A few or several? Construal, quantity, and argumentativity

Abstract

This study examines two seemingly similar quantifiers, a few and several, and argues that the differences between them go beyond the (slightly) different quantities they each denote. Specifically, we argue that several construes its nominal complement as composed of individuated entities, which renders them more prominent, and thus a stronger basis in support of a conclusion the speaker is arguing for. We base our analysis on two experiments and a corpus study. The experiments show that there is indeed an argumentative difference between the quantifiers, and the corpus study points to the discourse factors behind it. In comparison with a few, several is associated with a higher discourse prominence for its complement (greater individuation, significance) and with greater argumentative strength. Based on this data, we characterize the quantifiers’ prototypical discourse profiles. A typical instance of several occurs in persuasive genres, refers to a not-so-small quantity, construes the plural entity as composed of individuated entities, and contributes to a strong argument. A typical instance of a few occurs in non-persuasive genres, denotes a small quantity, construes the entities composing the plural entity as un-individuated, and contributes to a weak or neutral argument.

1. Introduction

It has long been recognized that quantifiers such as all and many affect discourse beyond the contribution they make through the quantities they denote (Moxey & Sanford, 2000). Thus, the sentence many people came to the party produces a particular rhetorical effect because the cardinality of the set of people who came to the party is larger than that of some reference set (e.g., the people expected to attend or those who typically come to this type of party). On this view, (1a) is more convincing than (1b) because several expresses a larger quantity than a few:

Indeed, prior research has found that several is associated with slightly higher quantities than a few (Newstead & Coventry, 2000; Newstead et al., 1987; van Tiel et al., 2021 inter alia). Nonetheless, we here argue that focusing on the quantities expressed by the quantifiers portrays only a partial picture of the discourse functions served by the quantifiers. While the larger quantity associated with several may render it argumentatively stronger than a few, we propose that other factors are involved. Specifically, the two quantifiers differently construe the entities they introduce, such that several construes its nominal complement as composed of individuated entities. Individuated entities are perceived as more prominent. Interlocutors pay more attention to prominent entities; therefore, they are perceived as important and carry heavier argumentative weight. This study demonstrates how cognitive constructs, such as quantity, focus, and individuation, are related to discourse.

We focus on a few and several because the two share a similar core which delimits a smallish quantity larger than one. Our first goal is to test our intuition that several has a stronger argumentative effect than a few, one which goes beyond the quantity it denotes. Our second goal is to show how this difference manifests itself in the distribution of the two quantifiers. Our corpus study explores the distributional patterns of the two quantifiers, as they are reflected in their association with specific adverbs, adjectives, and nouns. The combined results show that, indeed, quantity alone cannot account for (1) the inferential patterns licensed by the two quantifiers, nor for (2) their corpus distribution.

In the cognitive-functionalist tradition, meanings are represented as rich conceptual structures, often including detailed information about how language users experience the world. The current study follows this tradition by showing that even expressions that convey schematic relations, such as quantifiers, reflect our models of the world.

Section 2 introduces the theoretical frameworks that inform this study. Sections 3 and 4 provide the empirical basis for our analysis – two experiments and a corpus study, respectively. We discuss the picture that emerges from the combined results in Section 5 and conclude with Section 6.

2. Preliminaries

2.1. Theoretical background

We adopt a usage-based approach to language (Barlow & Kemmer, 2010; Bybee & Beckner, 2010). According to this approach, language use and language structure shape each other in a feedback loop, which is why meaning and distribution are closely related. Corpus analysis can then test such correlations.

We here focus on the two quantifiers as they figure in the left slot of the English nominal. It is possible to represent our objects of study as two partially specified constructions (Croft, 2001; Goldberg, 1995) in (2), or micro-constructions (Traugott, 2008):

A central concept in this study is argumentativity in language, the theory developed by Anscombre and Ducrot (1976, 1983). To illustrate, consider (3a) and (3b):

(3a) as a whole most likely supports the conclusion that I will not clean my house, while (3b) most likely supports the opposite conclusion. According to Anscombre and Ducrot, this is because but coordinations introduce two arguments in support of opposite conclusions, such that the hearer is invited to infer that the speaker is leaning toward the conclusion supported by the second argument. Anscombre and Ducrot’s theory focuses on how linguistic units license inferential processes, as a fundamental part of everyday interactions.

Argumentativity can be conceptualized as an intersubjective negotiation over modification of the common ground (Verhagen, 2005, 2015). In cognitive-functional frameworks, linguistic meaning is viewed primarily as evoking and modifying conceptual structures, rather than referring to actual objects and states of affairs (Croft & Cruse, 2004; Langacker, 1987, 1991, 2008; Talmy, 2000). Linguistic structures are seen as instructions issued by the speaker to the addressee (Harder, 1996; Langacker, 2008, p. 460), for example, to conceptualize an object of conception in a certain way. If so, we can consider argumentative direction as an instruction or cue to accept or reject certain inferences as valid in the common ground – the knowledge, beliefs, and attitudes shared by the speech event participants (Clark, 1996).

Argumentative strength is defined as follows:

Thus, saying that several is argumentatively stronger than a few means that for two sentences that differ only in whether they contain a few or several, an addressee who accepts a conclusion based on the sentence with a few will also accept a conclusion that relies on the sentence with several, but not necessarily vice versa.

