1Usage-based models of language posit that grammar is acquired, represented mentally, and accessed in a ‘bottom-up’ fashion, through exposure to usage events (Bybee, 2006; Langacker, 1988; 1999; 2000). Because usage events are intrinsically context dependent, grammar is a structured inventory of symbolic units whose architecture is shaped by the accumulation of linguistic experience. Given this focus on context and this holistic approach to linguistic experience, a fast-growing community of cognitive linguists have begun to realize that the assumptions of their own theoretical framework were essentially empirical and had to be tested (Gibbs, 2007; Gries, Hampe, & Schönefeld, 2005; Tummers, Heylen, & Geeraerts, 2005). As a result, two lines of research emerged at the crossroads of lexical semantics and construction grammar, the first one focusing on how sociolinguistic factors determine the choice of synonyms, and the second one relying on multifactorial methods to differentiate between near-synonyms (for an overview, see Geeraerts, 2010: 263-264). This paper belongs to the second research area.
2I propose quantitative techniques to capture subtle usage-based relations between two near-synonyms in British English: quite and rather. Although quite and rather can modify other adverbs (quite frankly; rather desperately), noun phrases (quite a sight; rather a shock), or even verb phrases (I quite understand; I rather enjoyed it), my investigation is restricted to the far more frequent contexts where these intensifiers are used as degree modifiers of adjectives:
- 1 Throughout this paper, examples from the British National Corpus appear with an indication of their (...)
(1) We are quite different people now, we need different things. (BNC-CEX)1
(2) Yes, and there’s a rather odd smell in this room... (BNC-J17)
3When quite and rather modify attributive adjectives, they can occur in pre-determiner position, a behavior that other intensifiers do not show:
(3) a. That has proved to be a quite difficult question to answer.
b. That has proved to be quite a difficult question to answer.
(4) a. That is a rather difficult question to answer.
b. That is rather a difficult question to answer.
(5) a. I know it’s a fairly difficult question.
b. ??I know it’s fairly a difficult question.
4Allerton (1987: 25) believes that the choice of pre-determiner position over the default pre-adjectival position is more than a matter of style or formality. In this paper, my first goal is to use quantitative corpus-based techniques to decide if the difference between the pre-adjectival and the pre-determiner patterns of quite and rather translates into a difference in meaning.
5Allerton observes further that some restrictions apply depending on whether the adjective that quite modifies is scalar or absolutive:
(6) a. I mean this is quite a good idea / ?a quite good idea actually.
b. This is ?quite an excellent idea / a quite excellent idea.
6More precisely, pre-determiner uses of quite and rather seem to inherit the emphatic and exclamatory meaning of <quite/rather + NP> (e.g. quite a pen!, rather a fool!). Typically, when the intensifier precedes the determiner, it has a modal meaning and it functions as a sentence-modifier, not a word-modifier (Allerton 1987, Bolinger 1972, Stoffel 1901). Regarding quite, however, Bolinger (1972: 137) observes the following:
(…) somehow along the way, the indefinite article has ceased to separate the two functions consistently, with the result that quite a and a quite, for example, form an alternating pattern with so slight a difference in meaning that outside factors may decide the choice.
7In the light of subsequent research on synonymy, Bolinger’s observation cannot be taken for granted. Bolinger himself claims elsewhere that “if two ways of saying something differ in their words or their arrangement they will also differ in meaning” (1977: 1). In the same vein, Clark’s Principle of Contrast stipulates that “every two forms contrast in meaning” (Clark 1987: 2, 1992: 172), and Goldberg’s Principle of No Synonymy states that “[i]f two constructions are syntactically distinct, they must be semantically or pragmatically distinct” (Goldberg 1995: 67).
8Admittedly, the semantic difference between the pre-adjectival and pre-determiner structures in (0) is slim. However, if we examine usage closely, pre-adjectival and pre-determiner uses do not overlap neatly. Indeed, a cursory search in the British National Corpus (World Edition) reveals that quite tends to intensify adjectives that denote importance in size (big, considerable, sizeable, substantial) when it precedes the determiner. There is no such preference when quite follows the determiner and precedes the adjective. Semantic differences between alternating patterns may be small, but not necessarily insignificant in terms of conventionally sanctioned usage. Therefore, my second goal, which is more general, is to inspect the near synonymy of quite and rather with respect to form and meaning.
9Most major studies on intensifiers have focused on the best way to infer gradability from collocational preferences (Altenberg 1991, Kennedy 2003, Lorenz 1999, 2002, Simon-Vandenbergen 2008). One reason behind this methodological preference is the following: because the vast majority of degree modifiers of adjectives function as ‘word modifiers’ (Stoffel 1901), they lend themselves to easy corpus extraction and quantification. However, the methodology used in those studies can be deemed faulty on (at least) two counts. Firstly, syntactic variables are seldom incorporated into the abovementioned collocation-based descriptions of adjective intensification (Lorenz 2002, Kennedy 2003). When they are (e.g. Lorenz 1999, Simon-Vanderbergen 2008), these variables are generally listed irrespective of their potential interaction with semantic features. Secondly, whether based on raw counts, coarse-grained relative frequencies such as percentages and counts per n-thousand words (e.g. Altenberg 1991, Paradis 1997), or far more reliable statistical association measures, collocation analysis cannot detect relevant patterns of interaction among multiple factors because it is meant for the exploration of one variable at a time. While the approach presented in this article shares substantial common ground with the above studies, it departs from them methodologically.
10I will therefore use multifactorial techniques to examine the co-occurrence of intensifiers and adjectives in the context of the syntactic patterns where they occur. These patterns are the pre-determiner vs. pre-adjectival positions of quite and rather and the predicative vs. attributive positions of the adjectives that these adverbs intensify. Results will show that the focal adjustment on the meaning of adjectives does not depend exclusively on the choice of an adverb but also on the syntax of both adverbs and adjectives.
