Linguistic networks: Connecting constructions within and between languages
This project uses Construction Grammar to develop a linguistic network that (a) accounts for Swedish grammatical constructions and (b) connects them to constructions in other languages.
Linguistic analysis is typically stratified into levels (syntax, lexicon, etc.) and tends to disregard the interplay between the levels. Accordingly, linguistic patterns that combine properties from different levels are notoriously difficult for second-language learning, language technology and translation alike. For example, a Swedish pseudo-coordination such as “Hon går och grubblar”, which integrates properties from several linguistic levels, is literally but incorrectly translated “She goes and ponders” in Google Translate, instead of “She is worrying”.
Construction Grammar, in contrast, treats language as a network. Structures are analyzed as constructions, i.e., conventionalized combinations of form and function. A construction may combine properties from several linguistic levels, which benefits accurate descriptions of, e.g., pseudo-coordinations. With this approach, it is also possible to capture connections between structurally distinct constructions in different languages and contribute to better machine translations.
We have previously developed a Swedish construction database that will now be converted to a linguistic network. With tools from language typology, we will also connect it to networks for other languages. These two objectives are closely integrated and mutually dependent.
Linguistic analysis is typically stratified into levels (syntax, lexicon, etc.) and tends to disregard the interplay between the levels. Accordingly, linguistic patterns that combine properties from different levels are notoriously difficult for second-language learning, language technology and translation alike. For example, a Swedish pseudo-coordination such as “Hon går och grubblar”, which integrates properties from several linguistic levels, is literally but incorrectly translated “She goes and ponders” in Google Translate, instead of “She is worrying”.
Construction Grammar, in contrast, treats language as a network. Structures are analyzed as constructions, i.e., conventionalized combinations of form and function. A construction may combine properties from several linguistic levels, which benefits accurate descriptions of, e.g., pseudo-coordinations. With this approach, it is also possible to capture connections between structurally distinct constructions in different languages and contribute to better machine translations.
We have previously developed a Swedish construction database that will now be converted to a linguistic network. With tools from language typology, we will also connect it to networks for other languages. These two objectives are closely integrated and mutually dependent.
Final report
This project is based on the constructionist idea of language as a network. We have developed a network model for Swedish, which accounts for a wide range of linguistic constructions, network relations between them, and a system for combining them into larger units such as sentences. We have also, in international collaboration, developed an associated language-neutral network model to align constructions in different languages. Both models are implemented in freely available linguistic resources: Svenskt konstruktikon (‘The Swedish Constructicon’) and MoCCA (Model of Comparative Concepts for Constructicon Alignment), respectively.
A large benefit with a constructionist perspective is the ability to capture the interplay between linguistic structures that are otherwise treated as separate levels. Linguistic analysis is typically stratified into levels (syntax, lexicon, etc.) and, hence, tends to disregard the interplay between the levels. Accordingly, linguistic patterns that combine properties from different levels are notoriously difficult for second-language learning, language technology and translation alike. For example, a Swedish pseudo-coordination such as “Hon går och grubblar”, which integrates properties from several linguistic levels, is literally but incorrectly translated “She goes and ponders” in Google Translate, instead of “She is worrying” (2026-03-31). Construction Grammar, on the other hand, treats language as a network. Structures are analyzed as constructions, i.e., conventionalized combinations of form and function. A construction may combine properties from several linguistic levels, which benefits accurate descriptions of, e.g., pseudo-coordinations.
The project addresses three research questions:
1) How can the grammar of a language, Swedish, be fruitfully accounted for as a network of constructions?
2) How can constructions be connected across languages through a linking system of shared features?
3) How can the somewhat conflicting demands of describing a language on its own terms and capturing cross-linguistically relevant properties in language-neutral terms be conciliated in a constructionist approach?