When comparing argumentative strength, we need to consider argumentative orientation. Consider the following examples in which the second clause provides evidence for the conclusions stated in the first:

One can argue that a few is argumentatively stronger than several, because (5a) is more convincing than (5c). Nonetheless, the reason for this argumentative difference is that a few introduces a weaker counterargument to the conclusion that the speaker is trying to establish that the exam was easy. At the same time, (5c) is less convincing overall, because several introduces a stronger counterargument to that same conclusion. The difference between (5a) and (5b) further illustrates the role of argumentative orientation by contrasting a few, which carries a positive argumentative orientation, with few, which carries a negative argumentative orientation. That is, it directly contributes to the conclusion opposite the one supported by a few, making the overall sentence in (5b) more convincing. ¹

Crucially, however, the difference between (5a) and (5c) is not due to a quantity difference alone. On the view that language is primarily an interactional phenomenon, utterances are rarely isolated events. They are rooted in discourse, comprised of chains of usage events linked by inferential steps. Speakers design their utterances so that their addressees can readily identify what is to be added to that common ground. Crucially, the addressee need not conceptualize the object of joint attention in a fully detailed manner. Instead, they need to understand which inferences are licensed by the speaker (see Pander Maat, 2006, for a similar analysis for gradable adjectives). Argumentativity is then a crucial cue.

2.2. Quantifiers and their meaning

The view that quantifiers express relations between sets is captured by Barwise and Cooper’s (1981) theory of generalized quantifiers. Some simplified examples for the meaning of quantifiers according to this theory are listed in (6).

There are several problems with this account. First, seemingly synonymous quantifiers, such as most and more than half, exhibit differences in the inferential patterns they license and in the way they are used in discourse (Ariel, 2004; Hackl, 2009). Moreover, the theory cannot account for data showing that “synonymous” quantifiers express different quantities, in a distributional pattern reflecting prototype effects (van Tiel et al., 2021).

Langacker (1991, 2016) offers a characterization of quantifiers from a cognitive linguistic perspective. In line with Milsark (1976), Langacker divides quantifiers into relative quantifiers, such as all and most, and absolute quantifiers. A few and several are of the latter type. While the primary function of relative quantifiers is grounding some subset (with respect to the whole set it is a part of), absolute quantifiers are adjectival. That is, they profile a relationship in which their complement maps onto a portion of a measurement scale, much like adjectives such as small or pretty. However, when in initial position, absolute quantifiers too function as grounding elements (Langacker, 2016). As such, they instruct the interlocutor to direct their attention to an intended discourse referent (Langacker, 2008, ch. 8).

Absolute quantifiers are semantically characterized along four parameters: the quantified mass (plural, in this case), the measurement scale (quantized or continuous), whether the quantity is assessed in reference to the origin of the scale or the scalar norm, and whether the assessment is made by scanning the scale in a positive or negative direction (Langacker, 2016). A few and several differ only with respect to one of these parameters – the type of scale. Several evokes a quantized scale, while a few evokes a continuous scale. Langacker does not elaborate on how this difference manifests itself in their use.

As for quantity, many factors may affect the specific quantities or proportions denoted by a quantifier, including the nature of the counted entity and the size of the reference set (Newstead et al., 1987). In these studies, a few is associated with smaller values than several. It is, therefore, tempting to conclude that this quantitative difference is responsible for all the differences between the two quantifiers. Nevertheless, there are several problems with this conclusion. First, there is a significant referential overlap between the two quantifiers. That is, they can often be used to describe the same quantities. Second, there are differences in the behavior of the two quantifiers when they are used in a context in which there is a salient background set. ² Consider the difference in felicity between (7a) and (7b):

Several sounds odd when associated with very large numbers (U.S. voters), while a few doesn’t. This is surprising under the view that reduces the differences between the quantifiers to the set size denoted by the quantified NP, since it is several, rather than a few, that is supposed to denote a larger quantity.

Third, as this study argues, quantity by itself is not enough to predict the inferential patterns licensed by the quantifiers and their distribution in the corpus. As argued in the previous subsection, the communicative intention of a speaker using a vague quantifier is not necessarily to (only) communicate a certain quantity. Facilitation of certain argumentative inferences is at least sometimes equally important. As Nouwen (2010, p. 240) argues, “Elimination of vagueness from a term, strips this term from some of its key communicative functions.” While quantity plays a role in facilitating these inferential patterns, we identify other factors as well: individuation and argumentative significance.

3. Experimental studies

We conducted the first experiment to test whether there is an argumentative difference between several and a few and, furthermore, whether this difference should be attributed to factors beyond the quantity expressed by the quantifiers (at least sometimes). The second experiment builds on some of the results of the first and aims to see whether several and a few patterns with purely quantitative items (i.e., numerals) with respect to dialog coherence.

Participants of both experiments were native speakers of American English, raised as monolinguals, with no diagnosis of dyslexia, autism, or mild cognitive impairment. The participants were recruited over the commercial platform Prolific (www.prolific.co.uk), and they provided informed consent before participation in the study.

Analyses were carried out using R (R Core Team, 2021). Regression analyses were conducted using the packages lme4, lmerTest, car, MuMin, emmeans, and ordinal (Barton, 2022; Bates et al., 2015; Fox & Weisberg, 2019; Kuznetsova et al., 2017). The data and the scripts used for the analyses are available at https://osf.io/kxaht/?view_only=00503082a44b4207bb505bd04f658bc2.