11This paper is organized as follows. Section 2 reviews previous approaches to English intensifiers and highlights some issues in the collocation-based methodologies that these works implement. Section 3 addresses the limitations of previous collocation analyses, advocates the need to include collocation data in multidimensional tables, and proposes a new methodology to handle such tables. Sections 4 and 5 present and discuss the results. Section 6 concludes on methodological issues and implications.
- 2 For example, the adverbs nearly and almost both express near-completion on a path (both literally a (...)
12From a Cognitive Grammar perspective (Langacker 1987, 1991, 2008), synonymous expressions have identical conceptual content and impose the same construal upon that conceptual content, while near synonyms share the same conceptual content but differ in terms of construal.2 If quite and rather are indeed near synonyms, we should expect them to share the same conceptual content but to differ in terms of construal. The conceptual content that quite and rather share is essentially functional: both intensifiers index the properties denoted by the adjectives they modify on a scale (‘upwards’ or ‘downwards’). The fact that both quite and rather alternate between pre-determiner and pre-adjectival positions is a sign that their functions overlap.
13Early signs of construal differences between quite and rather can be traced back to their respective etymologies. Quite is borrowed from Anglo-Norman quit/quite (‘without opposition’), which can be traced back to Latin quietus (‘at rest’). Its original meaning is the absence of obligation, the discharge of one’s duty. We may assume that ‘freedom from coercion’ grammaticalized into ‘absence of grading restriction’ by the time quite was first attested in Middle English, because it was then used as a boosting intensifier. ‘Absence of grading restriction’ further shifted to ‘presence of a property to some degree’ in the early 19th century, when quite started to be used as a moderating intensifier:
(7) Had Lawyer L--n staid at home, His honour might have pass’d, with some, For quite a decent country Squire. (1806, T.G. Fessenden, Democracy Unveiled II. Vi. 204) (OED online).
14The diachrony of quite offers valuable clues as to how the adverb intensifies adjectives. If we posit that an adjective denotes a property P, quite signals the following interpretation path: ¬P is considered and rejected in favor of P, which is then indexed on an intensity scale. This is consistent with Bolinger (1972: 105), quoted by Gilbert (1989: 7-8), who observes that quite cold and very cold do not have the same connotations:
(8) It is quite cold this morning.
(9) It is very cold this morning.
15He writes: “the first connotes an unexpectedness that the second lacks; it would be used of a cold day in summer, for example.”
16Originally, rather combines the adverb rathe ‘quickly, rapidly’ and the comparative suffix –er. It develops from later Middle English onwards, via the decline of rathe. The OED online states that ‘[t]he modifying function (…) probably arose from the reanalysis of the adverb’s syntactic role in sentences where contrast is anaphorically implied (…)’. The idea of contrast is present in (8):
(10) He was very negligent himselfe & of a Philosophic temper & was indeede rather negligent of his person. (a1684, Diary anno 1675 (1955), J. Evelyn, IV. 60) (OED online)
17P (‘negligent’) and ¬P (‘not negligent’) are compared and P is considered prevalent to some degree. The comparative nature of rather suggests that the intensifier selects P more than ¬P. This is consonant with Allerton’s description of rather as a sentence adverbial in prosodically isolated positions:
(11) a. She’s an attractive woman(,) rather.
b. She is – rather – an attractive woman. (Allerton 1987: 27)
Here, rather signals that the use of the tonic adjective attractive is preferred to the use of another adjective.
- 3 Here, the acceptability judgment is valid for British English only. American English accepts quite (...)
18There is more to the meaning of an intensifying construction than the meaning of the intensifier. Paradis (1997: 62) argues that the semantic relation between modifiers and adjectives is bidirectional: “(…) the adjective selects a degree modifier which in turn constrains the conceptualization of the gradability of the adjective definitively (…)”. We can interpret the bidirectionality hypothesis in two ways. One way is to assume that intensifying constructions are assembled in a linear way, the adjective being selected first and then modified in an online, post hoc fashion by the adverb. This could explain why some adjectives, because of their semantics, do not combine easily with all intensifiers, e.g. ?quite curious, ?rather perfect.3 However, from a usage-based perspective, the adjective and the intensifier are more likely to be selected simultaneously. No matter which interpretation of the bidirectionality hypothesis we choose, it does not take into consideration the influence of the syntactic context where adjectives and intensifiers combine.
19Altenberg (1991) is well aware of the syntactic restrictions governing the use of boosters and maximizers (1991: 128-129) but he does not discuss them in depth. Instead, he focuses on the collocational preferences of amplifiers regardless of the phrasal contexts (1991: 130). Both Lorenz (2002) and Kennedy (2003) adopt lexeme-centered approaches. In contrast, Lorenz (1999: 201-211) devotes a whole section to adjective intensification, information structure, and phrasal environment in a comparison of English learners’ tendencies and native speakers’ preferences in four corpora. He observes that adjective intensification occurs naturally with rhematic adjectives in predicative position to the detriment of attributive occurrences. However, Lorenz does not correlate the predicative and attributive positions with the semantic classes that the different adjectives fall into. Simon-Vandenbergen (2008) proceeds likewise. While her comparative study of certainly and definitely encompasses several kinds of distinctive variables (register, genre, dialogic context, and information structure), she does not examine their correlation. In sum, previous quantitative studies on degree modifiers of adjectives do not systematically incorporate syntactic variables in their collocational description of intensification. When they do, these variables are generally described irrespective of their potential interaction with semantic features.
