2.1 Weight of syntactic operations

A question of significant interest for syntacticians in the generative framework is which syntactic operations are possible, and which ones are not possible. And among those that are possible, which ones cost more than others. Cinque (2005; 2013) makes a proposal that assigns different weights to different syntactic operations, in order to derive the typological frequencies of different word orders in the DP. Experiments 1 and 2 of our study use linear regression to derive these weights automatically. Crucially, Cinque proposes that one kind of movement, recursive pied-piping of the NP, is cost-free. Our results confirms that this kind of movement is mostly cost-free, and that, in specific circumstances, it is more desirable than partial movement. In this way, we explain a fact that has remained unexplained in Cinque’s accounts: word orders produced by derivations in which no movement has occurred are attested by fewer languages than those in which movement has occurred.¹

2.2 Structural position of cardinal numerals

Having established the relative costs of movement to derive word orders in the DP, we turn to the category of some of the elements in the DP. Specifically, another question that has been the object of a lot of attention in theoretical linguistics in recent years is the syntactic category of cardinal numerals occurring in noun phrases, such as those shown in (1).

Since numerals can appear in various syntactic contexts, it is not clear what syntactic category fits them best. It has been argued that numerals are quantifiers, adjectives, nouns, or a combination of categories. Cinque’s (2005) proposal to account for Universal 20 assumes a high merge position for numerals, much like quantifiers. Stavrou & Terzi (2008; 2009), who examine numerals in Greek, argue that indeed, simple numerals like three in (1-a) constitute a subclass of weak quantifiers, like many in (1-b). In contrast, other researchers treat all numerals as nouns, like garden in (1-c) (Hurford 1975; 1987; 2003; Ionin & Matushansky 2004; 2006). Ionin & Matushansky (2006) further propose that numerals have the semantics of modifiers, like adjectives. Landman (2000), in fact, argues that all indefinite DPs are predicates of type <e, t>, which requires numerals to be adjectival, like big in (1-d). This would allow numerals to merge either higher or lower than other adjectives.

Experiments 3 and 4 of the current study compare two of these proposals from a typological perspective, namely merging numerals high versus merging them low, as well as a number of intermediate views where they merge high in some languages and low in others. The results suggest that, assuming numerals constitute a uniform syntactic category cross-linguistically, treating them as structurally high allows for a better prediction of the typological frequencies than treating them as adjectives, and structurally lower.

2.3 Roadmap

The rest of the paper will develop as follows. Section 3 presents the computational and experimental method. The following three sections each present an experimental question: section 4 presents Experiment 1, which tests the predictive power of Cinque’s (2005) proposal for Universal 20 using linear regression, and automatically determines the weights of the different syntactic operations. Section 5 presents Experiment 2, which replicates Experiment 1, but changes the encoding to treat lack of movement as a uniform factor suggesting that the right kind of movement is not only free, but also desirable. Section 6 presents Experiment 3, which also uses linear regression, this time to evaluate the predictive powers of an alternative to Cinque’s (2005) proposal where numerals are allowed to merge either low or high as a parametric choice, rather than having an exclusively high position as assumed in the proposal. Subsection 6.5 presents Experiment 4, which replicates Experiment 3 using discrete categories (very frequent, frequent, infrequent or zero) instead of raw frequencies to control for the inherent uncertainty about typological counts, and trains a Naive Bayes Classifier to classify different word orders into one of these classes. The results of Experiment 3 and Experiment 4 are consistent, and both point to the high merge position as the better choice.

3.1 Encoding linguistic proposals as vectors of features

In Cinque (2005), Greenberg’s Universal 20 is derived from independently motivated principles of syntax organised in a derivational explanation. Based on data as those shown in the second to last column of Table 1, Cinque remarks that there are 24 combinatorially possible orders of the four elements: N, DEM, NUM, ADJ. Only 14 of them are attested in the languages of the world (see also Dryer’s counts in the same table, Table 1, and Cinque’s new counts, last column). Some of the 14 orders are unexpected under Universal 20. It is proposed that the actually attested orders, and none of the unattested ones, are derivable from a single universal order of the basic constructive syntactic operator, the Linear Correspondence Axiom (Kayne 1994), and from independent conditions on phrasal movement.

