Version classiqueVersion mobile

New Perspectives in Biblical and Rabbinic Hebrew

 | 
Aaron D. Hornkohl
, 
Khan Geoffrey

Parts of speech in Biblical Hebrew Time phrases: A Cognitive-Statistical Analysis

Cody Kingham

Résumé

Though time phrases function adverbially, they can be headed by nouns, adjectives, and adverbs. This non-prototypical behaviour makes time phrases a useful context in which to test theories on parts of speech. The collocation behaviour of head words in time phrases can be measured statistically as a way of capturing their semantics. These objective statistical data provide a necessary control on more subjective linguistic hypotheses. In that vein, this article explores a new method of ‘cognitive-statistical’ analysis. The behaviour of words found in over 3,400 adverbial phrases in Biblical Hebrew is analysed using Principal Component Analysis (PCA), an unsupervised clustering technique. The resulting analysis uses the PCA model of word collocations to automatically classify their parts of speech. The end results prove promising for future classifications based on statistical analyses of word behaviour rather than on pure intuition.

Texte intégral

1.0. Cognitive-Statistical Approach

  • 1 See, though, the important work of van der Merwe (1997). There have also been numerous studies dedi (...)
  • 2 van der Merwe (1997, 49) asks similar questions:
    Is there any difference in the syntax of non-tempo (...)

1With the exception of a few small studies, phrasal time adverbials in Biblical Hebrew (BH) have received little systematic linguistic treatment.1 The reason for this may be because simple phrase adverbials present no obvious challenges for the classical linguistic method. As the grammars reiterate, time adverbials are simply adverbs or noun phrases that optionally indicate time reference (GKC §118a; Waltke and O’Connor 1990, §10.2a; van der Merwe, Naudé, and Kroeze 2017, §33.1). But this simple description leaves many unanswered questions, among them: What qualifies as an adverb or noun? How do adverb-headed time phrases differ semantically from noun-headed phrases? How are noun phrases utilised within the time adverbial function ?2

2These questions are not trivial, but go to the very foundation of grammatical inquiry. In recent years, Hebrew linguists have become more interested in foundational questions (e.g., Forbes 2009; Miller-Naudé and Naudé 2017). The interest parallels a sea change in general linguistics, largely driven by new challenges from the cognitive linguistic perspective (Geeraerts 2010; Scholz, Pelletier, and Pullum 2011). At the same time, advancements in machine learning and brain imaging seem to corroborate cognitivist claims that language categories are learned rather than innate (Perek and Goldberg 2017; Fedorenko et al. 2020). This calls into question the degree to which a priori categories can be assumed within individual languages (Haspelmath 2007). In BH, linguists have long recognised that expected categories such as ‘noun’, ‘adjective’, and ‘adverb’ do not always fit (Waltke and O’Connor 1990, §39.3.1a; van der Merwe, Naudé, and Kroeze 2017, 380–81). Given the new advancements, BH categories like ‘noun’ and ‘adverb’ are ripe for fresh evaluation.

  • 3 I agree with Ellis, O’Donnell, and Römer (2013, 30): “We take the Saussurian (1916) view that the u (...)
  • 4 On similarity-based categorisation see Sloutsky et al. (2015).

3The cognitive approach provides an elegant and powerful explanation for the huge variety of forms observed across world languages. The simple human impetus to form categories around observed data drives the creation of linguistic patterns. Language, then, is a cognitive tool for categorising the world (Geeraerts and Cuyckens 2012, §2). Humans connect categories built from experiential knowledge to linguistic sounds and signs (cf. Saussure).3 The concepts of prototypes (‘best example’), family resemblance, and gradient categories are central to the structure of categories (Lakoff 1987, 41). As it turns out, these structures line up with quantitative patterns observed in the world. Longtailed distributions (e.g., Zipf’s Law) pervade natural, biological, and sociological systems and may drive the formation of prototypes (Ellis, O’Donnell, and Römer 2013, 31–33; Piantadosi 2014). Hebraists will recognise this distribution from lexeme frequencies: a handful of terms (e.g., -ְ ו ‘and’, -ְ ה ‘the’, א ָ רמ ‘say’) occur very frequently, whereas thousands of terms occur only once (hapax legomena). Language learners likely use the most frequent items, the prototypes, as reference points for analogically learning new categories.4

