested our methodology by analyzing the formulaic "maþelode system" (named after the preterite 3sg of the Old English verb maþelian 'speak, discourse, make a speech') of speech introductions in Beowulf (Cook 1926;Creed 1957, p. 527;Mertens-Fonck 1978;Lord 1991, pp. 147-69;cf. Pàroli 1975, p. 182) and by looking for lexical, syntactic, and semantic parallels for this system in our corpus.We focused on the main pattern [PROPN X maþelode [NOUN PROPN.Gen or PROPN.Gen NOUN] X ], attested by the majority of the system's occurrences (15× of 26), i.e., by the following passages (in what follows, translations of Beowulf and Exodus are adapted from (Fulk 2010) and (Hostetter 2024), respectively):

"Argument Variated by Kenning" Construction
The maþelode system's main pattern instantiates (among others) an "Argument Variated by Kenning" construction, which in turn combines two constructions, both of which reflect poetic devices that are very well attested in both the Old English and other early Germanic traditions (henceforth, each example will be preceded by the pattern it instantiates, as it was extracted by our automatic query): • The construction [NOUN/PRON/PROPN X (. ..) [. ..]X ], by which a noun (or pronoun or proper name) evoking a concept X is followed (immediately or not) by a word or a phrase that appositionally refers to the same concept, X.For instance, in (21), the word winreced 'winehouse' is immediately followed by the phrase goldsele gumena 'golden hall of men', and both items refer to the same concept, HALL.This pattern reflects a broader (not restricted to nouns) and well-known poetic device of early Germanic poetics, widely attested in, e.g., Old English (Campbell 1962;Robinson 1985) and Old Norse (McTurk 1997) poetic texts, and variously referred to as either "parallelism" (e.g., Campbell 1962), "appositive style" (Robinson 1985), or "variation" (e.g., Paetzel 1913;McTurk 1997).The latter term shall be used in the remainder of this paper.
[winreced.NOUN.Acc X [goldsele.NOUN.Acc gumena.NOUN.Gen] X ] (21) [. ..] to þaes þe he winreced / goldsele gumena gearwost wisse [. . .] 'to where he well knew the winehouse was, the golden hall of men' (Beo.714b-715) • The construction [NOUN NOUN/PROPN.Gen] X or [NOUN/PROPN.Gen NOUN] X , by which a single concept, X, which may usually be expressed with a single noun (or proper name), is instead evoked by means of a bipartite noun phrase, consisting of a noun that is modified by another noun (or proper name) in the genitive case.
For instance, in ( 22), the noun phrase sinces brytta 'disperser of valuables' is used to refer to the concept LORD, in place of a simple noun, such as dryhten 'lord'.This pattern reflects the famous poetic device called "kenning" (from Old Norse kenning 'knowledge, recognition'), a widespread feature of early Germanic poetic traditions, such as the Old English (see, e.g., Marquardt 1938) and the Old Norse (Meissner 1921; see also above, Section 2.1) ones, but also attested in several other Indo-European languages (e.g., Krause [1930Krause [ ] 2014, p. 578ff;, p. 578ff;Waern 1951;Campanile 1977, p. 108ff;Watkins 1995, p. 44ff).In what follows, we use the term "kenning" according to Watkins' (1995, p. 44) broad definition as, "a bipartite figure of two nouns in a noncopulative, typically genitival grammatical relation (A of B) or in composition (B-A) which together make reference to, 'signify', a third notion C". 4   [sinces.NOUN.Gen brytta.NOUN.Nom] LORD (22) [. ..] þaet ðu geare cunne, sinces brytta [. . .] "so that you will know for certain, disperser of valuables [i.e., lord]" (Beo. 2070b(Beo. -2071a) ) In the maþelode system's main pattern, these two constructions are combined with a verb's argument-structure construction [X V (X.2) (X.3)], comprising a verb and its argument(s), X (and eventually X.2 and X.3).The resulting complex pattern may be represented as follows, with two possible word orders for the modifier in the genitive case (for the sake of clarity, in what follows, we will always use the first variant as the default order): 5 According to this generic pattern, which we may refer to as the "Argument Variated by Kenning" construction, an argument (a noun or pronoun or proper name) evoking a concept, X, is followed (immediately or not) by the verb it belongs to, which is then followed (immediately or not) by a kenning that appositionally refers to the same concept, X.Our query captured 108 valid occurrences of this pattern in our corpus, many of which are not directly relevant to our investigation, such as those with the variated element in the dative (23) or genitive case (24): 'a mortal wound he earned by lot, sword-strokes,' (Beo. 2385b(Beo. -2386a(Beo. , cobeowul,74.2384(Beo. .1940) ) [lissa.NOUN.Gen bidde.VERB sigora.NOUN.Gen gesynto.NOUN.Gen] (24) [. ..] and eow liffrean lissa bidde, sigora gesynto, [. . .] 'I ask for you the grace of the Lord of Life, the success of victory' coexodus,98.269.221)With the sole exception of (25), where the nominative brego is used absolutely in its vocative function, in all other occurrences that involve a verb of SPEECH (19×) or GIVING (7×) and a variated element in the nominative, the latter is always the subject of the verb and always refers to an ANIMATE entity (either a HUMAN being or the Christian God).We may thus represent this "ANIMATE Subject Variated by Kenning" construction of SPEECH and GIVING verbs, attested 26 times in our corpus, as follows: The two patterns discussed in the previous section and in the present one underlie all fifteen occurrences of the maþelode system's main pattern.Each of the following two sections instead discuss a sub-pattern that is attested only by some instances of the latter.