3.1. Experiment 1

Participants read sentences that reviewed products. All the target items were negative reviews, mentioning specific problems with the product. The quantifier specifying the number of problems was varied (e.g., the laptop arrived with [a few / several] dents). Participants were then asked to assess the writer’s satisfaction with the product. We assumed that if several is argumentatively stronger than a few, products reviewed with sentences containing several will receive lower ratings than those with a few. Next, in order to see whether raw quantities can account for argumentative strength, participants were also asked to assess the quantity expressed by the quantified NP for each item. The results from both tasks were taken into account in the statistical analysis.

3.1.1. Participants

Ninety-three adults participated in the study (47 females, mean age = 42.7, range = 20–75, SD = 14.5). Six participants were excluded due to failure to follow the experiment’s instructions, and five others because their answers to the debriefing question at the end of the study indicated they noticed experimental manipulation.

3.1.2. Materials

The experimental items were comprised of four sentence types, each containing one of two quantified nominals (a total of eight items). Each sentence described an item with a problem. The items were divided into two lists with four items each, such that each participant saw two of the sentences with several and two with a few. Items appeared in a pseudo-random order. Each participant saw one sentence from each type (only one nominal per type).

The experimental items and names used for coding the results are in Table 1.

Table 1. Items used in Experiment 1.

Figure 1. An experimental item and a CS evaluation task.

3.1.3. Task and procedure

Participants were told that they were going to read sentences taken from product reviews and then answer questions about them. For each experimental item, participants were asked to perform two tasks. The first was a customer satisfaction (CS) evaluation. Participants were presented with a sentence and asked, “Based on this review, how satisfied is the customer with the product they purchased?” To answer, participants marked their assessment with a ticker on a scale from “very satisfied” to “very unsatisfied.” The second task was to assess the quantity denoted by the quantified NP. Once again, participants were shown the sentence and asked, “Based on this review, how many dents did the laptop have?” Here they had to type their answer. Each occurrence of an item was separated by two or three filler items. The quantity assessment task appeared only after the participants completed all four CS evaluations. This simulated a more natural understanding of quantifiers in discourse and ensured that participants did not consciously think about quantities when performing the CS evaluation just because of the experimental manipulation.

We included 29 filler items, for which participants performed one of four tasks. The first task was a CS evaluation, identical to the one in the experimental task. In the second, participants were asked to mark on a scale how likely it is that the customer would recommend the product to friends or family. In the third, participants were asked how much they thought the customer paid for the product, and they needed to respond by typing in a value. In the fourth task, used as catch trials, participants were asked about some quantity mentioned in the sentence. All filler items were taken from actual product reviews on Amazon. Some sentences presented a positive attitude towards the product, some negative, and some mixed. Four sentences were repeated twice with different tasks.

3.1.4. Results and analysis

Items with a few received a higher average CS score (M = 30.79, SD = 19.17) than items with several (M = 18.9, SD = 16.32). Fig. 2 presents the means and interquartile ranges of the CS scores per quantifier. In all boxplots used in this article, the boxes indicate the interquartile ranges, the vertical lines are the medians, and the diamond indicates the mean. The whiskers extend to the interquartile range × 1.5 in each direction or the minimum/maximum values.

Figure 2. Boxplot representing the distribution of CS scores per quantifier. The boxes indicate the interquartile ranges, the vertical lines are the medians, and the diamond indicates the mean. The whiskers extend to the interquartile range × 1.5 in each direction or the minimum/maximum values.

Fig. 3 presents the distribution of CS scores grouped by items. The pattern in which items with several received lower CS scores than those with a few is repeated in every item. The means, medians, and SDs for each item are listed in Table 2.

Figure 3. Boxplot representing the distribution of CS scores grouped by items. The boxes indicate the interquartile ranges, the vertical lines are the medians, and the diamond indicates the mean. The whiskers extend to the interquartile range × 1.5 in each direction or the minimum/maximum values.

Table 2. Means, SDs, and medians of CS scores grouped by item.

As for quantity, items with a few received a mean quantity assessment of 3.71 (SD = 1.95), while items with several received a mean score of 5.47 (SD = 3.46). This finding is significant (t(293.34) = 6.059, p < .001) and replicates previous research on quantifiers and quantities. Fig. 4 presents the means and interquartile ranges of the assessed quantity per quantifier.

Figure 4. Boxplot representing the distribution of assessed quantity per quantifier. The boxes indicate the interquartile ranges, the vertical lines are the medians, and the diamond indicates the mean. The whiskers extend to the interquartile range × 1.5 in each direction or the minimum/maximum values.

3.1.4.1. Regression analysis

We fitted two mixed-effects linear models to see whether the quantifier has an effect beyond the quantity associated with it. The first is a model with fixed effects for z-scored quantity and random intercepts for participants and items. The second is a model with fixed effects for z-scored quantity and quantifier (factors were sum-coded). Both models included random intercepts for participants and items. We did not include interactions as we do not have a hypothesis regarding their interpretation. Variance inflation factors for the second model indicated no problems with collinearity. To compare the two models, a likelihood ratio test was performed using the function anova(), which revealed a significant difference between models (χ²(2) = 40.45, p < .001).

The mixed-effects model coefficients for the second model are presented in Table 3. We see that an increase in quantity has a negative impact on the CS score; when the assessed quantity of an item is higher, it is more likely to receive a lower CS score. Additionally, the choice of several as a quantifier also leads to lower CS scores.

Table 3. Linear mixed-effects model. Formula: CS score ~ quantity + quantifier + (1|participant) + (1|item).

R² marginal = 0.12, R² conditional = 0.44.