  • 5 Indeed, as noted above, these mental associations are learned from relationships observed in the wo (...)
  • 6 Famously summarised by Firth (1962, 11) as “knowing a word by the company it keeps.”
  • 7 Based on an ongoing analysis by this author: https://github.com/codykingham/noun_semantics/blob/mas (...)

4The link between statistical patterns and cognition opens the door to an empirical method for studying language categories (Stefanowitsch 2010). Probabilistic patterns in the world provide humans with training data, which they then re-externalise using language (Goldberg, Casenhiser, and Sethuraman 2004). Mental associations between cognitive concepts thus give rise to statistical associations between language patterns.5 These associations appear most clearly in the phenomenon of co-occurrence, or collocation, between related constructions.6 For instance, מ ֶ ל ֶ ך ‘king’ is statistically associated with צ ִ ו ָ ה ‘command’ in the Hebrew Bible.7 Conceptually this is due to the relatedness of the two terms with ‘authority’. Synonym relations can be identified by indexing which terms collocate with the same kinds of items; for example: רש ‘prince’. This statistical concept has been successfully extended to grammatical constructions, further demonstrating that lexicon and syntax are not fundamentally distinct (Stefanowitsch and Gries 2003). The collocation principle also underlies a number of recent advancements in machine translation (see, for instance, Devlin et al. 2019).

5The cognitive-statistical approach I propose here aims to work backward from statistical associations towards hypothetical cognitive associations, as illustrated below.

Figure 1: Mutual relationship between linguistic, cognitive, and statistical data

Figure 1: Mutual relationship between linguistic, cognitive, and statistical data
  • 8 These links may not be one-to-one (e.g., concept B > construction Z and Y).

6The right side of the chart represents cognitive concepts which are ultimately stored in the minds of language users, while the left side represents externalised language in the form of words, phrases, etc. Mental concepts have logical associations which connect them (e.g., king is connected with ‘authority’; ‘leaves’ with ‘green’). However, the problem for linguists is that these concepts are inaccessible to empirical methods, and research has shown that introspective linguistic judgments can be inconsistent (see Schütze 2016). One solution is to recognise that the way constructions co-occur (‘statistical associations’) are motivated by the underlying cognitive concepts they are linked with.8 Thus, by identifying statistical associations between language constructions, one can work backwards and begin to ask, why is this statistical association here? Cognitive linguistic theory can then provide hypotheses. In this way, the data serve as a more objective control for testing linguistic hypotheses (see Geeraerts 2010).

7In BH, Forbes has pioneered data-driven models to study word classes, using unsupervised clustering algorithms to automatically group words based on their co-occurrences (Forbes 2005, 105–39; 2009). The approach yields promising results, with clear tendencies and recognisable groups. The methods of analysis for this study are different, but similarly reliant on the principle of collocation.

2.0. Dataset and Scientific Framework

  • 9 This is the process known as ‘operationalisation’ and is the ‘secret’ to the success of the natural (...)

8The objective of this analysis is to measure the behaviour of words which head time adverbial phrases. What can powerful statistical tools tell us about the natural behaviour of such words? And do we find any evidence for parts of speech categories? To that end, the study requires a dataset that has wide coverage of the relevant linguistic features and is freely accessible for scholarly scrutiny. Data-driven research is not theory neutral, nor is it immune to bias (equally true in the natural sciences; see Kuhn 2000). Rather, scientific research aims to make its assumptions clear, gather and measure data methodically, test the assumptions against the data, and make the results available for public scrutiny.9 Any data-driven study that does not make its full dataset available does little to advance the field.