"HUMAN Subject Variated by CHILD-Kenning" Construction
Within the great majority (12×) of occurrences of the maþelode system's main pattern, the HUMAN subject X of the verb is variated by means of a patronymic CHILD-kenning construction [NOUN/PROPN.Gen CHILD.NOUN] X , i.e., a noun phrase with a noun or a proper name in the genitive case that modifies a noun evoking the concept CHILD, namely, bearn 'child' (one time with Unferð and nine times with variants of Beowulf ), as in ( 39)-( 40 While the same pattern is not attested with other SPEECH verbs in our corpus, it occurs four times with GIVING verbs, namely, in the same passages already listed above as ( 32)-( 33) and ( 35)- (36) and repeated below as ( 42)-( 45).In these occurrences, the verb's subject is either a noun, a pronoun, or a proper name, and the nouns used to evoke the meaning CHILD are sunu 'son', as in (42), dohtor 'daughter', as in (43), and maga/mago 'son', as in ( 44)-( 45 (Beo. 2173b(Beo. -2174a(Beo. , cobeowul,67.2172(Beo. .1774) ) This pattern is attested eight times with SPEECH verbs in our corpus.Again, the same pattern is also attested three times with GIVING verbs-two times with the same noun waldend 'ruler, lord' that is also attested with the SPEECH verb swerian 'swear' in (55): this correspondence corroborates the impression of similarity in the behaviors of SPEECH and GIVING verbs, already noted above in the previous sections and discussed in more detail below (in Section 4): On the other hand, a similar pattern is also attested eight times in our corpus, with the variated element either in the nominative case in vocative function (once), or in other cases, namely, the accusative (five times) and the dative (two times).Both the vocativenominative and the accusative patterns are attested with SPEECH verbs, namely, biddan 'ask, request' in (25) above, repeated below in (63), and frinan 'ask, inquire' in (26) above, repeated below in (64).The accusative and dative patterns, however, are also attested with several further verbs, such as witan 'know' in (65) and herian 'praise, honor' in (66).This allows for the identification of an even more generic "ANIMATE Argument Variated by LORD-Kenning" construction, attested 31 times in our corpus and underlying all the examples discussed in this section, by which an ANIMATE argument (in any case) precedes its corresponding verb, which in turn precedes a LORD-kenning referring to the same ANIMATE:

Discussion
Scholars have long noted the peculiar features of the maþelode system of speech introductions (see, e.g., McConchie 2000, p. 59), even going as far as suggesting that some peculiarities of this system may stem from the influence of a foreign tradition (the Homeric one, according to Cook 1926, p. 6).Our results, however, show that the maþelode system's main pattern rather belongs to a very well-attested construction continuum, which is deeply structured by the widespread Old English (and ultimately Germanic) poetic devices of variation and kenning.This continuum ranges from more fixed repetitions that exclusively involve the verb maþelian to more schematic patterns that are also attested by other SPEECH verbs, by verbs of GIVING, as well as by a number of further verbs of various semantic types.The occurrences (tokens) of this continuum may be represented as in Table 1.All the passages that we were able to identify in the corpus by means of our basic query are instances of the broader "Argument Variated by Kenning" construction, i.e., [NOUN/PRON/PROPN X VERB [NOUN/PROPN.Gen NOUN] X ].As explained above (in Section 2.2), our query allowed us to retrieve 108 valid instances of this pattern with any verb (i.e., any word tagged as either VERB or AUX in the corpus).As discussed in Section 3, eight of these instances involve verbs of GIVING, whereas 26 involve SPEECH verbs, obviously including the fifteen instances of the maþelode system's main pattern.In Table 1 and in what follows, for each of these four classes (maþelode, SPEECH verb, GIVING verb, and any verb), we detail the relative frequencies of three further sub-patterns discussed in Section 3: "ANIMATE Subject Variated by Kenning", its sub-type "HUMAN Subject Variated by CHILD-Kenning", and "ANIMATE Argument Variated by LORD-Kenning".
