2.1 Degree modifiers: classifications in Finnish and English

Paradis (Reference Paradis1997:87) describes most DMs as typologically unstable since they do not consistently align with the functional categories linguists have designated for them. The core function of DMs is to modify other words by presenting the degree of a feature as higher or lower than expected (Paradis Reference Paradis1997, Hakulinen et al. Reference Hakulinen, Vilkuna, Korhonen, Koivisto, Heinonen and Alho2004:§615, Biber et al. Reference Biber, Johansson, Leech, Conrad and Finegan2021:551). They can be categorized and labelled in multiple ways, for example as amplifiers (increasing intensity; e.g. very) versus downtoners (decreasing intensity; e.g. rather) (e.g. Quirk et al. Reference Quirk, Greenbaum, Leech and Svartvik1985:589–590). These umbrella labels can be further divided into subcategories: downtoners, for example, can be divided into diminishers and moderators.Footnote ¹ In this paper the focus is on DMs whose most frequently used function is to moderate the degree of the word they are connected to. They ‘seek to express only part of the potential force of the item concerned’ (Quirk et al. Reference Quirk, Greenbaum, Leech and Svartvik1985:598). However, especially in the cases of quite and ihan, they can also amplify the feature and therefore both of these functions are included as distinct levels in the analysis (see Desagulier Reference Desagulier, Glynn and Robinson2014:149 on the ambiguity of quite and Lauranto Reference Lauranto2025 of ihan; similar multifunctionality has been observed also with German ganz in Stratton Reference Stratton2020a and Norwegian ganske in Stratton & Sundquist Reference Stratton and Sundquist John2022).

Finnish and English belong to different language families, and they also diverge typologically in many respects. Finnish is primarily a synthetic language, relying heavily on grammatical agglutination. In contrast, English is predominantly an analytic language, relying more on word order and auxiliary items to convey grammatical relationships. While both Uralic and Indo-European languages share some similar synthetic ways to express degree (e.g. suffixes -hkO [halpa ‘cheap’, halvahko ‘cheapish’] and -ish for Finnish and English respectively), both language families are known to have a rich system of DMs as well (Jantunen Reference Jantunen2004:69). Previous research states that the classification of English DMs also applies well to Finnish DMs, although on a lexical level DMs are language-specific (Huumo Reference Huumo2021:89). This means that translating a specific DM from one language to another hardly works with a one-to-one correspondence. For example, the DMs under scope in this study can be used as translation pairs in some but not in other contexts: it was quite good/se oli ihan hyvä, she is quite tall/hän on ?ihan pitkä. In this case, the difference in use lies in that, unlike quite, the Finnish ihan ‘quite’ is typically not combined with open-ended scalar words such as pitkä ‘tall’ (Huumo Reference Huumo2021:89–90). From a language learner’s point of view, the use of such ambiguous expressions can be challenging.

According to Sinclair (Reference Sinclair1991:109−110), language use is a combination of open-choice and idiom principles. When following the open-choice principle, speakers can freely and flexibly construct expressions, the only constraint being grammatical correctness. Coexisting with this, the idiom principle suggests that speakers rely on semi-preconstructed, fixed expressions which are not analysable into segments and which, essentially, help streamline communication. Hasselgård (Reference Hasselgård2022:384) has noted that the combinations of DMs and the words they attribute range from relatively fixed, idiomatic combinations to free combinations that leave room for linguistic creativity.

2.2 Previous usage-based studies to learning near-synonymous DMs

This study approaches language and its learning from a usage-based perspective, according to which a language structure emerges from language use (e.g. Tomasello Reference Tomasello2003). For the language learner to be able to reach the goal of understanding and conveying messages in the target language, they must learn constructions of that language (i.e. the ways in which forms, meanings, and functions of these messages are combined). According to Goldberg (Reference Goldberg2019:2), constructions allow speakers to express messages effectively and apply their linguistic knowledge to new situations while respecting the conventions of the speech community. NSs are known to be able to use constructions in a productive and creative manner while intuitively avoiding expressions that would be considered ‘not quite right’ (ibid.). From the point of view of near-synonymous words, this intuition guides the speaker to prefer one expression over the other to convey the wanted meaning-in-context (Goldberg Reference Goldberg2019:25). How to model this kind of hidden knowledge to language learners to give them access to nuanced use and understanding of the target language remains an intriguing question.