  • 10 This is done using a custom Python parser which works within the existing phrase segmentation of th (...)

9The open-sourced BHSA of the Eep Talstra Centre for Bible and Computer (ETCBC) meets the proposed criteria (Roorda et al. 2019). The BHSA is accessed using Python and a corpus analysis package, Text-Fabric (Roorda 2017). It contains annotations for phrases interpreted to indicate adverbial time modification, with coverage for the whole of BHS. The ETCBC has made the annotations over a forty-plus year history (Oosting 2016, 192– 209; Kingham and Van Peursen 2018). As with any dataset, BHSA represents only one interpretation. It also has a few shortcomings. The analysis is based only on qere forms. Furthermore, relations below the phrase level are not always reliable, and there exists no tagging to explicitly indicate headship. For this reason, phrases must be pre-processed using a custom-built parser to isolate the necessary relations.10

  • 11 The whole project is archived at the following reference. For this article, data is stored in the v (...)

10It is important to note that the retrieval and tagging of large quantities of text always assumes a trade-off between precision and recall, i.e., a ratio of good or missed matches (Stefanowitsch 2020, 111–16). Realistically, large-scale studies require a compromise between perfection and coverage (Stefanowitsch 2020, 113). Especially with the parsing of internal phrase relations, edge cases may remain in this study’s sample. The remedy is good exploratory data analysis. Results should not be treated as a black box, but carefully examined in connection with the actual underlying data. The datasets, along with all the code for the analysis, are published online for scrutiny.11

  • 12 I used Croft’s (2001, 257–59) notion of semantic headship as a general guide. Objects of prepositio (...)
  • 13 A measure called deviation of proportions (Gries 2008, 415–19) was used to compare the sampled dist (...)

11The primary dataset is a subset of all phrases in the Hebrew Bible stored in the BHSA. The entire BHSA contains 4,041 phrases marked for adverbial time function in a clause. Many of the phrases contain embedded sub-phrases that are not always wellmarked. To ensure accuracy, the set is pruned down to a group of shorter, simple phrases. The resulting dataset contains 3,442 phrases that are parsed for head words and modifiers using a generally semantic definition.12 A test to see whether the subsampled phrases negatively affected any particular book’s representation showed no significant effects.13

  • 14 Throughout the study, I refer to proportions out of 1, which can be converted to percentages by mul (...)

Table 1: Time Adverbial phrase function frequencies and proportion of selected samples versus BHSA14

Table 1: Time Adverbial phrase function frequencies and proportion of selected samples versus BHSA14

12Throughout the study, I refer to the phrase label ‘Time’ in reference to the samples rather than the actual linguistic category. Thus, I do not assume a fully one-to-one correspondence. The tagged phrases can be viewed as tools of analysis that help to approach the object of study. This is on analogy with other scientific objects of measurement (e.g., a ruler, or a scale) which approximate rather than exhaust reality (Stefanowitsch 2010, 358–61). Alternatively the phrases and tags can be viewed as a hypothesis (one interpretation) in need of testing.

  • 15 This includes articles reflected only in the vocalised text of BHS.
  • 16 In this case the modifier term is the nomen rectum, or trailing term in a construct chain, rather t (...)
  • 17 The attributive adjunct construction involving the definite article, i.e., - ַה + word + -ַה + adju (...)

13For each of the sampled Time phrases, nine types of modifiers are tagged when found modifying the head: plural endings, pronominal suffixes, definite articles,15 construct relations (called ‘genitive’ here16), demonstratives or ordinals, cardinal quantifiers, and other quantifiers (e.g., .) כ ּ ל17 Another label, ‘Ø’, indicates no modifying specification of the headword except for the allowance of prepositions. I justify the choice of these modifiers more fully in the linguistic section. But, in brief, these modifiers are hypothesised to modify or select various attributes of a ‘noun concept’ (to be explained). Thus, the working theory here is that a higher degree of modification by nominal modifiers will indicate how noun-like a given lexeme is.