The order of merger in (2) is assumed, where only the overt NP or phrases containing the overt NP can move. The allowed syntactic operations (movements) are given in (3), omitting intermediate projections for simplicity.

The 24 possible permutations of demonstrative, numeral, adjective and noun are derived using these movements. Some orders require no movements, other require derivations of different numbers of movement steps. These derivations are summarized in Tables 2 and 3. The tables show two lines for each word order, for each movement step. The first line describes the word order movement operation, the second line gives the name of the type of movement according to our formal encoding.

Since each language in the sample has one dominant word order, and that word order is assumed to be the result of a derivation produced by some of the syntactic operations in (3), we treat these syntactic operations as binary parameters that a language either has or does not have.

In addition to the allowed movements, we define one parameter to describe movements like (4), which are argued to be impossible by Cinque (2005), and where a phrase not containing the overt NP is moved. These parameters are those in (5).

Since these parameters are binary, we can now encode the different word orders as a vector of values, by assigning either 1 or 0 to each parameter. Importantly, since the frequency of a word order is not correlated with the number of movements necessary to reach it, but rather with the type of movement necessary to derive it, we do not count the different occurrences of a certain movement. Rather, a parameter has a positive value, if it needs its corresponding movement, and we assign it the value 1, regardless of how many times it is needed for a given derivation.

To illustrate the encoding of three different word orders, step by step, consider the English word order DEM NUM ADJ N, the second most frequent order. This order can be obtained immediately by linearizing the structure (2), with no movements at all. This means that to reach this word order, no movements are needed. So all the parameters in (5) are set to 0 for this word order. This gives us the values in Table 4, which is equivalent to row a of Table 7.

The mirror order of English N ADJ NUM DEM occurs in 411 languages, an example is Sudanese Arabic. This order involves NP moving above ADJ; after this movement, the phrase [_YP [_NP N]] Adj] moves above NUM; finally, [_XP [_YP [_NP N]] Adj] Num ] moves above DEM. These movements are illustrated in (6).

Since all of these movements are movements of the NP with pied-piping (i.e. movements of the [NP[XP]] type), only the parameter (5-c) is set to 1, and all other ones are 0. This gives us Table 5, which is row x of Table 7.

Finally, to illustrate a less straightforward case, consider the word order DEM N NUM ADJ. This order is less frequent than the other two, it occurs in 37 languages, for instance Xhosa (Bantoid, Niger-Congo). In order to derive this order, the NP has to move above NUM, skipping the adjective, without pied-piping. This word order therefore requires movement of the NP without pied-piping, so the value of (5a) for this word order is 1. In addition, since the noun remains to the right of the demonstrative, the NP must stop after moving above NUM. This word order therefore requires partial movement (movement without going above DEM). So the value of (5-d) for this word order is 1 as well. Since no other movement is needed, all the remaining parameters are set to 0. So this word order is encoded as illustrated in Table 6, which corresponds to row c of Table 7.

Table 7 summarizes the encoding for each of the 24 possible word orders, as a vector of the six parameters defined in (5). Now that we have transformed the theory into a vectorial representation, we automatically find the relative weight of each element in the vector, a process of parameter fitting.

4.1 Materials

The primary data used in all four experiments is Cinque’s (2013) sample of languages containing the order of the four elements demonstrative, numeral, adjective, and noun in the DP in 1475 languages. The sample specifies the language, the word order, the genus, and the family. This data was collected by Cinque through examination of different grammars and resources, and converted into a table to allow computerized work.³ Except for one word order that turns out to be more represented than previously known, the overall tendencies in this large sample of languages is similar to earlier published samples by Cinque (2005) and Dryer (2006), as shown in Table 1. As is often the case with linguistic distributional data, the word order frequencies are skewed according to Zipf’s law. Only 5 of the 24 possible orders account for almost 70% of the world’s languages, and the remaining 30% share another 9 word orders, and 10 orders are unattested. This is summarized in Table 8.