In response to the first question, we have firstly developed a network model for Swedish, in which linguistic constructions are associated by shared properties (Blensenius & Lyngfelt 2025). Constructions are annotated for both formal properties (e.g., coordination, compound, noun phrase) and functional properties (e.g., comparison, resultative, interaction). Thus, they are linked to other constructions with the same properties, thereby establishing a multidimensional network of constructions. Secondly, we have developed a syntax model, accounting for the combination of constructions into larger units and a system of central syntactic constructions such as different clause types (Andréasson & Lyngfelt, in press). In combination, the network model and the syntax model provide a coherent construction grammar for Swedish, the first of its kind. It provides a platform for further constructionist research on and description of Swedish. Arguably, it also provides proof of concept of the idea of language as a network of constructions.
The second research question applies a similar network approach to crosslinguistic comparison of constructions. In an international collaborative effort, we have built a model to align constructions in different languages via shared properties, analogous to the afore-mentioned network model for Swedish, except that the connecting properties are defined language-neutrally instead of by Swedish standards (Lyngfelt et al. 2022). The model is based on so-called comparative concepts (CCs) of the kind employed in language typology. Hence, it is called MoCCA: Model of Comparative Concepts for Constructicon Alignment (Lorenzi et al. 2024). Specifically, the model is based on a set of CCs compiled by the typologist Bill Croft, who has also been directly involved in the development of MoCCA. The purpose of this model is to support non-biased language comparison. It is a well-known problem that corresponding constructions in different languages rarely, if ever, are fully equivalent, and, consequently, that cross-linguistic comparison tends to be conducted on the terms of the source language. For a non-biased comparison it is desirable to have a language-neutral base of comparison, and this is what MoCCA offers. Furthermore, the applicability of MoCCA goes beyond direct comparisons between constructions in particular languages, since the CCs are applicable to all languages. When a construction is linked to a CC it is thereby associated with all other constructions, regardless of language, that are connected to the same CC. Consequently, the model gets more powerful the more constructions and the more languages are linked to it.
The third research question concerns the relation between these two models, the Swedish construction network and the cross-linguistic MoCCA network, and how to accommodate the language particular and the language neutral. Our solution is to treat Swedish constructions on Swedish terms and afterwards link them to MoCCA where applicable. This approach has led to minor discrepancies, for instance that certain reflexive constructions in Swedish do not fit the language-neutral definition of reflexive in MoCCA (due to lack of ‘reflexive’ meaning), but as long as the differences are properly documented they do not pose any significant problem. The MoCCA guidelines have been updated to handle such differences, which further strengthens its potential for crosslinguistic comparison.
All in all, the project has fulfilled its purpose: to develop a Swedish construction network, on the one hand, and create a model to align constructions in different languages, on the other. The Swedish network model is implemented in the web resource Svenskt konstruktikon, in which we have also started linking Swedish constructions to MoCCA. The syntax model, however, is not implemented yet. Furthermore, there are of course many more Swedish constructions to describe and more network relations to establish, both language-internally and cross-linguistically. Morphology, in particular, is so far less well covered than syntax, and the integration between productive constructions and particular lexical units should be better accounted for. The presently established system points in itself to related constructions in need of coverage. It also provides both a context and a format for such further development.
Throughout the project, the results have been presented and discussed at different scientific conferences, both internationally and in Sweden, and the main findings have been reported in a number of scientific publications, prioritizing Open Access. We continuously collaborate with constructicon teams working on other languages, as well as with parties applying the material to language technology and language teaching.
A large benefit with a constructionist perspective is the ability to capture the interplay between linguistic structures that are otherwise treated as separate levels. Linguistic analysis is typically stratified into levels (syntax, lexicon, etc.) and, hence, tends to disregard the interplay between the levels. Accordingly, linguistic patterns that combine properties from different levels are notoriously difficult for second-language learning, language technology and translation alike. For example, a Swedish pseudo-coordination such as “Hon går och grubblar”, which integrates properties from several linguistic levels, is literally but incorrectly translated “She goes and ponders” in Google Translate, instead of “She is worrying” (2026-03-31). Construction Grammar, on the other hand, treats language as a network. Structures are analyzed as constructions, i.e., conventionalized combinations of form and function. A construction may combine properties from several linguistic levels, which benefits accurate descriptions of, e.g., pseudo-coordinations.