14The tagged modifiers and their respective frequencies with Time are shown below; modifiers can, of course, co-occur.

Table 2: Collocational frequencies between Time head and nominal modifiers

Table 2: Collocational frequencies between Time head and nominal modifiers

15A separate table is compiled which tabulates how frequently a given lexeme occurs with a given modifier. Words with a sample size <5 are dropped to ensure enough data is present for a reliable analysis. The resulting table contains 41 lexemes. It will be clear in the experimental results which lexemes are included, but the table below gives some preliminary impression, containing common words like םוי ‘day’ and תע ָ ה ‘now’.

Table 3: Excerpt of Time head x modifier co-occurrence matrix (original: 41x9) Using the Pandas Python package and a spreadsheet of individual observations, a tabulation is compiled which counts how frequently a given head word lexeme is used with a given modifier type. The result is a table with 41 rows and 9 columns (not including the ‘head’ column), containing a total of 369 data points. The excerpt below shows the first three rows of that table.

Table 3: Excerpt of Time head x modifier co-occurrence matrix (original: 41x9) Using the Pandas Python package and a spreadsheet of individual observations, a tabulation is compiled which counts how frequently a given head word lexeme is used with a given modifier type. The result is a table with 41 rows and 9 columns (not including the ‘head’ column), containing a total of 369 data points. The excerpt below shows the first three rows of that table.

16This dataset contains 369 individual data points (41 rows by 9 columns). We need a method that can ingest the data and identify tendencies and patterns in it. One might imagine the task of a biologist, who collects samples of butterflies, and groups them together based on their shared features. When the number of features and samples becomes quite large, it is helpful to approach the problem mathematically. Imagine, then, a sheet of paper with a horizontal and vertical line drawn down the middle (Cartesian plane). The lines separate the paper into four corners, or quadrants. Then, we select those features that tend to be most distinctive of particular groups and assign them to one of the four quadrants. For each sample in our dataset (e.g., butterflies or words) we place a dot in the quadrant which most strongly matches the sample’s own tendencies. So, a sample in the very corner of the paper would represent a case that is strongly aligned with the features we assigned to that quadrant, whereas a sample closer to the centre represents a case with no strong tendencies.

  • 18 For a mathematical description, see Abdi and Williams (2010).

17The method described above is a strategy used in statistics for clustering samples. In the illustration, we have four corners. But remember that our dataset contains nine features, the modifiers. How exactly should each feature come to be associated with a given corner? What weight should each feature have on placement on the sheet? In statistics this problem is solved using what is called ‘dimensionality reduction’. In the case of this dataset, each of the nine features represents a dimension. But we need to somehow compress those nine features into a two-by-two space, so as to know which quadrant should be associated with what and therefore achieve useful clusters. Rather than manually (and subjectively) deciding which features influence placement in the space, we can use an unsupervised clustering method. Principal Component Analysis (PCA) is one such method, which mathematically isolates those features that produce the greatest separation amongst the various samples (Rokhlin, Szlam, and Tygert 2010; Pedregosa et al. 2011). To return to the previous analogy, imagine that 50 percent of the butterflies have eye-shaped patterns on their wings, while 50 percent have stripes. In this case, the sharp divide along ‘wing pattern’ might lead to that feature being heavily weighted in the PCA model, so that butterflies with stripes are placed on one side of the ‘paper’, with those with eyes on the other side. The power of PCA is that several discriminative features can ‘cooperate’ together simultaneously. So, for example, ‘wing pattern’ combined with ‘proboscis length’ might yield stronger separation than ‘wing pattern’ by itself. Using methods from Linear Algebra, PCA is able to isolate such patterns.18

18PCA is well-suited for investigating parts of speech in time adverbials, a semantic function which is known to display mixed behaviour with respect to word class. We will apply the same approach to the time adverbial dataset. The model not only allows us to ask, what clusters exist in the dataset? but also, which features are most distinctive of particular groups? Afterwards, we shall ask the question of why?