Table 8

Zipfian distribution of word orders.

The different movement operations corresponding to each word order are shown in (5) and were discussed in the previous section. The features and the possible values of Cinque’s model are shown in Table 7. The data in Table 7 are the data that we use in our experiments. The values in the last column are the frequency property of the word order, the dependent variable that we are trying to explain.

4.3 Results and discussion

The linear regression model that was generated was the function in (7), whose goodness of fit is indicated by the correlation coefficient of 0.52.

The linear model in (7) can be read as a ranking of the different syntactic operations in terms of markedness. The weights are summarised in (8). Specifically, we note that large negative values of any of the parameters (5-a), (5-b), (5-e) and (5-f) are considered very costly.

If we interpret the weights as costs, so that high negative weights are high costs, we can rank the different movements in a partial order, as in (9) (where the symbol “<” means “less costly” and we use abbreviated symbols):

A system of weights can also be inferred from Cinque’s proposal (2005: 321), based on the markedness levels assigned to the movement types, as in (10):

Considering only the types of movements that are encoded in both accounts and simplifying labelling for readability, we have the two following partial orders:

We can see that the two orders are well correlated: tied orders in Cinque’s are ranked but adjacent by our ranking, and in only one case the rank is reversed (XPNP and Split) in the two orders. Kendall’s Tau_b is a measure of rank correlation that allows ties. A comparison of our two orders yields a Kendall’s Tau_b of 0.6, which is statistically significant at p < 0.5 even in such a small sample. We can conclude that our regression corroborates not only the nature of the movements but also, to some extent, the respective costs that were assigned in Cinque’s work.

The low cost of partial movement does not confirm Cinque’s assumption that partial movement is penalising. Rather, partial movement appears to be less costly than other operations. Partial movement was considered more marked than complete movement in Cinque’s system to derive the fact that the mirror image of the base order is very frequent. This is an order where movement is complete. However, Cinque’s system does not predict that the mirror image of the base order is more frequent than the base order. In Cinque’s model, no movement is unmarked. Since there seems to be a weak correlation between the weights assigned to complete and partial movement and typological frequencies, we investigate the relative weights of partial and complete movement. We develop therefore a new model, where we investigate different encodings of movement.

5.1 Materials

The same set of data was used for this experiment as for experiment 1. The encoding, however, involved a new parameter, which replaced Partial Movement. So the new set of parameters is shown in (12).

The materials are the same as those of the previous experiment, with modifications to the feature Partial Movement. In particular, in experiment 1 the partial movement feature was encoded as (13), such that complete movement and lack of movement were grouped together, and partial movement had a different specification.

In (13), partial movement is defined as involving any category, but receives a different value from no movement or complete movement. We also encode the two new models shown in (14) and (15). Both models consider that the distinction between complete and partial movement is relevant only for N. In these two models, the only distinguishing property is the fact that no movement receives the same value as complete movement, as in (14), and that in (15), no movement receives the same value as partial movement, defined for N only. The encodings of partial movements in the three sets of data is shown in detail in Table 9, and exemplified for one order in Table 10.

5.3 Results

The three models derived by the linear regression are shown in examples (16), (17), and (18).

The first observation about these models is that the NoMove feature, in any of its encodings, is among the least penalising and it ranks second least penalising, like in the model of experiment 1. In comparison to experiment 1, though, this encoding yields, in two cases, a much better correlation to the frequency data.

The comparison of the first encoding of complete movement compared to the second encoding of complete movement, anything moves above DEM vs N moves above DEM, shown in (17) and (18), shows that the two models yield in some cases the same correlation with an almost identical ranking of cost of features, but that results are mixed, and on average, the more general definition of partial movement yields better fit to the data.⁵

The comparison of the first and second encoding of movement compared to the third encoding of movement (N moves above DEM is as costly as not moving, partial N moves, not above DEM is as costly as not moving, where it is free) shown in (16), (17) and (18), and footnote 5 is much clearer. It shows that for these data no movement is always preferred. While lack of movement does not become particularly costly in the third model, the correlation coefficient is lower.