3.1.5. Discussion

Results from the experimental study indicate that several is indeed argumentatively stronger than a few – negative reviews based on several “problems” received lower CS scores than those based on a few. This means that participants interpreted the negative traits denoted by the quantifier’s complement as having more weight in their estimation of the speaker’s satisfaction with the products. Interestingly, however, while quantity was indeed a significant predictor, adding the quantifier as an additional predictor turned out to be significant and increased the variance predicted by the model. In other words, the difference in quantities is not sufficient to account for the difference in argumentative strength. Factors other than quantity must have influenced the participants’ evaluation of the reviews. The experimental results, however, do not tell us what these factors are.

3.2. Experiment 2

Number words lack argumentative orientation (Anscombre & Ducrot, 1989). ³ They can then serve as a baseline for assessing the argumentative work performed by quantifiers in discourse. We used the average quantity estimates from the previous study (four and six) to see how a few and several compare to their numerical counterparts. By comparing the behavior of quantifiers to that of numbers, we can learn whether the difference between several and a few is primarily quantitative or not. As a second point of reference, we sought out two quantifiers that express the same quantities. If such a pair exists, it would be reasonable to assume that its members will differ on another, non-denotational dimension. We chose a pair of multal quantifiers, a lot and many, which, according to previous studies, are associated with roughly the same quantities (van Tiel et al., 2021). Although the evidence for the quantitative equivalence is not very robust, we nevertheless decided to include the multal pair, at least as an approximation to a referentially equivalent pair of quantifiers.

3.2.1. Participants

Four hundred and eighty participants (240 females, mean age = 39.56, range = 18–80, SD = 13.58) took part in the study.

3.2.2. Materials

The experimental items were comprised of three pairs of quantity expressions paired with two different plural nouns, problems and books (a total of six items). Items appeared in dialogs. In each dialog, the first interlocutor predicates some quantity, and the second interlocutor replies with another quantity expression. Each item appeared in one of two conditions – negation or affirmation. In “negation” (8), the second interlocutor prefaces their response with no, while in “affirmation” (9), the response is prefaced with yes.

In addition, the order of the members of each pair alternated, such that half of the times one member appeared in A’s utterance and the second in B’s, and half of the times the order was reversed. We used a single-trial experimental design, in which each participant sees only one item in one condition.

3.2.3. Task and procedure

Participants had to rate the coherence of the dialogs on a seven-point Likert scale (1 – not coherent at all, 7 – totally coherent).

3.2.4. Predictions

We predicted that dialogs involving numbers would receive higher coherence ratings in “negation” than in “affirmation,” because four and six are clearly different in quantity, but carry no argumentative force. Thus, the negation can easily target the quantitative difference, while the affirmation is left with no target. We expected that the two multal quantifiers, many and a lot, would show the opposite pattern, with higher ratings in “affirmation” than in “negation,” inasmuch as they express the same quantity. Here, the affirmative response confirms quantitative information, while the negation has no argument (but see the discussion in Subsection 3.2.6). If a few and several are only different in quantity, they should behave like the number dialogs and favor “negation.” However, if they stand for roughly the same quantity, they should behave more like denotationally equivalent quantifiers (a lot and many, presumably) and favor “affirmation.”

3.2.5. Results and analysis

The results of the study are summarized in Table 4 and Fig. 5. Many and a lot were acceptable in the affirmation condition (M = 6.19, SD = 1.11), but much less so in the negation condition (M = 4.08, SD = 1.83). For the numbers, the coherence ratings were very low in the affirmation condition (M = 2.01, SD = 1.35) and very high in the negation condition (M = 6.23, SD = 1.35). Finally, for a few and several, the coherence ratings were similar in the affirmation (M = 4.78, SD = 1.45) and negation (M = 5.32, SD = 1.72) conditions, indicating that participants found the dialogs generally coherent, although not maximally so.

Table 4. Means, SDs, and medians of coherence ratings.

Figure 5. Coherence ratings by quantity expressions and condition. The boxes indicate the interquartile ranges, the vertical lines are the medians, and the diamond indicates the mean. The whiskers extend to the interquartile range × 1.5 in each direction or the minimum/maximum values.

A cumulative link model was fitted with coherence rating as a response variable and quantifiers, condition, and an interaction term between them as call variables. Factors were sum-coded. The model’s coefficients are presented in Table 5. The model revealed a significant effect of quantifiers on coherence rating, with a few and several leading to higher ratings than the grand mean (estimate = .59, SE = .12, p < .001) and four and six leading to lower ratings (estimate = −.8, SE = .13, p < .001). Additionally, the model showed a significant effect of condition, with affirmation leading to lower ratings (estimate = −.36, SE = .09, p < .001). The interaction between quantifiers and condition also had a significant effect, with a few and several in affirmation resulting in higher scores (estimate = .59, SE = .12, p < .001) and four and six in affirmation resulting in lower scores (estimate = −2.1, SE = .15, p < .001).

Table 5. Cumulative link model.

Formula: Rating ~ quantifiers × condition. Log-likelihood = −719.91; AIC = 1461.82.

Post-hoc pairwise comparisons with Bonferroni correction showed that in the affirmation condition, four and six received significantly lower coherence ratings compared to the negation condition (p < .001). Conversely, a lot and many received significantly higher coherence ratings in the affirmation condition than in the negation condition (p < .001). The difference was not significant for a few and several.