20Another issue is the choice of association measures to assess collocational preferences, namely raw frequencies (Lorenz 2002), basic standardized figures (Altenberg 1991, Lorenz 1999, Paradis 1997, Simon-Vandenbergen 2008), and pointwise mutual information (Kennedy 2003). Such measures do not filter away lexemes that are unrealistically too frequent or too rare regardless of the words they collocate with. Based on raw frequencies in the BNC (World Edition), the top collocate of quite and rather in pre-adjectival position is different (1 256 and 477 tokens respectively). Considering these collocation frequencies alone, one could easily jump to the conclusion that the degree of mutual association between different and both intensifiers is high. That would be overlooking the fact that different ranks among the top five adjectives in the whole corpus. If we compare these collocation frequencies to the overall frequencies of different (47 209 tokens), quite (16 087), and rather (5 521), we realize that different is distinctive of quite more than it is distinctive of rather.
21Assuming that both intensifiers and adjectives contribute to the meaning of the intensifying construction, a good place to start is to inspect co-occurrences of quite/rather and the adjectives that they modify in a large corpus and determine preferred collocations with a reliable association measure. However, because speakers do not make lexical choices regardless of morphosyntactic and semantic/pragmatic considerations, it seems implausible to assess near synonymy accurately on the basis of lexical collocation alone, as argued extensively by Divjak (2006, 2010). Accordingly, I investigate the co-occurrence of intensifiers and adjectives in the context of the syntactic patterns where they occur. Quite and rather manifest themselves across eight patterns, which are summarized in Table 1. I contrast these alternations in their respective adjectival preferences using a method from collostructional analysis (Hilpert 2008, Stefanowitsch 2013), namely multiple distinctive collexeme analysis, henceforth MDCA (Gilquin 2007, Hilpert 2006, Desagulier 2014). Collostructional analysis is a family of methods aimed at measuring the degree of attraction or repulsion that words exhibit to constructions. It consists of collexeme analysis (Stefanowitsch & Gries 2003), distinctive collexeme analysis (Gries & Stefanowitsch 2004b), and co-varying collexeme analysis (Gries & Stefanowitsch 2004a, Stefanowitsch & Gries 2005, Desagulier 2015).
Table 1. How quite and rather pattern with adjectives
22MDCA is a subtype of distinctive collexeme analysis, which is well suited for the purpose of finding subtle differences between pairs of near-equivalent constructions (e.g. the ditransitive vs. prepositional dative alternation). MDCA extends distinctive collexeme analysis to more than two constructions. It determines the probability of each lexeme’s observed frequency given its expected frequency in each construction. This probability is log-transformed and the resulting value captures distinctiveness (pbin). The co-occurrence between a lexeme and a construction is statistically significant if pbin is higher than 1.3 (p < 0.05). MDCA outputs a table that contains as many lines as there are distinctive collexemes. In addition to observed frequency, expected frequency, and pbin, the columns provide two values to help the linguist identify meaningful patterns of attraction: SumAbsDev and LargestDev. SumAbsDev gives the sum of all absolute pbin values for a particular collexeme – the higher the figure, the more the adjective deviates from its expected distribution. LargestDev indicates which construction each lexeme is a distinctive collexeme of.
23Thanks to MDCA, one can determine the probability of each adjective’s observed frequency given its expected frequency in each pattern. Only those adjectives whose absolute distinctiveness values are significant were kept, along with their respective observed frequencies in each of the eight quite/rather patterns.
24The obtained contingency table served as the basis for the constitution of a multidimensional dataset of non-negative ratio-scaled data. This dataset contains supplementary frequency data regarding the syntax and the semantics of the quite/rather constructions. Supplementary columns provide the frequency of co-occurrence between each adjective type and the following variables:
the syntactic position of the intensifier: pre-adjectival vs. pre-determiner;
the syntactic position/function of the adjective: attributive vs. predicative;
the sum total for each intensifier: quite vs. rather;
the text mode: spoken vs. written.
Supplementary rows group the frequencies of adjectives according to 59 semantic classes listed in the Appendix section. The semantic classes were inspired from Dixon & Aikhenvald (2004). Classes that matched the data better were added in an ad hoc fashion.
25The dataset was then submitted to correspondence analysis (henceforth CA), an exploratory statistical technique that takes the frequencies of multi-way tables as input, then summarizes and visualizes distances between the variables (Benzécri 1973, 1984, Greenacre 2007, Glynn 2014). More precisely, CA uses the frequencies of the dataset to:
compute a matrix based on χ2 distance and determine the probability of global association between rows and columns;
transpose the multidimensional distances to a two-dimensional plane that maps the correlations between the variables; each row and each column is thus represented as a point in the Euclidean space.
The graph can be used to visualize relative distances between (a) intensifiers, (b) adjectives and their respective semantic classes, (c) syntactic information regarding how intensifiers and adjectives pattern together.
26Multiple correspondence analysis, henceforth MCA, was also used at a later stage to accommodate for the inclusion of more detailed modalities regarding the text mode. As its name indicates, MCA is an extension of CA. It takes as input tables with nominal categorical variables whereas CA takes as input tables with counts. The general principles of CA outlined above also apply to MCA (Greenacre & Blasius 2006; Le Roux 2010).
27The data were taken from the British National Corpus (World Edition), which consists of 100 467 090 words of spoken and written British English divided among over 4 054 texts. The spoken component contains approximately 10 million words, and consists of transcripts of spontaneous conversation and context-governed recording samples. The written component contains approximately 90 million words, and consists of samples of many kinds of text material (newspapers, fiction books, unpublished memoranda, etc.) The whole corpus spreads across the period from 1960 to the early 1990s (Burnard 2000).
28As opposed to most spoken corpora, the BNC is a relatively large corpus. Its size and its sampling scheme increase the reliability and validity of our observations. As opposed to other editions, the world edition of the BNC is freely and publicly available. It is annotated for parts of speech and extractions are done through an online query interface (Davies 2004). One major disadvantage is that the interface offers no way of exporting query results for further edition with a text editor or spreadsheet software. Exporting para- or extralinguistic information can only be done manually (e.g. via copy and paste), which is not a viable option when the construction under investigation has a high token frequency.