5.4 Discussion

The model in (14), where lack of movement is encoded in the same way as complete movement, as proposed in Cinque (2005), which considers both unmarked, yields better results than (15), where lack of movement and partial movement have the same value. Assuming that lack of movement is sometimes costly would explain the fact that the base order DEM NUM ADJ N is less frequent than the mirror order, N ADJ NUM DEM, which requires four movements. There are, in fact, theoretical precedents to this position. Indeed, Shlonsky (2012) proposes that movement of the noun (or the NP) over any agreeing DP components is obligatory. In other words, some languages (namely those with agreeing modifiers) cannot do without movement. This is encoded with the weights assigned in (15): If movement is required in a significant number of languages, and allowed in the rest, then lack of movement will be more penalized than the right kind of movement. The low correlation of model (15) does not appear to support this proposal.

This conclusion is corroborated by the fact that the model in (13) is better than the one in (14), which shows that no movement and complete movement appear to pattern together. This is what we find in the distribution of word orders: the most frequent word order is generated by roll-up movement of different kinds of phrases, and the third and fourth most frequent word orders are generated by partial movement of the N. These orders together, however, do not make movement a less costly option than no movement, exhibited by the unique base order. In other words, our model of cost of operations derives the number of languages that exhibit a given order in a complex configuration of values that involve both the type of movement and the category being moved.

5.5 Conclusions

The two experiments described in the last two sections use a novel method to address the question of which syntactic operations are possible, which ones are not possible; and among those that are possible, which ones cost more than others. The results confirm previous theoretical proposals, but also explore new possibilities.

In particular, the ranking of costs among operations proposed by previous studies is mostly confirmed (Cinque 2005; 2013). The similarity of correlation between the two definitions of partial movement, movement of any category or movement of N, confirms that partial movement is penalised in any definition, and leaves the door open for a hitherto unexplained fact, namely that between the two fully harmonic word orders, the one that according to Cinque’s theory requires movement (N ADJ NUM DEM) is more frequently attested among the world’s languages than the one that does not require any movement (DEM NUM ADJ N).

The modelling of movement operations in Cinque’s theory is based on some basic assumptions on the structural dominance of the syntactic categories that occur in the DP/NP. Cinque’s (2005) account places numerals below demonstratives and above adjectives. For a similar assumption, see also Stavrou & Terzi (2008; 2009). But since numerals can appear in various syntactic contexts, it is not clear what syntactic category fits them best or what syntactic category would best fit our data. In the next two experiments, we investigate what syntactic category better fits the typological distributional data, by studying different structural positions for numerals.

6.1 Materials

Experiment 1 already details the encoding of the assumption that numerals merge always above adjectives. In order to encode the idea that a numeral can merge either below or above adjectives, we encode each word order as two vectors, one that corresponds to the movements assuming the base order [Dem [Num [Adj [N]]]], i.e. those in Table 7; and one that corresponds to the movements assuming the base order [Dem [Adj [Num [N]]]], i.e. those in Tables 16 and 17.

To differentiate the two vectors for each word order, we add the binary parameter Involves the numeral merging below the adjectives, which is added to the list of parameters in (20).

Now, instead of having one vector per word order, with this new encoding, any given derivation of a word order has its own unique vector, which is formed by assigning values to the attributes in (20). And every word order has two derivations (and therefore two vectors). Going back to our example illustration, for this proposal, the word order DEM N NUM ADJ has two associated vectors. The first vector represents the movements involved in deriving it from the base order [_WP Dem [_YP2 Num [_YP1 Adj [_NP N ]]]], and is given in Table 14. The second vector represents the movements involved in deriving it from the base order [_WP Dem [_YP2 Adj [_YP1 Num [_NP N ]]]], and is given in Table 15. This results in two sets of encodings, one for each base order, shown in Tables 16 and 17.

6.3 Results

As it turns out, treating numerals as high-merging gives rise to better predictions than any of the formulas allowing numerals to merge either higher or lower than adjectives in the dominant base structure of a language. That is to say that the assumption that numerals merge higher than adjectives in the dominant merge order of all languages makes better predictions than the assumption that numerals merge higher than adjectives in the dominant merge order of some languages, and lower than adjectives in the dominant merge order of others. The result are detailed in the following two paragraphs.