3.2.6. Discussion

The results of the study showed that, as predicted, the number dialogs were accepted in the negation condition and rejected in the affirmation condition. Additionally, the “yes” responses for the multal quantifiers received significantly higher ratings than the “no” responses, which is in line with the predicted patterns for referentially equivalent expressions. However, the “no” responses for the multal quantifiers were not completely rejected but received a low acceptability rate. We consider two possible reasons that can account for this. The first is that while the referential difference between the two quantifiers is small, it is nevertheless large or salient enough for some participants to be targeted by negation. The second interpretation is that there is some other difference between the two quantifiers, such as stance (with the newer a lot being more subjective than the older many), although the exact nature of this difference is unclear at this point. We find the second explanation more appealing in light of the high ratings in “affirmation,” which are unexpected under a salient quantitative difference.

Notably, a few and several did not pattern exactly the same as either pair of quantity expressions. However, the differences between the mean scores in both conditions indicated that a few and several are more similar to the multal quantifiers: the difference between the mean ratings is .46 for a few and several, 2.11 for many and a lot, and −4.22 for the number pair. If the reason for the response pattern of the multal quantifiers is that they mainly differ on a non-objective level, this could indicate that the primary difference between a few and several is not the quantity difference, but rather the non-objective (argumentative) difference.

Crucially, the fact that several and a few are so different from the numbers supports our argument that quantity alone cannot fully explain the difference between these two quantifiers. However, it is still unclear what other factors may contribute to this difference. We turned to corpus data in order to find out.

4. Corpus study

4.1. Motivation

As previously mentioned, the meanings of expressions both shape and are shaped by the contexts they occur in. We therefore hypothesized that discourse distribution might reflect different argumentative strengths for the two quantifiers as well as other meaning features possibly contributing to this difference. To test this hypothesis, we examined the lexemes and constructions that combine with the quantifiers to form the nominal, including the plural nouns serving as the nominal head, the adjectives modifying them, and other adverbs and constructions modifying the entire nominal. If the only difference between the quantifiers is in the quantities they communicate, we would expect that (1) there would be only few lexemes that prefer one quantifier over the other, and (2) any differences in preference would be explained by appealing to quantities. We also examined the genres associated with the quantifiers to determine whether register can explain the preferred patterns of association for the quantifiers.

This type of analysis allows us to utilize large datasets, including tens or even hundreds of thousands of tokens (the size of the dataset is discussed in the next subsection). It also has the advantage of limiting the degrees of freedom for researchers, leading to results that are easier to evaluate.

4.2. Data collection

Data were extracted from the Corpus of Contemporary American English (COCA) (Davies, 2008–). First, all sentences with the quantifiers were extracted. The sentences were then parsed using the spaCy library for Python (Honnibal et al., 2020), with the model “en_core_web_lg.” spaCy’s syntactic parser automatically identifies “noun chunks,” which correspond to noun phrases. In addition, it identifies the “root” of the chunk, which corresponds to our definition of head noun in the present study. The head nouns identified were cleaned, removing interjections, punctuation marks, and other tokens mistakenly identified as head nouns.

In order to collect the adjectives modifying the head nouns, all tokens from the noun chunk that were not tagged as an adjective or verb were removed (we kept tokens that were tagged as verbs because they included deverbal adjectives, such as dedicated). Additionally, we excluded cases in which more than one adjective modified the head noun, because incorporating them as part of the analysis is not straightforward, while at the same time they are very infrequent (a few hundred cases for each quantifier). Finally, we removed tokens that were tagged as adjectives but were actually determiners (e.g., many, few) or modifiers of adjectives (e.g., very, really).

Table 6 shows the total and normalized frequencies per 1 million words (in parentheses) of each quantifier in different constructions.

Table 6. Total and normalized frequencies per 1 million words (in parentheses) of each quantifier with different constructions.

4.3. Method

For the nouns and adjectives, we conducted a Distinctive Collexeme Analysis (DCA) (Gries & Stefanowitsch, 2004), which is part of the family of collostructional methods (Stefanowitsch & Gries, 2003). This approach is useful for identifying distinctive co-occurring lexemes that occupy a specific slot of each construction, thus distinguishing them significantly from each other. A collexeme is a lexeme that occurs in a particular slot within some pattern. In the present study, the analysis identifies the collexemes that distinguish a few from several.

We conducted two separate analyses, one for the head noun slot, and another for the optional adjective slot. We first calculated an association measure for each collexeme (i.e., noun or adjective that co-occurs with one of the quantifiers) based on a contingency table. The frequencies we relied on for computing the strength of association are (1) the frequency of the collexeme with a few, (2) the frequency with several, (3) the frequency of a few with all other collexemes, and (4) the frequency of several with all other collexemes. Table 7 is an example of such a table for the head noun minutes. Data were filtered to only include nouns that co-occur at least 20 times with at least one of the quantifiers, and adjectives that co-occur at least 15 times with at least one of the quantifiers.

Table 7. A contingency table for the noun minutes that serves as input for the DCA.

The numbers in parentheses indicate the expected frequencies.

The output of the DCA is a list of all the collexemes that figure in the selected slot, together with a number that stands for the strength of the association. We used the association measure log-likelihood, or G² (Dunning, 1993). A G² score of 3.85 corresponds to a p-value of .05. To determine the direction of the association, when the expected frequency of co-occurrence with several was higher than the observed frequency, the G² scores were multiplied by −1. Thus, a positive score indicates that a collexeme is distinctive of a few, and a negative score indicates that it is distinctive of several.

Table 8. Number of instances used in the current study.