3.0. Statistical Analysis

19The first part of this section details the results of the PCA experiment. The second half consists of digging into the results by looking at individual examples and texts. Afterwards, we ask what these results mean for the discussion on parts of speech.

  • 19 Normalisation involves adjusting the counts so that they are comparable. For instance, some words o (...)

20The analysis begins by normalising the raw frequency data contained in Table 3 (above) as a proportion of a given Time head’s total observed modifiers.19 A proportion is simply a decimal that can be read as a percentage when multiplied by 100.

Table 4: Excerpt of Time head x modifier proportion matrix (original: 41x9) The data from Table 3, containing the raw co-occurrence frequencies, are converted to a table containing proportions using a Pandas sum (across rows) and divide operation (across columns). The resulting decimal values across each row sum to 1, which corresponds with 100 percent of a given lexeme’s attested modifiers. For example, the ‘cardinal’ feature with י ֹ םו comprises 0.07 in 1 of all the modifiers observed with י ֹ םו , or 7 percent.

Table 4: Excerpt of Time head x modifier proportion matrix (original: 41x9) The data from Table 3, containing the raw co-occurrence frequencies, are converted to a table containing proportions using a Pandas sum (across rows) and divide operation (across columns). The resulting decimal values across each row sum to 1, which corresponds with 100 percent of a given lexeme’s attested modifiers. For example, the ‘cardinal’ feature with  י ֹ םו  comprises 0.07 in 1 of all the modifiers observed with  י ֹ םו , or 7 percent.

21It is worth pointing out that some of these modifiers co-occur in the same phrase. For instance, the phrase, םויב ְ אוהה ‘on that day (literally ‘in the day the that’)’ contains two separate modifiers: the definite article and the demonstrative. Each of these modifiers is counted as an individual instance of modification, each adding +1 to their raw frequencies in Table 3 (above). Another approach might count this combination separately, as a ‘definite + demonstrative’ pattern. This approach would identify whole constructions which are associated with particular Time heads, an interesting experiment in its own right. On the other hand, counting each modification separately can be seen as a measurement of compatibility between a given lexeme and modifier. There is theoretical warrant for this choice—we want to test the degree to which each of the modifiers alters or selects various aspects of a word’s meaning and, by extension, determine which aspects are most relevant. From this point of view, a value in Table 4 represents the proportion that one modifier comprises of a lexeme’s total attested ‘modifiability’.

  • 20 The lexicon parts of speech values come from the BHSA’s ‘sp’ feature, which is derived from Köhler (...)
  • 21 These are the loading scores, which are the eigen vectors × √eigen values following the discussion (...)

22The unabridged data from Table 4 are fed for analysis into the PCA algorithm, which identifies which features and feature combinations produce the strongest separations in the dataset. The resulting graphs (Figure 2, below) should be interpreted with respect to space, where closer samples are more closely related based on input features. The placement of each sample is similar to the butterfly and paper illustration given in the previous section. Location near an edge represents strong tendencies; location near the centre represents weaker tendencies. The left-hand graph plots the samples using parts of speech from the database and lexicon (subs=noun, inrg=interrogative, adjv=adjective, advb=adverb); this plot shows how the classical parts of speech compare with their PCA classification.20 The right-hand plot shows the samples as Hebrew text. The blue arrows show which features have influenced each lexeme’s placement on the graph, with longer arrows indicating stronger influence.21 Table 5 contains average proportions of lexemes with a nominal modifier versus without one based on the samples’ placement in the PCA graphs. It shows, for instance, that the words to the left of x=0 are used with null modification (Ø) 96 percent of the time.

