Dimensions of constructional meanings in the German Constructicon: Why collo-profiles matter

1 Constructicography and the meanings of constructions

Constructicography is still a very young linguistic subdiscipline which is dedicated to the systematic documentation and ‘archiving’ of lexical as well as grammatical constructions (cf. Boas, Lyngfelt and Torrent 2019; Lyngfelt 2018; Ziem and Flick 2019). Following a constructionist framework, constructicography ultimately aims at developing (digital) ‘dictionaries’ for constructions. Such reference works can serve a variety of purposes; they may, for example, serve as a linguistic documentation of a target language at a specific point in time, or as a rich resource for linguistic studies, including computational endeavors such as building construction parsers, or as pedagogical tools for foreign language teaching and learning.

Constructicography builds on the long traditions of lexicography and grammaticography, both of which have a great deal of (theoretical and practical-empirical) experience and bring with them a wealth of established concepts and tools (cf. Fuß and Wöllstein 2018, Mosel 2006, Osswald 2015). Just like lexicography, constructicography relies on corpus data which provide insights into the linguistic system peculiar to a variety of a speech community. At the same time, it differs from lexicography and grammaticography in that it is guided by the assumption that both lexicon and grammar form poles of a continuum (for an overview cf. Boas 2010, Broccias 2012, Ziem 2022). In this view, grammar and lexicon are not supposed to be different in nature. Instead, grammatical and lexical units are similar in kind and should thus be analyzed within the same framework (cf. in detail Ziem 2022).

Importantly, constructicographers are specifically interested in constructions that fall between lexicon and grammar: semi-schematic constructions, that is, constructions with both open and lexically specified slots (cf. for Russian: Janda et al. 2020, for Swedish: Bäckström et al. 2013, for German Boas and Ziem 2018). For illustration, consider the N₁ an N₁ construction in (1) which we will discuss in more detail later.

This construction includes a lexically specified slot (an ‘to’) and two open slots. Its function is to quantify the entity addressed in the noun slot (‘bumpers’) and it also expresses continuity (‘bumper to bumper to bumper…’).

Hitherto, neither grammars nor dictionaries feel responsible for such semi-schematic constructions, including constructional idioms.^[1] Usually, they are too schematic for a dictionary and too word-like for a grammar. Consequently, semi-schematic constructions fall through the cracks of a lexicon-and-grammar approach. They cannot be integrated into a lexicon; at the same time, they share only some properties with the grammatical units addressed in grammars. Although semi-schematic constructions are commonly considered ‘irregular’, their slots are subject to locally defined regularities, in particular constraints, which determine in a construction-specific way which semantic, pragmatic or formal conditions slot-fillers have to meet (cf., for instance, Fillmore, Kay and O’Connor’s (1988) seminal paper on pragmatic constraints of the let alone construction). Despite these peculiarities, semi-schematic constructions belong to the language system just as fully regular expressions and grammatical structures. This is all the more true since they occur in a large variety.

For any constructicographic enterprise it is of pivotal importance to document semi-schematic constructions. But how to capture their meanings? Is it possible to adapt lexicographic methods to also account for constructional meanings? If this is not the case, or only to a limited extent, is there a need to develop new methods to account for meanings of semi-schematic and fully schematic constructions? More generally, which dimensions of meaning could, and should, be addressed in a constructicon?

In this article, we try to provide at least preliminary answers to these questions. Relating to categories implemented in the German FrameNet Constructicon, we first give an overview of the broad range of semantic dimensions included in each construction entry. We argue that due to the conceptual complexity of semi-schematic constructions we need go beyond definition-like meaning description as usually provided in lexicon entries. Instead, it is necessary to offer various points of access to constructional meanings, ranging from semantic specifications of a construction’s slots to information about the construction family of which the target construction is a member and about so-called constructional collo-profiles specific to the construction addressed. In the second part of our paper, we solely focus on the role of collo-profiles as introduced by Herbst (2018, 2020) and Herbst and Hoffmann (2018), showing that collo-profiles are not only conclusive for determining constructional meanings but also helpful for explicating usage preferences, and thus serve as a rich semantic resource. As a consequence, we consider collo-profiles an indispensable component of construction entries. However, in the context of constructicography it is not feasible to conduct collostructional analyses (CA) for hundreds or even thousands of constructions. Being aware of this limitation, we introduce a procedure based on the large language model (LLM) BERT that allows to predict collo-profiles without having to extensively annotate constructs, i.e. instances of constructions, in a given corpus.