For the adverbs and constructions that mark argumentativity, we conducted a modified DCA to examine the association of the quantifiers with constructions that explicitly mark different aspects of argumentativity. The list of constructions was compiled by the authors. It includes two focus operators (only and just), the exceptive except for, and but in its adversative and non-adversative uses.

Only and just mark their complements as weak arguments. The same applies to exceptives that typically mark their complements as insufficient counterarguments. The non-adversative functions of but are “exceptive” (as in all neighbors but Cohen paid …) or restrictive focus (he is but a child). Their argumentative contribution is like the other focus particles. In contrast, but’s adversative function marks its second conjunct as argumentatively strong (see example (3)), which is the position we examined.

For each construction, we calculated a G² score in the same manner as in the DCA. Since but is polysemous, we randomly extracted 200 instances with each quantifier, counted the number of relevant senses, and extrapolated the total co-occurrence frequency. Finally, we examined the distribution of the two quantifiers across the different genres in COCA to see if it can account for the results of the DCA.

4.4. Results and analysis

4.4.1. DCA – results

4.4.1.1. Nouns

Overall, the dataset contained 1,376 noun types; 717 nouns were distinctive for several and 385 for a few. This means that only <20% of the nouns (271, 19.7%) did not have any preference for one of the quantifiers. Since the objective difference between the two quantifiers is rather minor, this result is surprising.

Table 9 presents the top 15 nouns distinctive for each quantifier, ranked by their strength of association.

Table 9. Top 15 distinct collexemic nouns for each quantifier.

4.4.1.2. Adjectives

Overall, the dataset contained 369 adjective types; 143 adjectives were identified as distinctive for several and 134 for a few. Again, only about a quarter of the adjectives were not distinctive for any quantifiers (92, 24.9%).

Table 10 presents the top 15 adjectives distinctive for each quantifier, ranked by their strength of association.

Table 10. Top 15 distinct collexemic adjectives for each quantifier.

4.4.1.3. Discourse prominence

Looking at the results of the DCA, we argue that they present evidence for several’s complements having a higher discourse prominence. This manifests itself differently in nouns and adjectives. For nouns, discourse prominence is achieved via individuation, which we measured as an association with the partitive construction. For adjectives, it is achieved by adjectives of significance and individuation. The analyses were carried out on the full list of significant collexemes of each type.

Prominence is related to strong argumentativity because prominent entities are more difficult to ignore. Thus, when evaluating the strength of the argument, they are more likely to be factored in. In addition to being more prominent, complements that are composed of more individuated entities contribute to a stronger argument. This is because when distinct sources support the same conclusion, it is perceived as more firmly established than when the conclusion is supported by a homogenous source, even if it is robust because it is a plurality.

4.4.1.3.1. Nouns – association with partitives

We checked whether the entities denoted by the relevant plural nouns are eligible to function as discourse referents. Looking at the list of nouns, we hypothesized that one of several’s features is that its complement tends to introduce more individuated entities. In other words, the entities are differentiated from each other in a meaningful way. One way this might manifest itself is in their ability to serve as distinct discourse referents. We operationalized this propensity by examining the noun’s attraction to the partitive construction (attracted/rejected/neutral). The partitive construction is composed of a quantifier/numeral, of, and a potentially modified noun. Some COCA examples of sentences with several and a few in the partitive construction are presented in (10):

The partitive construction carves out a new discourse referent out of an existing one. In (10a), for example, a subset of Dostoyevski’s books is singled out and becomes an entity on its own. The writer then continues to name these books and elaborate on how they support their argument. Thus, the construction is a reasonable strategy to introduce new entities into the discourse. If nouns distinct for several are typically perceived as more individuated, we predicted that they would more likely be attracted to the partitive construction than the ones distinct for a few. Nouns distinct for a few, however, would more likely be rejected from the partitive construction than the ones distinct for several.

For example, the noun moments, which is strongly distinct for a few, may be prominent enough as a whole entity to receive the status of a discourse referent. However, it is (typically) less meaningful for us to distinguish the individual moments composing this whole, as they are too small to be told apart insightfully. Of course, we could construe a plural entity composed of moments such that every moment will carry a crucial contextual weight and thus be individualized. However, this is not the typical way we perceive such entities.

To calculate the association with the partitives, we first extracted all partitive constructions with quantifiers or numerals in COCA. We then applied a collostructional analysis (Stefanowitsch & Gries, 2003) and calculated a G² score for each noun. Nouns that received positive, significant scores were tagged as “attracted,” significant negative scores as “rejected,” and the rest as “neutral.”

Fig. 6 presents the distribution of the association level with the partitive construction for the nouns distinctive of each quantifier. We see that for several, the distinctive collexemic nouns are rejected from and attracted to the partitive construction in equal proportions (about 40% each). In contrast, the nouns distinctive for a few are overwhelmingly rejected by the partitives (70.4%), while only a small minority (18.4%) is attracted to them.

Figure 6. Association of distinctive collexemic nouns with the partitive construction.

A logistic regression analysis was conducted to examine the relationship between the quantifier (the response variable) and its association with partitives (the predictor variable). The predictor variable was dummy-coded, with the neutral level as the reference level. The results of the analysis revealed that there was a significant effect of association with the partitives on quantifier when the level of association was “rejected” (estimate = −1.2, SE = .2, p < .001). However, no significant effect was found when the level of association was “attracted” (estimate = .13, SE = .22, n.s.). The coefficients are presented in Table 11. A post-hoc analysis of simultaneous pairwise comparisons using Bonferroni adjustment indicated that only “Attracted” and “Rejected” (Z = 8.74, p < .0001) were significantly different.