The linear regression model that was generated by assuming numerals merge high is the function in (21), with correlation coefficient of 0.751. Notice that while the features are the same as those illustrated in Table 7, we repeat these same features here ten times, producing the same number of data points as the datasets whose results are shown in Table 20, for a fairer comparison.

The linear regression models that are generated by the assumption that numerals are adjectives are shown in Table 19, depending on the proportions of languages given to each derivation, given a word order.

6.4 Discussion

The results of this experiment show a sharp degradation in data correlation as the proportions of high-adjectival word orders increases followed by fluctuation between 0.38 and 0.41 for the datasets that have a majority of high-adjectival word orders. This shows that the dominant merge order in any given language for a DP containing a demonstrative, a numeral, an adjective, and a noun must be DEM NUM ADJ N, and not DEM ADJ NUM N. In other words, going back to the question of whether numerals merge strictly higher than adjectives in the dominant order of all languages or not, the results of this experiment suggest, to a certain extent, they do.

It is important to note, however, that the results showing that including the second merge order deteriorates the typological predictions does not automatically entail that it is not a possible merge order. As alluded to in section 5.4, it is possible for various merge orders and movements to take place for semantic reasons. For example, it is possible to merge the numeral lower than some adjectives for scope reasons. That, however, will not affect the dominant order in the language, as it will only occur in the (infrequent) cases of a “high adjective”, like last, or of a collective interpretation of the adjective, like heavy in (22-c).

Whether these are cases of numerals merging lower than their usual position (Ouwayda 2013; 2014), or of adjectives merging higher cannot be entirely determined using typological facts, and require an in-depth semantic analysis that goes beyond this paper. We note, however, that assuming only the high positions for the numeral, rather than two positions with the lower one never occurring in the dominant order of a language, performs slightly worse.

6.5 Experiment 4: The structural position of numerals: classification

The previous experiments attempt to predict the actual counts of languages per word order. Not all languages, however, have been documented in the typological literature and, for those that have, there is some debate on what is the dominant word order. This experiment addresses the possible criticism that results of experiment 3 are too dependent on the actual counts of different word orders, which have a tendency to change when new languages are attested and documented. For this reason, we group the exact frequencies in discrete equivalence classes, and we repeat the experiment as a classification task. The experiment concentrates on predicting the ranking of word order counts and confirms the results of experiment 3.

7.1 Related computational work on Universal 20

In this work, we have proposed using statistical models and classifiers to automatically model and evaluate some quantitative and qualitative aspects of proposals concerning Universal 20. Two previous pieces of work have also used linear regression modeling and classification to evaluate different explanations of linguistic proposals, specifically concerning Universal 20.

Cysouw (2010a) uses Dryer’s (2006) typological data on the order of the four elements demonstrative, numeral, adjective, noun, in the DP. Cysouw’s model is factorial, and not derivational. The features used are all observed, i.e. they can be directly observed in the data, and are based on the position of the elements with respect to each other and to the edge of the phrase. This factorial explanation does not provide a generative process that explains how the different word orders could arise from a common grammar, but it identifies the predictive properties of the frequency distributions of word order and their relative importance. Dryer also proposes a factorial explanation based on general principles of symmetry and harmony (Dryer 2006). Extending these factorial approaches to probabilistic modeling, Merlo (2015) uses Bayes Networks to determine the generalization ability of each of Cinque’s, Dryer’s, and Cysouw’s formalizations. She shows that while all three proposals fare well, Dryer’s formalization has the best generalization ability.

