Volume 22, Issue 1

Tree-based methods are being both successfully applied and critically discussed in corpus linguistics. In this article, we would like to contribute a few aspects to this discussion from a methodological point of view. These aspects include the interpretation of interaction effects in single trees and random forests, as well as more general aspects like stability and overfitting. In particular, we have conducted a simulation study to investigate an approach suggested by Gries for computing the importance of interactions in random forests more systematically than the previous literature. The evidence of this simulation study shows that, even when interaction predictors are explicitly added, the permutation variable importance is not suited for distinguishing between main effects and interaction effects or between interaction effects of different orders. We also discuss the use of partial dependence (PD) and individual conditional expectation (ICE) plots for illustrating the functional form and potential interaction effects, and other means of interpretable machine learning.

A lot of importance is indirectly attributed to the orthographic word: it constitutes the basis of any task that is preceded by tokenization and presents material for stimuli in psycholinguistic experiments. But in many writing traditions, the orthographic word is representative of isolated entries in the lexicon and largely ignores phonological processes of production. This study examines near-naive word separation in Swiss German using a corpus of text messages, revealing distinct patterns of orthographic segmentation driven by phonological processes such as assimilation and epenthesis. Compared to Standard German, Swiss German exhibits fewer orthographic words, suggesting heightened representation of prosodic dependencies in writing. Writers prioritize phonology over syntax when deviating from standard German space insertion conventions. These findings increase doubts about the meaningfulness of orthographic representation for word-based comparative linguistic research and highlight the importance of integrating phonological information into natural language processing models.

Switch-reference is a family of grammatical devices whose primary function is to indicate whether two linked clauses have coreferential pivots, where the pivot is a prominent argument in each clause. In some languages, in addition to their function of reference tracking, switch-reference markers can be used to indicate whether the events or situations described by the two linked clauses differ with respect to some parameter, such as time, place or actuality. This phenomenon is known as non-canonical switch-reference. Whether canonical and non-canonical switch-reference marking are distinct grammatical phenomena is still an open question. In this paper, we investigate uses of switch-reference markers in a corpus of Mbyá Guaraní (Tupian) narratives, and we argue that the alternation between canonical and non-canonical uses is an epiphenomenon of the multifactorial and probabilistic nature of switch-reference marker choice. In this perspective, there is only one grammatical process of switch-reference marking and the distinction between canonical and non-canonical switch-reference marking is matter of language use.

Hungarian shows variable front vowel harmony, particularly in suffixed back vowel + [ɛ] nouns. The study aims to address two main research questions: (1) To what extent does stem-level information (similarity across stems) predict suffix variation for back vowel + [ɛ] stems in Hungarian corpus data? (2) Do suffixes themselves predict suffix variation beyond the stem-level information? We draw on a dataset of 200 noun stems, 4,501 suffixed forms and 4 × 10 6 tokens, based on the New Hungarian Webcorpus, and use a K-Nearest Neighbours learner and a hierarchical generalised linear model to address these questions. We find that the majority of back vowel + [ɛ] stems show variable vowel harmony, that this depends on stem similarity and that similarity effects are amplified by vowel-initial suffixes. This points to a model of Hungarian vowel harmony in which stem- and suffix-level information are lexically specified.

Part-of-speech annotation, as an exercise in categorisation, necessitates a category schema, based on some model or theory of the grammar of the language. Such a model may (sometimes, must ) deviate from traditional approaches for human understanding, as exploration of theoretical issues arising from a Georgian POS schema illustrates. Consistency on classifying by form versus function is problematised by difficult pronoun/demonstrative and adjective/noun distinctions. Adverb subcategorisation illustrates exclusion of semantic/derivational distinctions that traditional approaches readily admit. Variation in plural inflection has implications for how diachronicity is handled, as does “zero case”. Postpositions make necessary a specific approach to cliticisation in which enclitic elements are handled as separate tokens bearing their own analysis. Suffixaufnahme provides a case study in inclusion versus exclusion of a rare but current phenomenon. Verb morphology illustrates how simplifying assumptions help favour abstraction of categories over descriptive exhaustiveness. Divergence between the resulting model and traditional characterisations do not invalidate either, but evidence how a model’s design is inseparable from its purpose. With regard to these select issues of Georgian grammar, this discussion aims both to demonstrate the overall argument regarding theorisation/schematisation for a specific, practical purpose (POS annotation) and to justify solutions proposed to problems at hand.

This article showcases elastic net regression as a means to build fairer models of morphosyntactic variation. Elastic net allows lexical items to appear on the same level as traditional, high-level predictors, enabling fuller models of variation. We apply elastic net regression to 1,296,574 Dutch verbal cluster tokens from the SoNaR corpus, analysing a morphosyntactic alternance in Dutch subordinate clauses. Our results show morphosyntactic preferences among verbs, indicating that semantic effects are indeed at play. Further analysis shows that semantic patterns for either word order exist, though it remains difficult to glean any semantic generalisations. Still, the elastic net technique shows that the inclusion of lexical items as full predictors in a model is useful, as much of the variation left unexplained by high-level predictors can be explained in lexical terms.

The pervasiveness of the internet has given web language use a central role in society. However, the lack of multilingual corpora and scalable methods has led to the focus on English in web language research. To address this gap, the present paper sets itself in the register research tradition and explores French and Swedish web registers from a cross-linguistic angle. Methodologically we combine keyword analysis with multilingual deep learning, suggesting an approach that enables computational comparisons across languages. Specifically, we extract keywords for French and Swedish web registers, then associate the keywords with fastText word embeddings, and finally, cluster these key embeddings. The findings indicate that there are topical and functional clusters, and they are linguistically motivated and multilingual. The same clusters occur within the same registers in both languages pointing to shared topical and functional similarities – the registers are strikingly similar. The dissimilarities, in contrast, indicate that certain registers like Narrative blogs are to some extent different in French and Swedish. Moreover, grammatical specificities such as the location of adjectives explain some dissimilarities.

This paper investigates the role of corpus-based frequency metrics in predicting native speaker choice in Estonian, a Finno-Ugric language with rich inflectional morphology. We test whether frequency affects the choice between two constructions which represent the alternants of a morphosyntactic alternation in Estonian – the adessive case construction and the postpositional peal construction. Both constructions are used, inter alia , to express the spatial relationship of one object located on top of another object, e.g. Raamat on laual/laua peal ‘The book is on the table’. We further compare which out of several well-established frequency metrics yields the most accurate predictions and whether frequency on its own outperforms the predictiveness of significant semantic and morpho-syntactic variables identified in previous research (e.g. Klavan, Jane. 2012. Evidence in linguistics: Corpus-linguistic and experimental methods for studying grammatical synonymy (Dissertationes Linguisticae Universitatis Tartuensis). Tartu: University of Tartu Press, Klavan, Jane. 2020. Pitting corpus-based classification models against each other: A case study for predicting constructional choice in written Estonian. Corpus Linguistics and Linguistic Theory 16(2). 363–391). The comparison of several mixed-effects logistic regression models showed that out of seven frequency metrics considered overall, only those related to the postposition peal are predictive. Adding previously established semantic and morphosyntatic variables to the winning frequency metrics significantly improved model fit and outperformed any simpler model. We conclude that frequency affects the choice between two constructions in different ways and that it does not explain away other functional predictors.