Figure 2: PCA analysis of head word tendencies The PCA analysis is performed using the scikit-learn Python package. The function takes in the modification proportions per head lexeme, which contains nine dimensions, and reduces the dimensions to two vectors, each of which can then be assigned as the X and Y coordinates of the respective head lexeme. The first vector (X-axis) accounts for the most variation in the dataset, and the second (Y-axis), the second-most. Together the two components account for about 80 percent of variance in this sample.

Figure 2: PCA analysis of head word tendencies The PCA analysis is performed using the scikit-learn Python package. The function takes in the modification proportions per head lexeme, which contains nine dimensions, and reduces the dimensions to two vectors, each of which can then be assigned as the X and Y coordinates of the respective head lexeme. The first vector (X-axis) accounts for the most variation in the dataset, and the second (Y-axis), the second-most. Together the two components account for about 80 percent of variance in this sample.

23The PCA graph reveals a clear separation of terms along a gradient.

Table 5: Mean proportions for null (Ø) versus nominal modifiers by placement along x-axis

Table 5: Mean proportions for null (Ø) versus nominal modifiers by placement along x-axis

24Table 5 shows that terms to the right of 0 occur with a nominal modifier on average 96 percent of the time, and on the left of 0 with null modification also with 96 percent on average. This reflects the strong tendencies seen visually in the graph. In cognitive terms, the items on the extreme ends of the plot represent potential prototypes. The separation confirms the intuition that Time adverbial heads exhibit a broadly binary tendency between nouns and adverbs.

  • 22 This is a crude description of PCA, but gets the point across. Abdi and Williams (2010) present a m (...)

25In order to determine which features have influenced the model most strongly, we can look at the loading scores (indicated by the blue arrows). One might imagine these various values as measurements of ‘pressure’, pushing the samples into different directions. The features which exert the most ‘pressure’ in a given direction, combined with how often a word occurs with said feature, end up determining the word’s placement on the graph. The ‘amount of pressure’ in actuality corresponds to the separation (variance) a given feature accounts for in the dataset, so that features with more discriminatory power are ‘weighted’ more heavily.22 The X (horizontal) and Y (vertical) values of the loading scores are provided below in descending order:

Table 6: PCA Loading Scores (rounded to two decimal places)

Table 6: PCA Loading Scores (rounded to two decimal places)

26For example, the definite feature exhibits positive values for both horizontal and vertical axes (0.17 and 0.21, respectively). Therefore, words which frequently occur with definite articles will tend to be pushed toward the upper right-hand corner of the graph.

  • 23 The loading score includes direction from the eigen vector and magnitude from the eigen value. The (...)

27Null modification (Ø) is by far the strongest contributor to the separation (see also the length of its arrow), pushing samples associated with it to the left on the horizontal axis (X loading is −0.48).23 The result is a tight cluster of lexemes which often correspond to prototypical adverbs, yet with several interesting cases. These lexemes and their proportional tendencies are noted below.

Table 7: Adverb-like lexeme cluster (PC1 < 0) (with modification proportions per lexeme shown across rows; each row sums to 1, i.e., 100 percent)

Table 7: Adverb-like lexeme cluster (PC1 &lt; 0) (with modification proportions per lexeme shown across rows; each row sums to 1, i.e., 100 percent)

28Note that the Ø column contains the largest proportions; for example, 99 percent (0.99) of תע ָ ה ‘now’, 95 percent of לוע ָ ם ‘forever’, and 99 percent of א ָּ ז ‘then’ go unmodified. Several Ø values are at 100 percent ( ת ָ מ ִ די ‘continuously’, כ ּ ן ‘thus’, מ ָ ח ָ ר ‘tomorrow’, etc.). The words in the sample are certainly semantically diverse. Deictic terms like מ ָ ח ָ ר ‘tomorrow’ communicate positions in time, while terms like ת ָ מ ִ די indicate durations of time. Interrogative particles are also included ( א ָ ן ‘where/when’ and מ ָ ית ‘when’). These differences will be touched on briefly in the linguistic section. Of special note are those lexemes which some lexicons consider nouns, namely לוע ָ ם ‘forever’, מ ָ ח ֳָ תר ‘next day’, and ֶ עגר ‘moment’ (see Clines 2009, 315, 216, 414). These cases will be dealt with individually below. For now, it is important to note that the effect of rarely modifying a word sharply distinguishes it from other terms.