Cinque’s (2005) proposal derives the grammatically possible and impossible word orders, as traditional in generative accounts, but also derives the exceptions, and the different degree of markedness of the various possible orders. The proposal is carefully calibrated to obtain exactly this distribution of possible and impossible word orders, and their frequencies. But Cinque’s account faces an empirical problem: the frequency distribution of the attested orders is Zipfian, with a few common orders accounting for most of the languages, and a long tail of rarer attestations. Given this distribution, it is reasonable to wonder whether unattested orders are really grammatically impossible or if they simply have not yet been attested, but are, in fact, possible. On the latter hypothesis, Cinque is “overfitting” the data and the predictions are too strong. This observation is, indeed, confirmed by Cinque’s (2013) database that we use here: some of the order that Cinque (2005) indicated as unattested are, in fact, attested in this larger sample.

To avoid developing too strong a theory, other methods set out to explain the frequency differentials between pairs of word orders. In contrast to Cinque’s single universal base order, Abels and Neeleman (Abels & Neeleman 2009) propose a simplified movement theory. This alternative proposal makes much more use of base orders. Abels and Neeleman (2009) describe a system where eight word orders are base generated by linearising an unordered hierarchical structure demonstrably homomorphic to Cinque’s (2005). The remaining six orders are generated by leftward movement. They argue that their account is more parsimonious and more empirically adequate than the LCA or Cinque’s.

Since our account does not hinge on the LCA per se, which we do not explicitly represent, nor on any specific assumptions on the internal structure of the phrases, which differ in the two accounts, it is not clear if our method can say anything interesting about the difference of the two theories.

Abels and Neeleman’s account however is very restricted and makes interesting quantitative predictions. Specifically, it predicts, first, that all eight base orders should be roughly equivalent in frequency (counted by genera or by languages) and, second, that all six derived orders should be less frequent or, at most, as frequent as the base order from which they are generated. Typological counts are shown in Table 21.

A strict interpretation of the first prediction appears not to be fulfilled, especially if we look at the counts for genera. We can however notice that the base orders are all frequent, and none is rare, which seems instead to support the theory. Prediction two is largely confirmed. In only one case, the derived order does not appear to be convincingly more costly than its corresponding base order (NUM ADJ N DEM, 3 genera and 40 languages, compared to the derived N NUM ADJ DEM, 7 genera and 35 languages).

Future work will develop a systematic comparison of these two approaches. Since the two approaches have a very considerable difference in complexity, with Cinque’s allowing several types of movement, the comparison will have to include some measure of model complexity (for example by using Bayes factors) and is beyond the scope of the current article.

One of the clearest observations concerning Universal 20 above is that the two most frequently attested word orders are harmonic. This is a wide-spread observation in typology. While it is well-known that disharmonic patterns in the order of words exist, and, in fact most languages are not fully harmonic (Dryer 1992: fn12), typologists have long noted a preference for harmonic word orders, that is word orders in which the operand and the operator (i.e, complements and heads) are in the same sequence. The greater frequency of harmonic word orders follows in Cinque’s and our model from the smaller costs of the operations that are used to derive them. Other explanations for the greater attestation of harmonic word orders are rooted in grammatical principles (FOFC, for example, proposed in Biberauer et al. 2008); general learning and regularisation biases (Culbertson & Smolensky 2012; Culbertson et al. 2012); parsing or processing principles (Hawkins 2004; Sheehan 2013), such as, recently proposed Dependency Length Minimisation effects in the Noun Phrase (Gulordava & Merlo 2015; Gulordava et al. 2015).

A fundamental assumption of this paper is that frequencies within a language and across languages are an aspect of language that is systematically related to its formal properties and that requires explanation, both in its numerical magnitude and its distribution. Like Yang (2011), we attract attention to the explanatory relevance of probabilistic accounts and expected frequencies of linguistic facts.

In this respect our work is also related to other computational probabilistic proposals for language universals. In a widely discussed, and controversial, paper, Dunn et al. (2011) investigate the correlations between word orders that have been used to argue in favour of the existence of language universals (Dryer 1992). They claim that to demonstrate that these correlations reflect underlying cognitive or system biases, the languages must be sampled in a way that controls for features linked only by direct inheritance from a common ancestor. Correlation between two word orders may suggest that both are responding to some common evolutionary force or that one acts as a selective force for change in the other. To answer these questions, they develop a computational model that simultaneously accounts for phylogenetic uncertainty and estimates the posterior distributions of the parameters of the model of word order evolution.