Table 11. Logistic regression model. Formula: Quantifier ~ association

These findings suggest that the critical level of association is rejection. Nouns rejected by the partitive construction are more likely to be distinctive for a few. This indicates that the nouns associated with a few are indeed less prominent than those associated with several. We propose that this results from their lower individuality, which explains why they are worse candidates for strong arguments.

4.4.1.3.2. Adjectives

Adjectives were tagged for prominence (prominent/insignificant/neutral) based on two criteria. First, adjectives that convey significance (e.g., significant, important) were classified as associated with prominence. Some of these are adjectives from the domain of FOCUS, such as notable and prominent. Focused entities indicate discourse prominence; the more important something is, the more attention we pay to it (Frännhag, 2010, p. 112).

The second criterion is individuation. These adjectives include different, independent, and distinct, but also similar, which indicates the likeness (on some dimension) of entities that are actually distinct from one another. The relation between individuation and prominence was discussed in the previous subsection.

Insignificance is likewise effected via different routes. First, there are adjectives that profile a low region on a scale of importance, such as minor and secondary. Second, there are adjectives that indicate exceptions. Adjectives such as token or exceptional mark the nominal as an exception, which is not enough to undermine the conclusion supported by the typical case. Other adjectives convey a spatial configuration that underlies an inference of insignificance – scattered, isolated, and stray are some examples. Once again, the nominal is construed as insufficient to support a claim by itself (consider, for example, the type of conclusion established on the basis of a few isolated incidents).

Table 12 summarizes the types of adjectives for each quantifier. For several, about 20% of the adjectives were tagged as conveying prominence, almost 80% as “other,” and only a single adjective was found to convey insignificance. Conversely, for a few, only a single adjective was found to convey prominence, while about 14% of the adjectives were adjectives of insignificance. Even impressionistically, it is evident that several’s adjectives are much more likely to convey prominence than those of a few.

Table 12. Types of adjectives by quantifier with percentages in parentheses

We fitted a logistic regression model with quantifier as the response variable and adjective type (prominent/insignificant/other) as the predictor variable. The predictor variable was dummy-coded, with “other” as the reference level. The results of the analysis revealed a significant effect of adjective type on quantifier when the adjective conveys prominence or insignificance. When an adjective conveys prominence, it is more likely to be distinctive for several. Conversely, when an adjective conveys insignificance, it is more likely to be distinctive for a few. The coefficients are presented in Table 13. A post-hoc analysis of simultaneous pairwise comparisons using Bonferroni adjustment indicated that all comparisons are significantly different (see Table 14). All in all, together with the results for the nouns, we see that nominals modified by several tend to have greater discourse prominence, rendering the entire nominal argumentatively stronger.

Table 13. Logistic regression model. Formula: Quantifier ~ adjective type

Table 14. Pairwise comparisons of adjective types. p values are Bonferroni-adjusted

4.4.2. Adverbs

Table 15 presents the results of the adverbials examined, which show that out of all selected adverbs, only the second conjunct of the adversative use of but (the one construction that marks the relevant slot as a strong argument) is distinctive for several. The other adverbials, all marking weak argumentativity, are distinctive for a few (although except is only nearly significant). This finding is in line with the results thus far – a few is more highly associated with the markers of weak argumentativity, while several is more highly associated with a marker of strong argumentativity.

Table 15. Argumentative constructions and adverbs and their association with the quantifiers.

4.4.3. Genres

We next examined the distribution of several and a few in different genres in COCA (see Fig. 7) to see whether their distribution can be explained by reference to genre formality. If so, the DCA results might reflect several’s preference for more formal genres (academic writing and newspapers) and a few’s preference for less formal genres (blogs, fiction, and TV/movies). It is then possible that certain words, such as factors, turned out to be distinctive for several simply because these words are more prevalent in the same genres as several.

Figure 7. Distribution of tokens in different genres of COCA.

To determine whether genre alone could explain the data obtained from the DCA, we examined the co-occurrence of the distinctive collexemes with the two quantifiers in each of the eight genres in COCA. We calculated the expected and observed co-occurrence frequencies of the collexemes with each quantifier and then checked for each collexeme how many genres the relationship between the observed and expected frequencies aligned with the results of the DCA. A genre was considered aligned with the results of the DCA for a collexeme distinctive of several/a few if the observed frequency of co-occurrence with several/a few was higher than the expected frequency in that genre. Collexemes that did not co-occur with either quantifier in a genre were also considered non-aligning for this analysis. The results are presented in Tables 16 and 17, respectively.

Table 16. Collexemic nouns per aligning genres.

Table 17. Collexemic adjectives per aligning genres.

For a few, we see that the observed occurrence frequency of 83% of distinct collexemes is higher than the expected frequency in at least six genres (out of eight). For several, this number is 78%. That is, the vast majority of distinctive nouns are distributed in alignment with the results of the DCA in almost all genres. For the adjectives, the proportion of collexemes with at least six aligning genres is lower: 56% for a few and 64% for several.

To see if the results of the analysis for the adjectives still hold when the genre is factored in, we took only the adjectives that are distributed in alignment with the DCA results in at least six genres and used the same regressions model as in Subsection 4.4.1.3.2. This time, only insignificance was significantly different from the baseline “other” category (estimate = −2.6, SE = 1.05, p < .05). This means that several’s preference of adjectives conveying prominence is partially mediated by genre. However, while genre is a relevant factor, it cannot explain the totality of the results.