First, the fifteen occurrences of the maþelode system's main pattern are all structured by the more specific "ANIMATE Subject Variated by Kenning" sub-construction, namely, [ANIMATE.NOUN/PRON/PROPN.Nom X VERB [NOUN/PROPN.Gen.NOUN.Nom] X ].The latter pattern is by no means limited to the maþelode system, however, as it is also extremely frequent not only among SPEECH verbs in general (73.07%), but also among verbs of GIVING (87.50%), and it is actually very well-attested by verbs in general, accounting for more than half of all results of our basic query (53.70%).
An investigation of the two further sub-patterns that underlie the system produced similar results.The "HUMAN Subject Variated by CHILD-Kenning" construction (a sub-type of "ANIMATE Subject Variated by Kenning"), i.e., [HUMAN.NOUN/PRON/PROPN.Nom X (SPEECH/GIVING.)VERB[NOUN/PROPN.Gen.CHILD.NOUN.Nom] X ], accounts for most of the maþelode occurrences (80%), which also make up all twelve instances of this subpattern with SPEECH verbs (46.15%).Again, this sub-pattern is also among the most frequent ones for both verbs of GIVING (50%) and verbs in general (25%), and its frequency is perhaps best explained by the suggestion that it reflects ancient (West-)Germanic traditional poetic heritage (Mertens-Fonck 1978, pp. 435-36): as long noted (cf.e.g., Cook 1926, p. 3), passages such as (41), Wiglaf maðelode, Weohstanes sunu 'Wiglaf made a speech, Wihstan's son' (Beo.1862), find quite close parallels in texts in other early Germanic traditions, such as the Old High German epic poem Hildebrandslied (e.g., line 7: Hiltibrant gimahalta Heribrantes sunu 'Hildebrand said, Heribrand's son').As noted by Lord (1991, p. 168) in his comparative analysis of maðelode formulas in Beowulf and Elene, such highly formulaic expressions seem to be reserved for traditional figures, such as Beowulf and Hroðgar, who likely belonged to the Old English heroic tradition, whereas they seem to be lacking for the biblical characters of the poem Elene, who were not clearly identifiable by their patronymics or epithets, because they were not part of the same inherited tradition.The same may also be true for the "ANIMATE Argument Variated by LORD-Kenning" sub-construction, namely, [ANIMATE.NOUN/PRON/PROP X VERB [NOUN/PROPN.Gen LORD.NOUN] X ], which only accounts for 20% of the occurrences of the maþelode main pattern, relatively close to the percentage of the same sub-pattern for verbs in general (28.70%).This pattern's frequency, however, increases considerably if one takes into account SPEECH verbs in general (34.61%), and is even higher for verbs of GIVING (37.5%).