This model is computationally sophisticated, but it aims to explain linguistic data that are relatively simple. Our approach shifts the focus of the investigation, and is computationally simpler, but linguistically more detailed. Although very simple, Naive Bayes belongs to the same class of models as it is a Bayesian model with latent variables, but our encoding of word orders is more distributed and encodes a derivational theory. To a large extent, our model confirms Cinque’s approach and its differential cost of movement operations. These differential costs were set up to explain a pervasive universal asymmetry between prenominal and postnominal modifiers, the asymmetry between the very restricted choice of word orders among prenominal modifers and the much larger set of options for postnominal modifiers. This asymmetry is not captured by any of the other models that allow symmetric base orders.

7.2 Related theoretical work

The syntactic category of numerals has been in debate in recent years, with proposals claiming that numerals are quantifiers (Stavrou & Terzi 2007, 2008), adjectives (Corbett 1978; Landman 2000), or nouns (Hurford 1975; 1987; 2003; Ionin & Matushansky 2006). We have focused here on comparing the predictions of treating numerals as quantifiers, or treating them as adjectives. The claims that numerals are quantifiers or adjectives have been supported by intralinguistic facts, including agreement, case marking, and scope.

The idea that at least some numerals are adjectives is based on a number of properties of numerals that are typical of adjectives. Corbett (1978) notes that, like adjectives, low numerals in Russian, for example, agree with the following noun (23) in certain features including gender, number, case, and animacy.

Stavrou & Terzi (2008), in contrast, propose that numerals are a subclass of weak quantifiers, endowed with a [card] feature and are located in the specifier of a number phrase. They base this on a number of similarities between numerals and weak quantifiers. They note that many of the resemblances between numerals and adjectives also apply to quantifiers. For example, like adjectives, numerals and quantifiers can both occur in existential there-sentences (24). They can both be preceded by determiners (25).

Also, in negated contexts, both numerals and quantifiers have scope ambiguities, as in (26). The sentence can mean either ‘There are three/many women I didn’t see’ or ‘I didn’t see many/three women, I saw few/five’.

In addition, they show that numerals and quantifiers share properties that adjectives do not have, namely that cardinals and quantifiers are both able to license bare subjects in Greek (27). They can both appear without a noun, while an adjective cannot (28). They can head a partitive construction which an adjective cannot (29). And they both allow for split topicalization in Greek which an adjective cannot (30).

Many researchers view numerals as not having a uniform category, including Corbett (1978); Shlonsky (2004); Zweig (2005); Corver & Zwarts (2006); Danon (2012), but rather as being split into adjectives and nouns, or adjectives and quantifiers. Our reasoning in this paper assumes that numerals do form a uniform category. This is the same assumption that is made when they are referred to uniformly in the typological data.

There is much intra-linguistic evidence for more than one position for numerals. Researchers, including Zabbal (2005); Ouwayda (2013; 2014) propose that numerals do not have a single merge position in the DP, but rather can merge either higher or lower than adjectives. In Ouwayda (2013; 2014), for example, there are two functional projections that can host a numeral: A high one, which behaves much like quantifiers, and is associated mainly with a distributive interpretation, and another low position associated with a collective interpretation (like in sentence 20c)). It is important to note, however, that the two merge positions motivated in these works are semantically motivated. For Ouwayda (2013; 2014) for example, the difference is in distributivity. Similarly, in Pereltsvaig (2006); Ionin & Matushansky (2006), and Matushansky (2015), Russian numerals can have different positions depending on whether they are interpreted approximately or exactly. It is therefore crucial to distinguish between assuming multiple positions for numerals that are decided parametrically versus multiple positions for numerals that are semantically motivated within a given language.

Our work does not rule out the possibility of a lower merge position that is motivated semantically. It specifically rules out the possibility that different languages have different merge positions. The confirmation that the high merge position for numerals corresponds to the dominant order in all the languages, combined with intra-linguistic evidence of word orders that require a low merge position for numerals, lends support to the idea that the low merge position for numerals is semantically-motivated.