Comparing L1 use to L2 use is not unusual in corpus-based research, and DMs have also received attention in previous corpus-based L2 studies. Some studies have identified a tendency for learners to over-use DMs (e.g. Lorenz Reference Lorenz and Granger1998, Hinkel Reference Hinkel2003, de Haan & van der Haagen Reference de Haan, van der Haagen, Granger, Gilquin and Meunier2013), while other studies have found the opposite (e.g. Granger Reference Granger and Cowie1998, Schweinberger Reference Schweinberger2020). For example, Wirtz (Reference Wirtz2025) concludes based on an overview of earlier research that these opposing results could be explained by L2 learners’ different exposure times to the target language, and that increasing language proficiency can lead to mirroring or even exceeding typical intensification rates observed among L1 speakers. Additionally, Pyykönen (Reference Pyykönen2023a, Reference Pyykönen2023b) studied epistemic expressions which are functionally related to DMs in that both can be used to express the speaker’s confidence in a proposition, and found that the frequency and types of epistemic expressions used by Finnish learners of English is influenced by proficiency level and task type.

In L2 Finnish (F2), the use of DMs has been examined by Jantunen (Reference Jantunen2015), who used keyword analysis to compare the prominence of different kinds of DMs in different varieties of F2. The comparison was between two adult learner groups, one of whom had learned Finnish mostly outside Finland and another who had learned Finnish mostly in Finland. The results of Jantunen’s (Reference Jantunen2015) study indicate that while DMs were frequent across the board, it seemed that informal or more colloquial DMs were more common especially in the context of Finnish learned in Finland, suggesting that learners who are likely exposed to a higher amount of (informal) spoken language are more likely to transfer informal expressions into their written output as well. A similar tendency of NNSs over-using DMs typical for conversational style in more formal, written contexts has also been observed in English (e.g. Hinkel Reference Hinkel2003, Schweinberger Reference Schweinberger2020).

The use of DMs is linked to both lexicon- and grammar-related variables, and these interactions have been studied in previous L2 studies as well, most exploring L2 English (E2). For example, Schweinberger’s (Reference Schweinberger2020) findings highlight the importance of considering a multitude of features in the surrounding context of the DM. Applying the same multivariate MuPDARF method as this study, Schweinberger (Reference Schweinberger2020) found that the semantic category of the adjectives amplified with the amplifier very was connected to the likelihood of E2 speakers making non-target-like choices. Additionally, the syntactic function of the modified adjectives proved important to the target-like use. Thus, to gain a comprehensive view of DM use, it is important to examine a multitude of constructional features related to them, which we do in the present study by employing the MuPDARF approach.

3.1 Corpus data: native and non-native written academic Finnish and English

For Finnish, we make use of two corpora (Ivaska Reference Ivaska2014) compiled at the University of Turku: The Corpus of Advanced Learner Finnish (LAS2) and The Corpus of Academic Finnish (LAS1). Both corpora have been lemmatized and annotated grammatically. Additionally, a professional evaluator assessed at least two texts of each writer in the LAS2 for proficiency level according to the CEFR scale, and most texts in the current dataset were found to be at the B2 and C1 levels.

For E2, we make use of the University of Helsinki’s E-thesis and the University of Turku’s Academic Lexis in L2 Speech and Writing (ACALEX) corpora. For L1 English (E1), we draw data from the Michigan Corpus of Upper-Level Student Papers (MICUSP), which contains samples of academic writing collected in the United States; our results thus only pertain to US English. Of the English corpora, E-thesis is tagged for part-of-speech and syntactic function following a slightly different terminology to the Finnish data; MICUSP and ACALEX are unannotated. The necessary syntactic annotations were coded and unified manually using the same scheme as the Finnish data by one of the authors. Additionally, E-thesis contains no record of the L1 of the writer. To determine if the authors’ L1 was Finnish, we used the method adopted in for example Gao (Reference Gao2016) and Wu & Pan (Reference Wu and Pan2024): after retrieving all the instances of quite/rather/fairly from the data, a list of names was retrieved for all the writers. This list was then independently annotated by both authors with respect to whether the name was a traditionally Finnish one, which, based on the Finnish linguistic landscape, is a fairly good indicator of the writer’s L1. The precision of the method is likely rather good, as although some F1 writers might have been left out, the remaining writers are very likely to be F1 speakers. After the annotation, the authors agreed on 93.5% of the cases, and after only including the cases where both authors agreed, we were left with altogether 255 MA theses. The reported token count (Table 1) for the E-thesis is relatively high compared to others as it also includes the lists of references, which have been left out of the other corpora.