29For each of the eight alternating patterns listed in Table 1, all adjectival collocates were extracted from the BNC, amounting to 3,374 adjective types distributed across 28,790 construction tokens. To test whether the difference between the observed frequency and the expected frequency of each adjective in each construction was statistically significant, I performed an MDCA using Coll.analysis 3.2 (Gries 2007).
30Usually, lexical semanticists gather enough information from the top 10 or 20 collexemes and can safely ignore the rest of the output table (e.g. Hilpert 2008). That would be irrelevant here because the ten adjectives with the highest pbin and SumAbsDev values are distinctive of only one construction (Table 2). Therefore, the only information we can glean concerns the preference of <quite ADJ> for adjectives that denote modality in the broad sense – factual (clear, right, true), epistemic (possible, likely, sure) and dynamic (prepared, capable, easy) – and a positive mental disposition (happy). A more comprehensive way of inspecting the results is to focus on distinctiveness values for each of the 8 constructions (Table 3), but then the mass of information makes it hard to generalize over the table.
Table 2. Adjectives with the highest pbin and SumAbsDev values.
Table 3. Output of MDCA (sampled and sorted according to pbin value).
31If we focus on adverbs and disregard their syntax temporarily, some broad semantic tendencies emerge:
quite constructions co-occur significantly with adjectives that denote modality (sure, right, possible), dimension or position in space (big, large, substantial), proximity in time (new, recent), singularity (distinct), and difference (different, other);
rather constructions co-occur significantly with adjectives that denote oddities (unusual, eccentric) and retrograde leanings (conservative, formal, old);
quite and rather contrast significantly in the expression of value judgments; quite attracts adjectives with a positive connotation (happy, good, extraordinary, remarkable) whereas rather attracts adjectives with a negative connotation (tired, slow, limited, gloomy, negative, eccentric, grim, loud, conservative, low).
32The last point is particularly striking. By itself, surprised does not have a negative connotation. Once modified by rather we may assume that the cause for surprise is unpleasant:
(12) My earliest memory is of standing in the nursery of Byron House in Highgate and crying my head off. (…) I think my parents were rather surprised at my reaction, because I was their first child and they had been following child development textbooks that said that children ought to start making social relationships at two. (BNC-FYX)
- 4 These are only tendencies. The BNC contains examples where formal and loud have neutral interpretat (...)
33Presumably, the same kind of negative bias is at work with adjectives such as formal, loud, or low: 4
(13) Joseph seemed to be curiously affected by the game. I was convinced it was not a romantic passion. I had rather formal notions of desire. Boys fancied you when your make-up was just right and your freckles didn't show and your hair was tidy. (BNC-FU7)
(14) He was not short of that peculiarly British brand of slightly disdainful, rather loud self-confidence. (BNC-CS4)
(15) Even the unemployed place government policy rather low in a list of factors responsible for high unemployment levels. (BNC-B7G)
34On the one hand, MDCA reveals coarse-grained semantic tendencies that would be beyond the reach of contingency tables based on raw frequencies. On the other hand, the above interpretation is partial because it is based on a small number of collexemes and it disregards how their meanings correlate with the syntax of the intensifiers. Insofar as MDCA is not designed for the cross-comparison of variables, the structural limitations of Table 3 render further exploration uneasy. To assess the constructional near synonymy of quite and rather more deeply and more systematically, we need to (a) inspect more construction tokens, (b) describe the data with syntactic and semantic variables that cover several levels of granularity, and (c) determine those variables beforehand (i.e. not heuristically). If we do so, we face another problem: as we include more data and more variables, we need to build and inspect several tables at once. In such conditions, making cross-comparisons is tedious and unproductive. Therefore, rather than inspect and compare many collostruction-based frequency tables manually, I use correspondence analysis, a method that is well suited for the purpose of handling extensive datasets without compromising the complexity of the object of study.
35From Table 3, we can build a contingency table that confronts two variables – the eight quite/rather constructions from Table 1 and their distinctive adjectives – taking advantage of the fact that the table offers a way to filter away those adjectives that are below a certain distinctiveness threshold (i.e. pbin > 1.3, p > 0.05). However, studying the relationship between these two variables only is not particularly helpful for two reasons. First, my main objective is not just to see which adjectives pattern with which constructions. Rather, it is to determine to what extent the constructional profiles of quite and rather differ and to what extent they are similar. Second, the level of granularity overviewed in Table 3 is too high. If we keep only those adjectives whose absolute distinctiveness value is higher than 5 we are potentially looking at 4 344 construction tokens (i.e. 543 adjectives × 8 constructions). This means that we should sort the data using categories that are broad enough to capture the general profiles of quite and rather and reflect the semantic and syntactic properties of each pattern.
36To determine the semantic profile of a construction, I examined its distinctive adjectives, grouped by semantic classes. To determine the syntactic profile of a construction, I looked at whether the intensifier occurs in pre-adjectival or pre-determiner position and whether the adjective is predicative or attributive. Given that the profile of a construction may be sensitive to text mode, I also inspected variation according to the written vs. spoken distinction.
37The above remarks translate into a dataset, sampled in Table 4. It contains:
the raw frequencies of the 543 most distinctive adjectival collexemes (pbin > 5) of all 8 quite/rather constructions;
8 supplementary columns that correspond to the following variables: the syntactic position of the intensifier, the syntactic position/function of the adjective, the text mode, the sum total for each intensifier;
59 supplementary rows that group adjective frequencies according to semantic classes (positive value, negative value, atypicality, difference, etc.)
38CA uses these frequencies to display the rows and columns as points on a two-dimensional map. Interpreting the geometric positions of the points is a way of interpreting the similarities and differences between rows (i.e. adjectives), the similarities and differences between columns (i.e. quite and rather constructions), and the association between rows and columns. CA was performed in R (R Core Team 2015) with the CA function of the FactoMineR package (Husson et al. 2009).