2 Constructicographic approaches to constructional meaning in the German Constructicon

In the last decade, there has been a growing interest in developing and implementing rich language resources based on constructionist assumptions (Lyngfelt et al. 2018). In the spirit of construction grammar as well as (computational) lexicography, the German FrameNet Constructicon project (cf. www.german-constructicon.de), among others, specifically tries to explore and document the continuum between lexicon and grammar on a large scale.

From a constructicographic point of view, a constructicon – a morphological blend of construction and lexicon coined by Jurafsky (1992: 302) – is the result of examining and documenting both the full range of lexical and grammatical constructions of a target language (including meaning-bearing units of all kinds) and the ways in which they are interconnected and form a complex network of relations. It is the entire network of constructions that makes up the constructicon of a language. However, the term constructicon is polysemous as it does not only refer to a lexicon-like resource of constructions but also to the linguistic knowledge of speakers in a language community. With respect to such a “mental constructicon”, Goldberg (2003: 219) goes so far as to claim that the totality of our knowledge of language must be captured by a network of constructions. In the remainder, we will use constructicon exclusively in the former sense.

Constructicographic analyses of constructions are very demanding and labor-intensive. Since, at least in the case of German as a target language, partially schematic constructions have so far received little attention in grammaticography (Dobrovol’skij 2011, Ziem 2018, Finkbeiner 2018), we restricted ourselves to this type of construction in the first phase of the project. The guiding principle was Fillmore’s (1988: 36) claim that it is precisely the ‘irregular’, ‘idiosyncratic’, and supposedly ‘marginal’ constructions that must be used to measure the viability of an approach. In this area between lexicon and grammar, we find many hundreds of idiosyncratic, semi-schematic, and (partially) idiomatic constructions. They include, for instance, grammatical phrasemes (let alone, for…sake), phrasal idioms (cf. Ziem 2018, 2020), support verb constructions (making a request, making an assertion, etc.) as well as a multitude of other units, such as headless constructions (e.g., Was mich ärgert, sind solche Sätze ‘What annoys me are such sentences’), prepositional phrases with bare nouns (mit Sorge ‘with concern’, in Trauer ‘in sorrow’), and many more. The range of constructional idioms alone includes examples as diverse as those documented in (2).

Generally, semi-schematic constructions with varying productivity are particularly interesting and methodologically challenging since they do not only occur with great variety but also exhibit properties that are usually described lexicographically, while other properties, in particular those addressing construction elements (slots) coming with specific constraints, require special attention for a comprehensive description of a construction’s function. In order to capture functions of constructions as comprehensively as possible, each entry contains a range of semantic information, including the following.

(a) Meaning of construction elements (slots). Constructicographic investigations start with a careful analysis of the slots of the target construction (construction elements). To illustrate, consider for example the N₁ an N₁ construction as illustrated in (1). This construction is both in formal and functional terms a daughter of the more abstract prepositional reduplication construction. At the same time, it is a sister of the N₁ über N₁ construction as exemplified in (2b), and it also relates to the non-schematic construction Seite an Seite, cf. (2a). Despite its formal similiarities to this construction, it differs from it in that its construction elements are open for a wealth of fillers, namely any kind of count nouns; this is because (2a) is idiomatic even though there are functionally related non-idiomatic instances attested in corpus data. (3a) illustrates the idiomatic, (3b) the non-idiomatic one.

(1) also differs from (3a-b) in that its nominal fillers may be modified to some extent (e.g., blue bumpers to blue bumpers). This is because Seite an Seite is a non-schematic phraseological unit. Thus, it comes with no surprise that it should get its own entry in a constructicon, just like the semi-schematic N₁ an N₁ construction. The meanings of both constructions do not match, but they do overlap: While both units code spatial contiguity (proximity), only the semi-schematic construction codes continuity and thus quantifies the nominal referent. In addition, there is also only partial overlap between N₁ an N₁ and other preposition reduplication constructions such as (2b). Both fulfill a quantifying function, and in both cases the slots of the constructions only license count nouns; however, only (3b) codes continuity. Hence, it is this semantic dimension that makes the N₁ an N₁ construction functionally unique. Such a slot analysis is an essential starting point of any constructicographic meaning description.