Table 1.

Overview of the corpora

To ensure the comparability across the different datasets (i.e. FI/EN and L1/L2), we limited our scope to academic undergraduate texts from two major fields of study: the humanities and social sciences. In addition to this, we matched the datasets according to academic genre, one of the variables in our statistical model. In our data, we use the labels ‘essay’ (a shorter text investigating an idea or an experience) or ‘thesis’ (a text that involves original research or the plans to conduct one) for the different types of texts. The Finnish data also include texts labelled ‘exams’ (a short text showcasing the writer’s knowledge about a certain topic, often shorter and more free-form than essays). The exams were written in a supervised setting. Our hypothesis is, that while the chosen operationalization of genre will not give a full picture of general genre differences, the stylistic conventions between essays and theses, for example, could show signs of similar, albeit more subtle, patterns that have been found between academic texts and spoken conversation (Biber et al. Reference Biber, Johansson, Leech, Conrad and Finegan2021:560–564). An overview of the datasets is shown in Table 1. As seen in Table 1, the proportions of genre between the corpora vary, which is accounted for in the MuPDARF model.

3.2 Data extraction and annotation

The DM groups examined in this study are aika/melko/ihan ‘quite’ and quite/rather/fairly. The selection of the target expressions began by retrieving all adverbs in the F2 corpus (LAS2). From this list, we narrowed our scope to the most frequent downtoning DM (aika) that had sufficiently frequent near-synonyms (melko and ihan) to include in a quantitative analysis. As an external support for the near-synonymity, we used Kielitoimiston sanakirja (KS 2020), an easily accessible, general online Finnish dictionary. KS (2020) uses aika to define the meaning of melko and vice versa. Neither variant is mentioned in the definition of ihan, which is described mainly as an amplifying adverb but having downtoning use in everyday spoken Finnish. After this, we selected a comparable DM group in English. A widely used Finnish–English learner dictionary, MOT sanakirja (n.d.), gives quite as a translation for all the chosen Finnish DMs. To determine the synonyms of quite to be used in the analysis for English, we retrieved all adverbs in the grammatically annotated E-thesis corpus. Of the synonyms given for quite (‘as in “pretty”, to some degree or extent’) in the Merriam-Webster dictionaries (2025), we found that fairly and rather were the most frequent, and thus we decided to use them in our analysis.

After this, we retrieved all the instances of the target expressions in the L1 corpora as well. The retrieval was conducted using AntConc 4.0.11 (Anthony Reference Anthony2022) for LAS1, LAS2, ACALEX, and MICUSP. A web-based interface was used for E-thesis. All instances of the target DMs, along with their immediate surrounding context (i.e. the sentences they were used in) were then saved into a spreadsheet file. As the E-thesis dataset was substantially larger than the other datasets, a random sample of 300 uses was drawn to match the number of results from the E1 data. The frequencies of the target DMs are shown in Table 2 along with their distributions within the datasets.

Table 2.

The frequencies of the target DMs in the data

To investigate which speaker-, text- and grammatical context-specific factors influence the choice of DMs in both L1 and L2 Finnish and English, each instance of the target expressions was coded for 12 variables based on earlier research. All variables included in the analysis are shown in Table 3 and will be more closely discussed in what follows.

Table 3.

An overview of the variables used in the analysis

As the choice of DM is closely related to the expression it modifies (e.g. Ito & Tagliamonte Reference Ito and Tagliamonte2003, Jantunen Reference Jantunen2015, Schweinberger Reference Schweinberger2020, Stratton Reference Stratton2020a), many of the variables used in the analysis characterize the modified expression in the R1 position (first word after the DM excluding possible articles). These include its lemma (e.g. quite well , melko hyvin ), syntactic function (funr1), part of speech (posr1), and case. Given that some of the lemmas were used very infrequently, we decided to include only the more frequent ones in the FI and EN datasets and classify the lemmas with fewer than five instances as ‘other’ (see Table 3). funr1 covers the syntactic function of the modified expression in the R1 position: predicative (e.g. I’m fairly sure ), adverbial (e.g. This corresponds fairly well with the information), or a modifier (e.g. There has been quite a consensus ). Adjectival, adverbial, and nominal modifiers were conflated into one category as the number of R1 expressions acting as adjectival and adverbial modifiers was extremely low.