Moreover, we propose that these genre findings actually lend further support to our argumentative analysis, since “argumentative” lexical items should occur more frequently in more persuasive genres. Indeed, several is more prevalent in academic writing and newspapers, where persuasive discourse is often central. On the other hand, a few is more prevalent in blogs, fiction, television, and movies, where persuasion does not figure prominently.

4.5. Discussion of the corpus data

The results of the corpus study point to two major differences between several and a few. These relate to discourse prominence (a few’s complements tend to be less prominent and less individuated) and argumentative strength. In addition, several is more prevalent in persuasive genres than a few. Together, these findings support our argument that several is argumentatively stronger than a few. Finally, there are good reasons to believe that genre is responsible for some (though far from all) of the differences between the two quantifiers, in addition to the previously established difference in quantity.

5. Analysis

To summarize, combining previous research with our findings shows that:

1. a. While both a few and several are associated with quantities considered small, several is associated with a higher quantity than a few.

2. b. A few is argumentatively weaker than several.

3. c. Several construes its complement as internally composed of more individuated entities than a few (and thus as more prominent).

4. d. Several is prevalent in persuasive genres, while a few is prevalent in non-persuasive ones.

We suggest that these tendencies define distinct prototypical discourse profiles (Ariel, 2008) for a few and several. A discourse profile specifies optional contextual factors that accompany the use of a construction. Note, however, that distinct discourse profiles must not be taken as grammatical rules. Several can denote a relatively small quantity of an individuated entity in a spoken conversation without contributing to an exceptionally strong argument. But a prototypical instance of several (or a few) manifests its prototypical discourse profile.

What about the core meaning of each quantifier, which is stable across different contexts? We assume that the two share some features, for example, denoting a quantity larger than one. However, building on Langacker’s suggestion that several evokes a quantized scale, we could incorporate this concept into several’s core by postulating that the quantity needs to be countable in addition to being larger than one. ⁴ Although it is possible to interpret the corpus findings as supporting this analysis, it is difficult to say at this point whether this finding actually has implications on the countability of several’s complements.

Now, there is some symmetry between the components of the two discourse profiles, but when we examine the diachrony of the two quantifiers, we see that each quantifier took a different route to acquire its profile. Lack of space precludes a detailed account based on diachronic data. Instead, we rely on generally established principles of grammaticization and try to reconstruct the diachronic evolution of the expressions from the synchronic data.

Breban (2008) describes the path leading several from the original Latin sēpar “separate, distinct” into a grammaticized adjective first and then into a quantifier. According to Breban, in its adjectival form, several indicated that the plural entity is a set of distinct individual instances. ⁵ That is, the individuating components existed prior to argumentativity and quantity. The fact that individuation still prevails in the quantifier use of several follows naturally from Hopper’s (1991) principle of persistence, according to which traces of original lexical meanings persist despite diachronic changes.

As for quantity and argumentation, we suggest that individuation led to strong argumentation (via prominence, see Section 4.3.3). Next, the use of several to form strong arguments most likely biased it toward quantities that are not so low. Intensity schemas, such as “the more X, the more Y,” often underlie argumentativity (topoi according to Anscombre and Ducrot). Using several to forge a strong argument that supports Y may invite the inference that the quantity of X is relatively larger. Thus, the general path we suggest is:

$$ \mathrm{Individuation}>\mathrm{Strong}\ \mathrm{argumentativity}>\mathrm{Larger}\ \mathrm{quantity} $$

According to the OED, a few originates from “the same Indo-European base as […] ancient Greek παῦρος little, small, classical Latin parvus small, classical Latin paucus few.” This explains the salience of the small quantity component of a few. A small number is more compatible with weaker argumentativity, which is in turn compatible with lower individuation and prominence.

This analysis also accounts for the difference demonstrated by (7a) and (7b). Several might sound odd when there is a large, backgrounded complement set, because there is a connection between individuation and set size. Typically, as the size of a set increases, the individuation of its members decreases.

Table 18 summarizes our analysis. Two points are worth mentioning. First, although we suggested a grammaticization path where one feature led to another, the three components of each discourse profile coexist synchronically as a bundle. Even though for several individuation came first, we found no evidence of its primacy over the two other components (strong argumentativity and larger quantity). Second, we treat the distributional patterns presented in the bottom part of the table as lower-level generalizations, because they naturally follow from the two distinct discourse profiles.

Table 18. Result summary and analysis.

6. Concluding remarks

Our findings paint a picture in which quantifiers typically thought of as logical or formal entities are, in fact, associated with rich conceptual structures, much like other linguistic units so analyzed by cognitive linguists. Such analyses highlight the importance of detailed representations of cognitive models of the world for language meaning and use. For the argumentative function to be fulfilled, interlocutors need to have an underlying implicit knowledge of how we decide what type of evidence we may find more convincing or less convincing. This knowledge goes far beyond associating a quantifier with a set of specific cardinality. An example of a model that can account for such knowledge is Fillmore’s (1976) semantic frames, according to which semantic meaning is comprised of detailed models of situations in which components have different roles.

The results of this study demonstrate how cognitive concepts relate to discourse and vice versa. For example, in the case of several, individuation, which relates to how we perceive the internal composition of a plural entity, motivates discourse prominence, which in turn influences the quantity associated with the quantifier. We hope this research will inspire further investigations into the relationship between cognition and discourse.