Table 4. Input sample for CA (total = 9 632 cells)
In white: active rows (nrow = 543) and active columns (ncol = 8);in shades of grey: supplementary rows (nrow = 59) and supplementary columns (ncol = 8).
39The script outputs the following result:
**Results of the Correspondence Analysis (CA)**
Row variables have 543 categories, column variables have 8 categories
The chi-square of independence between the two variables is equal to 24522.65 (p-value = 0).
40χ2 has a very high value (24 522.65) and it is associated with the smallest possible p-value, implying that there is a relationship between the row variables and the column variables. Yet, we should be wary of using the magnitude of the χ2 value to quantify the effect of the correlation between variables, because this value depends on the sample size. The total sample size (43 216) is small relative to the number of active cells in the table (543 x 8 = 4 344). This means that the table contains many cells with small or null values. We are far from meeting the assumptions of the χ2 test, which stipulate that 80% of the sample size should be greater than 5, and the remaining 20% greater than 1. However, because CA is exploratory, it can be applied to tables even when the conditions of validity of the χ2 statistic are not met (Greenacre, 2007). Indeed, CA does not so much prove the existence of a relationship between columns and rows as the capacity of a limited set of data to make that relationship visible (Husson et al. 2011). Given that the p-value is 0, the significance of the deviation of the table from independence is undeniable. In other words, the choice of adjectives and the choice of active variables are globally interdependent in the dataset.
41Central to CA is the concept of profile. To obtain the profile of a row, each row frequency is divided by its row total. The average row profile is the average profile of all construction tokens put together. In a similar fashion, one obtains the profile of a column by dividing each column frequency by the column total. The average column profile is the average profile of all adjectives put together. Distances between profiles are measured with inertia (Φ2), i.e. “the weighted average of squared χ2-distances between the row profiles and their average profile (similarly, between the column profiles and their average)” (Greenacre, 2007: 32). More concretely, CA interprets inertia geometrically to assess how far row/column profiles are from their respective average profiles. The total inertia of a table is the χ2 statistic divided by the sample size. It is with inertia that CA measures how much variance there is in the table. In our dataset, Φ2 has a value of 1.13, which is relatively high. Therefore, we can expect data points to be more spread out on the map than if Φ2 were lower. This is because there are substantial differences between the data profiles.
42Figure 1 maps the first two dimensions of the data. Dimension 1 is represented by the horizontal axis, and dimension 2 by the vertical axis. Eight adjectives are plotted for illustrative purposes, as will appear below. The adjectives expensive and pleasant (in blue) are not associated with any particular construction and are very close to the average profiles (i.e. where the horizontal and vertical axes intersect). Given that <quite ADJ> is relatively close to the average column profile, it has a more representative profile than the other constructions.
43The syntactic profiles of quite have negative coordinates and cluster on the left-hand side of the plot, whereas the syntactic profiles of rather have positive coordinates and cluster on the right-hand side. This means that there is a clear divide between the syntactic profiles of quite and rather (in black) along the first dimension.
Figure 1. CA plot: a simultaneous representation of all active columns and 8 active rows.
44Dimension 2 shows that the patterns of quite and rather subdivide into three levels in parallel fashion (Table 5). This parallel becomes clearer if we abstract away from dimension 1 (i.e. the horizontal axis).
Table 5. Column clusters in dimension 2.
The relative distance to the average column profile is what determines each level (the higher the level index, the larger the distance from average).
45To some extent quite and rather follow a functional division of labor. For example, pleasant co-occurs with <quite ADJ>, whereas its antonym unpleasant co-occurs with <rather ADJ NP>, <DET rather ADJ NP>, and <rather ADJ>. The same three patterns attract odd, which connotes strangeness negatively, whereas <DET quite ADJ NP> and <quite ADJ> attract extraordinary, which connotes strangeness positively. Next, <DET quite ADJ NP> prefers costly over its less formal equivalent expensive, a sign that the choice of a syntactic pattern may be influenced by register. Little can be said about the proximity of <quite DET ADJ NP> and considerable besides a preference for this pattern to intensify adjectives that denote large proportions. If we projected more adjectives, we would see that considerable clusters with near-synonyms (compelling, massive, extensive, big, large, etc.), which would corroborate our claim.
46In CA, supplementary rows and columns can also be plotted to help interpret the active rows and columns. As opposed to active elements (white rows and columns in Table 4), supplementary elements do not contribute to the construction of the dimensions. They can still be positioned on the map, but only after Φ2 has been calculated with the active elements. Given the number of rows in our dataset, plotting all 543 adjectives would make little sense, as the myriad of points would clutter the graph. To make these data points invisible while still being able to inspect their general profiles, the supplementary rows were projected instead (i.e. the rows in grey in Table 4). These rows group the 543 adjectives into the 59 semantic classes described above. Each row cell is the sum of the frequencies of all construction tokens whose adjectives illustrate the semantic class. Because of their intermediate granularity, the supplementary rows are easier to interpret than the active rows.
478 supplementary columns (i.e. the columns in grey in Table 4) were also projected to: (a) see if the relationship between the syntax of the intensifiers and the syntax of the adjectives brings new insights on our dataset, (b) see if the mode (spoken vs. written) plays a role in the choice of a construction over another, and (c) summarize the general profiles of quite and rather to evaluate how close or distant they are. Figure 3 provides a superimposed representation of supplementary rows (in blue) and supplementary columns (in black).
48As expected, Figure 2 confirms most of the tendencies revealed in Figure 1. The position of the supplementary column variables on the first dimension corroborates the broad divide between quite and rather – quite constructions clustering to the left, and rather constructions to the right. Because quite constructions are closer to the average column profile than rather constructions, the profile of rather stands out in our dataset. As will appear below, the choice of rather over quite does not depend on the mode (since both spoken and written are close to the average column profile) but on the meaning of the intensified adjective.
Figure 2. CA plot: a simultaneous representation of supplementary columns and supplementary rows.