The variable case is only included in the Finnish data as, in Finnish, unlike in English, the modified expressions often appear in an inflected form. In our dataset, there are five possible cases for the modified expressions: nominative (e.g. vaatimustaso on aika kova ‘the requirement level is quite tough’), partitive (e.g. opiskelijat olivat aika eritasoisia ‘the students were of quite different levels’), genitive (e.g. sain melko hyvän kokonaiskuvan ‘I got a pretty good overall picture’), essive (e.g. olen pitänyt osaamistani melko vahvana ‘I have considered my skill set quite strong’), and translative (e.g. tehtävä osoittautui aika haastavaksi ‘the task turned out to be quite challenging’).

The variables negation and word order are related to the broader syntactic structure where the DM appears in. Both variables have been found to be important either for DM use (e.g. Jantunen Reference Jantunen2004, Bylinina et al. Reference Bylinina, Tikhonov, Garmash, Bouamor, Pino and Bali2023) or for distinguishing between L1 and L2 language use (e.g. Ivaska Reference Ivaska2015, Kekki & Ivaska Reference Kekki and Ivaska2022). Of these, negation is related to whether the sentence the DM is used in is negated or not (e.g. not quite the same vs. to be rather sceptical). word order, in turn, refers to the placement of the DM and the expression it modifies in the sentence: before the predicate verb (token_verb; e.g. Also fairly clear is the relationship between gender and power) or after it (verb_token; e.g. There is rather little evidence).

Finally, three variables related to the semantic qualities of the DMs and their context of use were included in the analysis, as semantic variables have been reported to impact the likelihood of NNSs making non-targetlike DM choices (e.g. Ito & Tagliamonte Reference Ito and Tagliamonte2003, Jantunen Reference Jantunen2015, Schweinberger Reference Schweinberger2020). First, degree is related to whether the DM was used as a downtoner or an amplifier. Although the target expressions are mainly used as downtoners, some of them (ihan, quite, rather) can also be used as amplifiers (e.g. KS 2020, Merriam-Webster 2025). sentiment, on the other hand, pertains to the emotional tone and message of the DM’s context of use (i.e. the sentence in which the DM is used). More specifically, sentiment is divided into the categories positive (e.g. seems quite sensible, ymmärsin […] melko hyvin ‘I understood quite well’), negative (e.g. the rather peculiar description, suunnistus oli melko summittaista ‘orienteering was quite random’), and neutral (e.g. it is quite likely, ei tarkoita ihan samaa ‘does not mean quite the same’). Finally, semantics was used to categorize the modified expressions into semantically related groups using the categorization proposed by Jantunen (Reference Jantunen2004:137–138). For example, expressions such as confident were placed under ‘certainty’ and easy under ‘difficulty’. As many of the original sixteen categories in Jantunen (Reference Jantunen2004) occurred extremely infrequently in our datasets, we decided to include only the eight most frequent classes in the FI and EN datasets and conflate the other categories under the group ‘other’. The remaining groups in each of the datasets are listed in Table 3.

To ensure that the coding scheme for the three semantic variables (degree, sentiment, and semantics) was consistent, a random sample of 200 DMs (100 in Finnish and 100 in English) was coded by both authors independently; the initial agreement rates from the two annotations are shown in Table 4. All instances where disagreements arose were examined more closely and resolved by devising a general code of conduct for similar cases.

Table 4.

Agreement in the coding of the semantic variables

3.3 MuPDARF method

Methodologically, we make use of the MuPDARF approach (Gries & Deshors Reference Gries and Deshors2014, Deshors & Gries Reference Deshors and Gries2016, Gries & Deshors Reference Gries and Deshors2020). The method has been developed to compare NNS production with NS production, but it has also been applied to other types of research designs, such as comparing different varieties of world Englishes (Deshors & Gries Reference Deshors and Gries2016) and closely related Finnic languages (Ivaska & Tamm Reference Ivaska and Tamm2024).

This approach allows us to explore the linguistic choices an NS makes under the exact same circumstances as an NNS: in other words, if they make similar choices, and if they do not, the constructional factors that explain it are examined by considering the multiple factors affecting the speakers’ language use (Gries & Deshors Reference Gries and Deshors2014, Reference Gries and Deshors2015). As language use, especially the use of near-synonymous expressions, is rarely an either/or situation but rather a continuum between choices, the more recent applications of the method also model the so-called middle ground where the NS would likely be satisfied with either (or, in our case, any) option (Gries & Deshors Reference Gries and Deshors2020). We conduct the MuPDARF analysis to both languages separately and then compare the results to shed light on (i) possible features shared by language learners of both languages and (ii) possible language-specific differences of near-synonymous DM use. The general steps of the method are shown in Figure 1 (adapted from Ivaska & Tamm Reference Ivaska and Tamm2024:21).