(b) Relations among constructions: constructional networks. Importantly, it is not possible to get a full picture of a construction’s function by only looking at its ‘internal’ properties. Even though analyses of construction elements also rely on comparing constructions, they do not take the habitat of the target construction into account. For this, it is necessary to consider the relations holding between a target construction and its neighbors. A close glance at the direct neighbors promises new insights into the peculiarities of the target construction in several respects. To exemplify, our target construction N₁ an N₁ has one mother and three sisters, some of which have sisters again; Table 1 provides an overview.

Each construction shares both basic form- and meaning-related characteristics: formally, they all instantiate one and the same reduplication schema provided by the fully schematic N1_PREP_N1 construction serving as the mother construction; functionally, they all express continuity combined with some sort of quantification. Such local network structures are visualized by the ConstructionGrapher (for an overview cf. Ziem and Willich in press). To illustrate, Figure 1 shows the habitat of the N₁ an N₁ construction (for the sake of better visualization restricted to two edges here).

Figure 1:

The N₁ an N₁_continuity construction in its local habitat (cf. https://gsw.phil.hhu.de/constructicon/constructiondata?id=349)

(c) Family structures. The habitat of a construction, i.e., its local constructional network, is not only structured by (different types of) relations holding among constructions. On top of that, they are organized in families (cf. Ziem and Willich, 2023). Such family structures provide additional clues to constructional meanings. In total, there are three types of families, each providing new insights into the nature of the target construction: (a) Construction families address constructions that share essential commonalities both in terms of form and meaning; in the case of N₁ an N₁, the constructions summarized in Table 1 form members of this family. (b) Form families relate to structural commonalities among constructions. For instance, the N₁ an N₁ construction belongs to the large family of reduplication constructions comprising a variety of phenomena, ranging from phraseological units, such as x hin x her (‘x or no x’) to intensifying reduplication constructions, such as Adj_Adj_NP.^[2] Finally, (c) meaning families gather constructions with strong semantic similarities regardless of their form gestalts. Constructions belonging to the same meaning family evoke the same frame or at least closely related frames. Specifically, in the case of the N₁ an N₁ construction it is the frame Gradierbare_Nähe (‘Gradable_proximity’) that shapes the meaning of the construction under investigation. Interestingly, the local network structure provided by the FrameGrapher opens up a broader, more coarse-grained view on the meaning of the N₁ an N₁ construction in that the evoked frame is part of a large group of frames coding spatial relations between entities, including the image schema frame Nähe (‘proximity’) as well as the frame Lokative_Beziehung (‘local relation’) which is essential for any frame in the large space-coding frame family.^[3] To conclude, family structures allow us to take a different perspective on properties specific to the target construction in general and, more specifically, on constructional meaning. At a higher level of granularity, they capture local network structures in which constructions are organized.

(d) Frame-approach to constructional meaning. As mentioned above, constructional meaning is determined by the frame a construction evokes. One benefit for including frames is to get an accurate sense of what other meaning-bearing entities in the lexicon-grammar continuum (ranging from lexical to phraseological and grammatical units) have in common with the target construction. At the same time, looking at other units evoking the same frame also reveals what distinguishes the construction from others. To this end, mappings between construction elements and elements of the evoked frame are instructive. Figure 2 illustrates these relations.

Since a frame entertains relations to other frames (and may be a member of one or more frame families), every frame evoked by a construction provides access to rich semantics, reaching far beyond the meaning provided by the frame itself. Consequently, taking the evoked frame and its habitat into consideration makes it possible to take a very sophisticated perspective on constructional meaning.

Figure 2:

Mappings between construction elements and frame elements (cf. https://gsw.phil.hhu.de/constructicon/constructiondata?id=349)

(e) Collo-profiles. Before constructicography emerged about half a decade ago, the most influential method to analyze constructional meaning relates to so-called collostructions (a blend of collocation and construction). In one variety, collostructional analysis (CA) measures the association strength between construction elements (slots), or the entire construction to which they belong, and instances entering the construction elements (Stefanowitsch and Gries 2003). With such a focus, CA combines John R. Firth’s (1951/1957) notion of collocation and Goldberg’s (2003) notion of construction (also adopted here). Crucially, collostructions help to identify collo-profiles which in turn provide insights into paradigmatic properties of construction elements by specifying the strength with which a construction element attracts linguistic units (Herbst 2018, 2020). Thus, collo-profiles are powerful clues to a better understanding of constructional meaning: if a collo-profile changes, the construction’s meaning changes.