The project addresses three research questions:
1) How can the grammar of a language, Swedish, be fruitfully accounted for as a network of constructions?
2) How can constructions be connected across languages through a linking system of shared features?
3) How can the somewhat conflicting demands of describing a language on its own terms and capturing cross-linguistically relevant properties in language-neutral terms be conciliated in a constructionist approach?
In response to the first question, we have firstly developed a network model for Swedish, in which linguistic constructions are associated by shared properties (Blensenius & Lyngfelt 2025). Constructions are annotated for both formal properties (e.g., coordination, compound, noun phrase) and functional properties (e.g., comparison, resultative, interaction). Thus, they are linked to other constructions with the same properties, thereby establishing a multidimensional network of constructions. Secondly, we have developed a syntax model, accounting for the combination of constructions into larger units and a system of central syntactic constructions such as different clause types (Andréasson & Lyngfelt, in press). In combination, the network model and the syntax model provide a coherent construction grammar for Swedish, the first of its kind. It provides a platform for further constructionist research on and description of Swedish. Arguably, it also provides proof of concept of the idea of language as a network of constructions.
The second research question applies a similar network approach to crosslinguistic comparison of constructions. In an international collaborative effort, we have built a model to align constructions in different languages via shared properties, analogous to the afore-mentioned network model for Swedish, except that the connecting properties are defined language-neutrally instead of by Swedish standards (Lyngfelt et al. 2022). The model is based on so-called comparative concepts (CCs) of the kind employed in language typology. Hence, it is called MoCCA: Model of Comparative Concepts for Constructicon Alignment (Lorenzi et al. 2024). Specifically, the model is based on a set of CCs compiled by the typologist Bill Croft, who has also been directly involved in the development of MoCCA. The purpose of this model is to support non-biased language comparison. It is a well-known problem that corresponding constructions in different languages rarely, if ever, are fully equivalent, and, consequently, that cross-linguistic comparison tends to be conducted on the terms of the source language. For a non-biased comparison it is desirable to have a language-neutral base of comparison, and this is what MoCCA offers. Furthermore, the applicability of MoCCA goes beyond direct comparisons between constructions in particular languages, since the CCs are applicable to all languages. When a construction is linked to a CC it is thereby associated with all other constructions, regardless of language, that are connected to the same CC. Consequently, the model gets more powerful the more constructions and the more languages are linked to it.
The third research question concerns the relation between these two models, the Swedish construction network and the cross-linguistic MoCCA network, and how to accommodate the language particular and the language neutral. Our solution is to treat Swedish constructions on Swedish terms and afterwards link them to MoCCA where applicable. This approach has led to minor discrepancies, for instance that certain reflexive constructions in Swedish do not fit the language-neutral definition of reflexive in MoCCA (due to lack of ‘reflexive’ meaning), but as long as the differences are properly documented they do not pose any significant problem. The MoCCA guidelines have been updated to handle such differences, which further strengthens its potential for crosslinguistic comparison.
All in all, the project has fulfilled its purpose: to develop a Swedish construction network, on the one hand, and create a model to align constructions in different languages, on the other. The Swedish network model is implemented in the web resource Svenskt konstruktikon, in which we have also started linking Swedish constructions to MoCCA. The syntax model, however, is not implemented yet. Furthermore, there are of course many more Swedish constructions to describe and more network relations to establish, both language-internally and cross-linguistically. Morphology, in particular, is so far less well covered than syntax, and the integration between productive constructions and particular lexical units should be better accounted for. The presently established system points in itself to related constructions in need of coverage. It also provides both a context and a format for such further development.
Throughout the project, the results have been presented and discussed at different scientific conferences, both internationally and in Sweden, and the main findings have been reported in a number of scientific publications, prioritizing Open Access. We continuously collaborate with constructicon teams working on other languages, as well as with parties applying the material to language technology and language teaching.