Figure 1.

Flowchart of the implemented MuPDARF approach.

To begin our statistical analysis, we used the ranger (Wright & Ziegler Reference Wright and Ziegler2017) package in R to train two language-specific multi-class random forest models to predict which one of the near-synonymous DMs (aika/melko/ihan; quite/rather/fairly) the NSs were likely to use in academic written texts.Footnote ² The model uses the variables described in Section 3.2 to make predictions as to how likely it is that a certain DM is used in each case (e.g. if the genre is X and the lemma is Y, the DM choice is most likely to be Z). The resulting model for Finnish is as follows:

$$\eqalign{ & {\rm{FI\!:}}\;{\rm\small{{CHOICE}}} \sim {\rm\small{GENRE}} + {\rm\small{LEMMA}} + {\rm\small{FUNR}}1 + {\rm\small{POSR}}1 + {\rm\small{CASE}} + {\rm\small{NEGATION}} + {\rm\small{WORD ORDER}} \cr & + {\rm\small{SENTIMENT}} + {\rm\small{SEMANTICS}} + {\rm\small{DEGREE}} \cr} $$

The English model is the same, except for the variable case, which was left out as irrelevant.

According to the approach, if the model can predict DM use in the NS data sufficiently accurately, it can be used on the NNS data to predict which of the near-synonymous DMs NSs would choose in each context of occurrence, as well as to determine whether the NNS choice is canonical, or ‘nativelike’.Footnote ³ To evaluate both models, we compared their performance to a baseline determined by how well the model would perform in a situation where it always predicted the most common DM in the data. For Finnish, this level of chance was 86.0%, and with an out-of-bag (OOB) error of 0.075 and a (micro-averaged) F-score of 0.959 the model performed extremely well. For English, the level of chance was 62.4%, which was also outperformed by the model with an OOB error of 0.24 and a (micro-averaged) F-score of 0.762. Thus the fit of both models was good enough to proceed to the next step of the analysis.

Next, we trained two language-specific random forest models that focused on the NS-likeness of the choices observed in the NNS data. More specifically, the model made it possible to analyse whether and how the canonical NNS choices diverge from the non-canonical NNS choices. With the addition of the chosen DM and the NNSs’ L1s for the F2 data, the predictor variables are the same as in the first models for both languages (again, the English model is the same, except for the variable case):

$$\eqalign{ & {\rm{FI: F}}1{\rm{ - likeness}}\sim{\rm\small{{CHOICE}}} + {\rm\small{{L}}}1 + {\rm\small{{GENRE}}} + {\rm\small{{LEMMA}}} + {\rm\small{{FUNR}}}1 + {\rm\small{{POSR}}}1 + {\rm\small{{CASE}}} \cr & + {\rm\small{{NEGATION}}} + {\rm\small{{WORDORDER}}} + {\rm\small{{SENTIMENT}}} + {\rm\small{{SEMANTICS}}} + {\rm\small{{DEGREE}}} \cr} $$

This step accounts for the cases where the NS would actually be comfortable using two or all of the near-synonymous DM variants under scope by quantifying the middle ground, which makes use of the degree of certainty that the model predicts as the non-canonical L2 choice. Following Gries & Deshors (Reference Gries and Deshors2020:80), we defined the middle ground as ‘the predicted probability that corresponds to one unit of log loss, i.e. to classifications/predictions that, if wrong, lead to log loss values of 1 [or more]’. In most previous studies, the focus has been on cases where there are two choices, and therefore the middle ground has been calculated using the cut-off point of 0.5 +/− log loss. As our study examines the choice between three different DMs, the cut-off point was set to 0.333 +/− logloss. This meant that our middle ground probability estimates varied between 0.226 and 0.440 for Finnish, and 0.033 and 0.633 for English. Thus the predictions with probabilities that landed within the middle ground are likely to be cases where more than one near-synonymous option is possible. Finally, when looking into the variable-specific results, we explored the NS-likeness of the NNS choices by omitting the middle ground and only examining DM choices that the model confidently predicted as either nativelike or non-nativelike.