In the next section, we stick to this dimension of constructional meaning and show in detail how to determine the collo-profile for the schematic N₁ PREP N_{1 (continuity)} construction with the help of the Large Language Model BERT.

3 Integrating collo-profiles in the German Constructicon

4 Why collo-profiles matter for constructicography: conclusions and outlook

Documenting the meaning of a construction is one of the key tasks of constructicography – and it is a specifically challenging task since it is not sufficient to solely rely on lexicographic procedures. Lexicographic analyses fall short not least because fully and partially schematic constructions are semantically much more complex than lexical constructions. Therefore, the schematicity of constructions has to be taken into account at the outset. What is the function of the construction elements (slots) of a target construction? What do they contribute to its meaning? What is the relationship between the construction elements? Which instances typically occur in the slots, and to what extent are they informative for the construction’s meaning? How is the target construction related to its neighbors, and how do their construction elements relate to one another? It goes without saying that the dimensions of constructional meanings addressed are manifold – nonetheless, constructicon entries should attempt to cover them all.

With such a goal in mind, the present paper has introduced various dimensions of constructional meanings with a specific focus on collo-profiles. Collo-profiles are considered a semantically prominent dimension of constructional meaning in that they not only provide information about usage preferences of constructions but also serve as a helpful indicator for semantic peculiarities of constructions. However, given that a constructicon does not consist of dozens but of many hundreds and even thousands of constructions, collostructional analyses as proposed by Stefanowitsch and Gries are simply not feasible on such a large scale. Drawing on the neural-network-based LLM BERT, the present article has introduced a way to cope with this challenge.

Specifically, we have developed an automated system that generates collo-profiles based on a limited number of annotated instances. To this end, we use an adapted version of Wada et al.’s approach and predict lexical substitutes for the original fillers of the example sentences, which we then filter and count. For illustration, we used the N₁ PREP N₁ (‘N₁ PREP N₁’) and the x macht y ADJ (‘x makes y ADJ’) construction. The results achieved show considerable agreement with a CA on the same data. However, as expected, there are also shortcomings; one of them concerns the handling of predicted inappropriate collexemes.

So far, our method only provides equivalent functionality to one variant of CA, namely the simple collexeme analysis. It would be very promising to extend the framework presented here and also incorporate procedures mimicking distinctive collexeme analysis and covarying collexeme analysis.

While the procedure for predicting collo-profiles has already been implemented in the German Constructicon, there are still some issues we need to address in the near future. One of them relates to developing new methods for identifying as many new constructions as possible and for subsequently creating collo-profiles for them following the procedure explained above. The aim is that each construction entry should come with an illustration of its collo-profile. For the identification of new constructions, we currently use distributional semantics and neural embeddings that measure the degree of attraction between the slots of a construction (construction elements) and those linguistic expressions that are licensed to fill them (Barteld and Ziem 2020).

Another future task concerns applications of collo-profiles in different domains. For example, we intend to use collo-profiles for detecting semantic similarities and differences between constructions in different languages. For such cross-linguistic alignments of constructions, we are developing an interface that takes into account language-specific as well as cross-linguistic requirements (Czulo et al. in press). The long-term goal here is to link constructions across constructicons in German, Brazilian Portuguese and Swedish and Japanese (for an overview see Lyngfelt et al. 2018), yielding a multilingual resource (Czulo et al. in press). By doing so, frames evoked by constructions may be included as well. This is possible since numerous projects use similar annotation procedures for building their repositories (Lyngfelt et al. 2018; Boas, Torrent and Lyngfelt 2019); specifically, the Brazilian, Japanese, and German constructicon closely follow the agenda formulated in Fillmore et al. (2012) and thus share a battery of basic features.

Yet another domain of application relates to language pedagogy. Including collo-profiles in construction entries invites to specifically address special users, most prominently language learners. Even though the German Constructicon currently focuses on documenting (relations between) constructions in German, it goes far beyond a mere reference work: As the coverage of constructions progresses, the emerging repository becomes more useful as a reference tool for various practical purposes; it may serve, for example, as a resource for language learning and teaching by helping to create learner-specific gap-filling tasks based on formal and semantic constraints as evidenced in the construction entries.

Needless to say that collo-profiles as introduced and elaborated prominently by Thomas Herbst are invaluable for any kind of constructicon.