Thus, on the one hand, our data suggest that the formulaic constructions of the maþelode system's main pattern result from the combination of several more schematic constructions, which in our corpus often occur with other verbs of various semantic types as well.On the other hand, however, they also reveal a very clear correspondence between the behaviors of SPEECH verbs, such as maþelian, and of verbs of GIVING, both of which occur disproportionately often (SPEECH 73.07%; GIVING 87.50%), compared to verbs in general (53.70%), within the "ANIMATE Subject Variated by Kenning" sub-construction.In principle, this correspondence may be explained by their underlying semantic structure, as both SPEECH and GIVING verbs are usually three-place verbs, i.e., their meaning implies three main semantic arguments (speaker, receiver, and message for SPEECH; donor, recipient, and theme for GIVING), two of which (speaker and receiver for SPEECH, donor and recipient for GIVING) are most often ANIMATE, with the subject being almost always ANIMATE.
Very similar frequencies for these two semantic classes, however, are also attested with respect to the more specific "HUMAN Subject Variated by CHILD-Kenning" construction (SPEECH 46.15%; GIVING 50%), frequencies that are disproportionately higher than those of the same sub-pattern for verbs in general (25%).The same is also true, although to a lesser extent, for the "ANIMATE Argument Variated by LORD-Kenning" pattern (SPEECH 34.61%; GIVING 37.50%; all verbs 28.70%).In order to explain why SPEECH and GIVING verbs behave similarly and occur so often within these constructions in our corpus, it may be relevant to take into account the heroic ideology that informs Old English epic texts, such as Beowulf.
As is well known, one of the main tasks of lords and chieftains in the Old English tradition was the distribution of wealth to their subjects and companions, as exemplified by the kenning sinces brytta 'disperser of valuables' for LORD in ( 22) above.The importance of the custom of gift exchange within the heroic society depicted in Beowulf has been extensively discussed by scholars, such as Donahue (1975, especially pp. 24-29).As argued by Bjork (1994, p. 998), however, just like the exchange of material objects, the exchange of words was also part of this heroic ideology, and in Beowulf, speech may in fact be seen as "a link in the chain of the gift".In any case, the importance of "word-exchange" among the skills of heroic characters is evident from several passages of Beowulf itself (Bjork 1994, pp. 996-98), allowing for the assumption that SPEECH, just like GIVING, was regarded as another prerogative of heroic and/or lordly characters in Old English poetry.This may explain why SPEECH and GIVING verbs display remarkably similar frequencies for the "HUMAN Subject Variated by CHILD-Kenning" construction and the "ANIMATE Argument Variated by LORD-Kenning" construction, which served the purpose to characterize the protagonist of the SPEECH or GIVING event as a heroic figure of renowned genealogy and as the chief of a community, respectively.
In the present contribution, we hope to have shown the potential of a constructionist and corpus-based approach to Old English formulas, such as the maþelode system of speech introductions.Our investigation should obviously be considered as a pilot study, as it relies on a relatively limited corpus, and it exclusively focuses on the main pattern that underlies most instances of the maþelode system.The analysis we carried out may certainly be improved with the aid of a larger corpus and more data in general.For instance, a more refined investigation will need to also take into account information on the metrical structure and the syntactic context of each occurrence.As for the former, metrical annotation would allow the extraction of more refined patterns from our corpus; as for the latter, since our results are exclusively limited to the occurrences that could be extracted on the basis of morphological annotation (part-of-speech class and case), syntactic annotation would allow us to explore further patterns attested by the maþelode system, e.g., those in which the verb (not the noun) is variated (Mertens-Fonck 1978, p. 435).It would also be possible to extend our analysis to larger patterns, e.g., to compare passages in which the maþelode line is part of a larger group of lines vs. those in which it is not (Mertens-Fonck 1978, pp. 436-37).Future research in the directions proposed in this study may also focus on other patterns attested by both maþelian and other SPEECH verbs in Old English, in order to see if differences in the constructions/patterns in which each verb occurs may parallel the differences in their broader usage (on which, see, e.g., Rissanen 1998, pp. 163-64).Finally, it would be interesting to investigate whether SPEECH and GIVING verbs display similarities in other contexts as well.

Figure 1 .
Figure 1.A sentence from the Beowulf section of our corpus formatted in CoNLL-U.

Figure 1 .
Figure 1.A sentence from the Beowulf section of our corpus formatted in CoNLL-U.

Table 1 .
The construction continuum of the maþelode system's main pattern.