49Dimension 2 confirms the gradient of typicality inferred from Table 5. Typically, the intensifiers occur in pre-adjectival position, and the adjectives in the predicative position. Since quite and rather can only occur in pre-determiner position when the adjective is attributive, we have reasons to expect a strong correlation between these two variables. Such is not the case as the attributive variable is also seemingly attracted to the pre-adjectival position (bottom left part of the plot). When quite and rather occur in pre-determiner position, we find no significant proximity with any of the semantic classes of adjectives. The extreme position of the pre-determiner variable on the vertical axis makes this syntactic pattern stand out in our dataset. This is due to the overrepresentation of <quite DET ADJ NP> in the pre-determiner context (see Figure 1).
50Regarding the semantic properties of quite and rather constructions, three configurations emerge. In the first configuration, each intensifier has an area of semantic specialization. In the second configuration, some meanings can be intensified indiscriminately by quite or rather. In the third configuration, quite and rather divide up the task of intensifying two complementary aspects within a single conceptual domain.
51In the first configuration, quite intensifies adjectives that denote:
modal meanings (factual, dynamic, epistemic): correct, false, true, able, aware, certain, likely, etc.
adequacy and inadequacy: legitimate, acceptable, fair, unfair, inappropriate, unsatisfactory, unsuitable, etc.
cost (low or high): cheap, expensive, costly
simplicity and complexity: easy, simple, difficult, complex, etc.
typicality: normal, ordinary, usual, familiar, etc.
Rather intensifies adjectives that denote:
formality and informality: formal, strict, casual, informal, etc.
physical properties: ugly, fat, stout, fit, nasal, etc.
physical stimuli: quiet, loud, silent, sweet, bitter, etc.
meanings that are connoted negatively (stupidity, dullness, unclearness, unevenness, sterility, bad condition, excess, repulsion): dumb, dull, cloudy, unequal, barren, battered, excessive, gruesome, etc.
In the second configuration, quite and rather intensify the following meanings indiscriminately:
energy (connoted negatively): ferocious, violent
dimension/position: large, little, long, short, high, low, etc.
social/psychological properties (connoted negatively): desperate, shameful, busy, emotive, half-hearted, pessimistic, sad.
- 5 The position of the semantic class ‘superiority’ is marginal. This is an effect of the low frequenc (...)
52The above meanings correlate with the choice of an adjective in attributive position, especially in the patterns <quite ADJ NP>, <DET quite ADJ NP>, and <rather DET ADJ NP> (see Figure 2).5
53Table 6 illustrates the third configuration: when two connotations are available for a given conceptual domain, rather tends to select the negative alternative.
Table 6. The functional division of labor between quite and rather.
54All in all, the most typical profile in the dataset involves a construction in which:
quite occurs in pre-adjectival position;
the adjective occurs in predicative position;
the adjective denotes a meaning with either neutral or positive connotations.
55Incidentally, no striking relationship emerges between rather constructions and any of the syntactic properties of both intensifiers and adjectives. Despite its relative isolation, the attributive variable is not as atypical as the pre-determiner variable.
56Apart from showing that, contrary to popular belief, the pre-determiner pattern is not significantly associated with written language (and the formality traditionally associated with it), text mode does not seem to play any discriminating role in the light of the ‘spoken’ and ‘written’ variables. A more relevant option is to apprehend stylistic variation through more detailed modalities (fiction, news, conversation, etc.). Such information is available separately from the BYU-BNC for each query result, but given the number of tokens involved, manual extraction is not a viable option.
57One interesting feature of exploratory methods is that one can look at the same dataset from different angles to make more tendencies appear. To this aim, one can even ignore irrelevant variables or add some variables one believes will prove relevant. To see how the pre-adjectival vs. pre-determiner alternation behaves in the light of more detailed contextual information, I imported contextual tags of the XML edition of the British National Corpus and augmented the dataset with three levels of contextual information instead of one for each observation: text mode, text type, and information regarding each text.
58To remain within the strict limits of the pre-adjectival vs. pre-determiner alternation, I kept only those patterns that involved both alternants, namely <a(n) quite/rather ADJ NP> and <quite/rather a(n) ADJ NP>. Following the same logic, I ignored the cases where the adjective was in predicative position. Finally, the 6 variables corresponding to the remaining syntactic patterns listed in Table 1 were brought down to two levels (‘pre-adjectival’ and ‘pre-determiner’) within a unique variable (‘construction’).
59I obtained a dataset of 3 086 observations sampled in Table 7. This dataset consists of six categorical variables:
construction: 2 modalities (pre-determiner and pre-adjectival),
intensifier: 2 modalities (quite and rather),
text mode: 2 modalities (written and spoken),
text type: 8 modalities (academic writing, non-academic writing, fiction, news, other published writing, unpublished writing, conversation, and other spoken),
text information: 66 modalities (e.g. S parliament, S interview, W pop lore, W newsp brdsht nat: editorial, etc.),
semantic classes: 59 modalities (e.g. adequacy, accuracy_suitability, difficulty_complexity, etc.).
60The dataset was submitted to multiple correspondence analysis.
Table 7. A sample from the second extraction, with contextual information
61First, all the variables were treated as active except for ‘text_info’. The variable ‘sem_class’ was ignored. Figure 3 maps the first two dimensions of the data. The plot differs from what Figures 1 and 2 show. In dimensions 1 and 2, there is a strong correlation between (a) quite and the pre-determiner construction in the lower right corner, and (b) between rather and the pre-adjectival construction in the upper left corner on the other. Along the horizontal axis (dimension 1), pre-determiner quite is correlated with the ‘spoken’ modality of the ‘text_mode’ variable, and more specifically with the two modalities of the ‘text_type’ variable (i.e. other spoken and conversation). In contrast, along the same axis, pre-adjectival rather is correlated with the ‘written’ modality of ‘text_mode’ and all the modalies of ‘text_type’ (i.e. academic writing, non-academic writing, fiction, news, other published writing, and unpublished writing). We should be wary of concluding that pre-determiner quite prefers spoken contexts and pre-adjectival quite prefers written contexts. Indeed, along the vertical axis (dimension 2), pre-determiner quite is found in one spoken context (conversation) and also four written contexts (other published writing, fiction, unpublished writing, and news). Along the same axis, pre-adjectival rather is found in one written contexts (academic writing) and also one spoken context (other spoken). Non-academic writing is specific to neither pre-determiner quite nor pre-adjectival rather.
Figure 3. MCA plot : a simultaneous representation of pre-adjectival vs. pre-determiner syntax (active, in black), intensifiers (active, in red), text modes (active, in green), text types (active, in blue), and text information (illustrative, in cyan)
62Considering the respective contributions of dimensions 1 and 2 to inertia, it appears that, within the strict limits of the pre-determiner vs. pre-adjectival alternation, quite occurs preferentially in pre-determiner position. This pattern tends to appear both in the written and spoken components of the corpus, with a preference for the latter. It is found in all text types except academic writing. Rather occurs preferentially in pre-adjectival position. This pattern tends to appear both in the written and spoken components of the corpus, with a preference for the former. It is found in all text types except conversation. Once again, we find that the pre-determiner construction is not exclusively associated with written language in the BNC.
63Next, all the columns were treated as active except for the variable ‘sem_class’. The variable ‘text_info’ was ignored. Figure 4 maps the first two dimensions of the data. Along the horizontal axis, pre-determiner quite is correlated with positive connotations (‘value_desirable’, ‘dimension_position’, ‘importance’, ‘psych_stim_good’, ‘adequacy_suitability’, ‘physical_property_good’), with the exception of ‘cost_high’. As expected, pre-adjectival rather is correlated with adjectives having negative connotations: ‘atypicality_odd’, ‘dullness’, ‘unclearness’, ‘tension’, ‘repulsion’, ‘condition_bad’, ‘discomfort’, ‘luck_bad’, ‘psych_stim_bad’, etc. However, along the same axis, rather also co-occurs with adjectives whose meanings are neutral or even positive such as ‘difference_contrast’, ‘atypicality_extraordinary’, ‘singularity’, ‘epistemic’, or ‘physical_property_neutral’. Along the vertical axis, quite co-occurs with adjectives that have positive connotations: ‘value_desirable’ and ‘physical_property_good’. It also co-occurs with adjectives that denote neutral values such as spatial properties (‘dimension_position’) or social or psychological properties (‘soc_psych_prop_neutral’). Interestingly, in one context (age), quite is found with adjectives denoting opposite properties: ‘age_young’ and ‘age_old’. Along the same axis, rather co-occurs exclusively with adjectives that have negative connotations: ‘unclearness’, ‘atypicality_odd’, ‘dullness’, ‘value_undesirable’, ‘unevenness’, ‘epistemic’, ‘discomfort’, ‘singularity’, ‘physical_property_bad’, ‘repulsion’, and ‘luck_bad’. As seen above (see Table 6), when two connotations are available for a given conceptual domain, rather tends to select the negative alternative.
Figure 4. MCA plot : a simultaneous representation of pre-adjectival vs. pre-determiner syntax (active, in black), intensifiers (active, in red), text modes (active, in green), text types (active, in blue), and semantic classes (illustrative, in cyan)
64Notwithstanding exceptions, the division of labor outlined above still holds. For example, pre-determiner quite is found with positively connoted modalities (e.g. ‘value_desirable’, ‘psych_stim_good’, or ‘physical_property_good’) whereas pre-adjectival rather is found with their negative counterparts (‘value_undesirable’, ‘psych_stim_bad’, ‘physical_property_bad’). Admittedly, not all the oppositions listed in Table 6 are found in the first two dimensions of the MCA plot in Figure 4 (for example, the ‘luck_good’ vs. ‘luck_bad’ is absent from this second dataset). It suggests that the behavior of the alternation differs slightly from the more general behavior of quite and rather.
65Coming back to Allerton’s claim, we may concede that the choice of pre-determiner position over the default pre-adjectival position is sensitive to context or register. Yet, this choice is indeed more than a matter of style or formality both within the strict frame of the pre-determiner vs. pre-adjectival alternation, and beyond, i.e. in the broader frame of adjectival intensification involving quite and rather.
66I set out to capture usage-based relations between quite and rather from the perspective of their near synonymy. I argued that linguists should be wary of reducing the study of near synonymy to a matter of choices between lexemes only because I assumed that just as lexical choices reflect construals, so do their constructional expressions.
67In the plots generated with CA and MCA, the differences between quite and rather are summarized graphically by their respective projections on the map. Affinities with syntactic and semantic profiles can be assessed spatially. I believe the methods presented above tease out successfully the subtle mix of similarities and differences that holds between two near synonyms.
68Because CA and MCA are based simultaneously on collocation data and formal and semantic variables, using them is a clear improvement upon previous lexeme-based collocation studies on intensifiers that treat form, meaning, and context separately. Under the assumption that syntactic structure is an index of conceptual structure, the study of quite and rather has more to gain from correlating form and meaning than from relying on large, separate tables, each table illustrating one variable.
69On the one hand, given that natural languages avoid true synonymy (Cruse 1986: 270), we should not be too surprised to see that quite and rather differ in some respect (cf. the first configuration in the previous section):
(16) It has been said that ‘the problem of turbulence has been solved’. Aside from the question of what is ‘the problem of turbulence’ (there are many), this can give a quite/?rather false impression. (BNC-J12)
(17) About 8 p.m. that evening Taff and I sat at the side of our slit trench eating what I thought was a rather/?quite dubious lump of meat. Taff said it was lamb; (BNC-A61)
70On the other hand, given that speakers generalize over recurring experiences of language use, similar linguistic information can be stored at multiple levels in the inventory of linguistic signs. Distinct linguistic units that are not necessarily synonyms may thus share identical semantic content. Therefore, that quite and rather share substantial common ground should come as no surprise either (cf. the second configuration):
(18) (…) taboos persist and elicit sometimes quite violent reactions if they are broken (…) (BNC-ECY)
(19) ‘Often,’ he notes, ‘we resembled a rather violent community welfare body rather than a group of revolutionaries.’ (BNC-APP)
71The most remarkable result to emerge from the data is that when quite and rather operate within the same conceptual domain, they often impose complementary focal adjustments on the representation that the conceptual domain evokes, following what I called a division of labor (third configuration, Table 6). When adjectives have a negative connotation, they tend to occur in rather constructions:
(20) Three-tiered walls and arcades of massive pointed arches soar upwards in a quite breathtaking fashion. (BNC-B3K)
(21) They met towards the end of January, and he kept the news from Theo until April. When he did let it out, he did so in a rather peculiar fashion, linking it to a quarrel with Mauve and casting it in a dramatic mode, with himself in the first and then third person. (BNC-CBN)
72In contrast to confirmatory statistical models, which provide accuracy scores of explanatory power, CA and MCA are exploratory: these methods merely tell us that the distributions in our dataset are statistically significant and not due to chance. Therefore, we should not infer any causal link between the variables.
73Finally, one anonymous reviewer considers that quite and rather are genuine intensifiers only when these adverbs express high degree, as opposed to when they express a moderate degree, as in well, she's quite beautiful... but not quite (the reviewed provided that example). The same reviewer wishes a distinction had been made between these two degrees. Following the literature on the topic, which already provides a response to this critique, I consider that quite and rather are intensifiers, or “degree words” (Bolinger 1972), as long as they index an adjectival property on a scale, regardless of the specification of degree. When quite and rather set the qualities that gradable adjectives denote to a moderate level, they function as ‘compromisers’ (Paradis 1994), ‘moderators’ (Paradis 1997), or ‘downtoners’ (Nevalainen & Rissanen 2002), along with adverbs such as fairly, mildly, moderately, partially, pretty, relatively, etc. When quite and rather set gradable adjectival qualities to a high level, along with greatly, really, so, strongly, too, very, etc., they function as ‘boosters’ (Paradis 1997). Determining whether degree modifiers are moderators, boosters, or maximizers is indeed relevant, but it cannot be done using only corpus-based techniques, mainly because semantic context is not always helpful. This has to be done experimentally, as exemplified by Paradis (1994, 1997), who interprets <intensifier + ADJ> collocations in combination with intonation patterns in the prosodically annotated London-Lund Corpus of Spoken English.
74Notwithstanding the above caveats, and pending confirmatory statistics and experimental verification, I believe that my approach contributes some new understanding of intensifiers (within the frame of the pre-determiner vs. pre-adjectival alternation and beyond), constructions, and collocation analysis. Once distinctive collexemes have been identified, the data can be used again safely in multi-way tables to map correlations between intensifiers, adjectives, and information concerning context of use and constructional idiosyncrasies.
75I have combined statistical methods to compare the constructional profiles of quite and rather, taking into account both semantic and syntactic factors. First, I asserted the need for better statistics in the collocation-based study of intensifiers. Regardless of the specific quantitative method that one adopts, it is particularly important to filter away co-occurring pairs that are unrealistically too frequent or too rare. It is also important to compute collocation strengths with statistics that do not violate the distributional assumptions that are specific to language data. Next, I augmented the collocation data with syntactic and semantic variables that license the use of quite and rather in context. Finally, I used two multifactorial methods to explore and view the same dataset from two different angles: one focusing on the differences between quite and rather, and the other on the pre-determiner vs. pre-adjectival alternation. Each time, similarities and differences were synthesized in one graph.
76This has led me to assess the near synonymy of not just two lexemes (quite and rather), but a network of construction types and subtypes. All assume different profiles depending on what conceptual domains (or what parts of a specific conceptual domain) intensification bears on. Through the choice of a syntactic profile, the speaker adopts a focal adjustment on the meaning of an adjectival property and influences the way that the hearer interprets this conceptual representation in context.
77Based on a combination of MDCA and CA, results show that quite and rather constructions are near-synonyms. On the one hand, both subdivide into three syntactic levels following a gradient of prototypicality, from pre-adjectival patterns to pre-determiner patterns. Both also share enough semantic properties to intensify adjectives with similar meanings (e.g. negatively connoted energy and social/psychological properties, and dimension/position). On the other hand, quite and rather constructions are distinct enough to divide up the task of intensifying certain categories of adjectival meanings. The most striking tendency is that quite constructions have a preference for the intensification of positive meanings (e.g. quick, harmless, lucky, etc.), and rather constructions have a preference for the intensification of negative meanings (e.g. slow, threatening, unfortunate, etc.).
78Based on a combination of MDCA and MCA, results also show that quite displays a neat preference for the pre-determiner pattern, and rather for the pre-adjectival pattern in the strict frame of the pre-determiner vs. pre-adjectival alternation. Finally, although relevant to some extent, context does not fully account for the pre-determiner vs. pre-adjectival alternation. The abovementioned division of labor plays a central role in the choice of one alternant over the other.
79The approach described in this paper can be extended to other intensifiers and other forms of intensification in English. It can also be extended to any linguistic paradigm, providing one wants to show how the paradigm is structured. This is immediately relevant to those linguists who subscribe to an inventory approach to the mental representation of grammar, as formulated in Goldberg (2006, 2009) and Cognitive Construction Grammar in general (Langacker 2009). Cognitive Construction Grammar posits that linguistic knowledge consists of a network of families of constructions, and advocates a usage-based model to explain how constructions are stored in the taxonomies. This paper should therefore be seen as a quantitative corpus-based contribution to a usage-based description of constructional networks.