29The other side of the graph contains lexemes broadly considered to be nouns. The placement of these terms has been defined primarily by the frequent presence of some form of modification. There are notably two subcategories of ‘nouns’ indicated in the graph: those frequently modified by definite articles and demonstratives versus those modified by genitives, plurals, and suffixes. This is a meaningful separation between more-singular and more-plural items. The two groups will be now be described.

30The group modified by definite articles and demonstratives is pushed into the upper right-hand corner of the plot. Definiteness is the second strongest influence on the model. The cluster and its modification proportions per lexeme is shown below.

Table 8: Noun-like lexeme cluster with strong definite/demonstrative influence (PC1 > 0, PC2 > 0) (with modification proportions per lexeme shown across rows; each row sums to 1, i.e., 100 percent)

Table 8: Noun-like lexeme cluster with strong definite/demonstrative influence (PC1 &gt; 0, PC2 &gt; 0) (with modification proportions per lexeme shown across rows; each row sums to 1, i.e., 100 percent)
  • 24 E.g., Gen. 7.11; Deut. 3.14; Josh. 20.6; 1 Sam. 7.6; 2 Sam. 3.38; 2 Kgs 20.1; Isa. 28.5; Ezek. 39.1 (...)

31Note that diverse modifiers are indeed compatible with these words, but there is a clear trend in this group to be modified more often by the definite article. The cooperation with the demonstrative (see blue arrows in) may be explained by the attributive adjunct construction with -ְ ה , so . םויה ְ זה ֶ ה24 See also that the important time word םוי ‘day’ has the definite article as a lower proportion (35 percent) of its total modification compared with the rest of the cluster. This aligns with its intermediate placement in the graph with respect to the two noun sub-clusters.

32The second noun-like cluster is primarily influenced by plural modifiers, genitive (construct), and suffixes.

Table 9: Noun-like lexeme cluster with plural, genitive, suffix (PC1 > 0, PC2 < 0) (shown with modification proportions across rows; each row sums to 1, i.e., 100 percent)

Table 9: Noun-like lexeme cluster with plural, genitive, suffix (PC1 &gt; 0, PC2 &lt; 0) (shown with modification proportions across rows; each row sums to 1, i.e., 100 percent)
  • 25 E.g., Exod. 12.14, 17, 42; 40.15; Lev. 18.18; Judg. 16.30; 1 Sam. 17.33; 2 Sam. 18.18; 1 Kgs 18.12; (...)

33In what way do these modifiers interact? Plurals and suffixes frequently co-occur in such constructions as וענ ּ ךיר ‘your youth’ or יח ָ וי ‘his life’, and דל ּ ר ּ ת ּ כי ֶ ם ‘unto your generations’, all terms related to life cycles.25 The PCA graph also places מ ָּ ו ֶ ת ‘death’ nearby, but on the basis of its co-occurrence with genitive modification. Typically, the genitive item is a person, as in ְ א ַ ח ֲֵ מ יר ֹ֣ תו ְ א ַ ב ְָּ הר ָָּ ם ‘after the death of Abraham’ (Gen. 25.11), ְ ע ַ מ־ד ֹ תו ְ ה ַ כ ֵֹּ ה ֣ ה ן ַ ג ָּ ד ָֹּ ל ‘until the death of the high priest’ (Num. 35.25), ְָּּ בו ְ מ ֹ֤ תו ְ א ָּ ב ִ ו הי ְ א ִ מ ָָּּ ה ‘and upon the death of her father and mother’ (Est. 2.7). There is a subtle similarity between these cases and the suffix group, in that both utilise modifiers to relate the life cycle term to a particular individual. Here we can see how the statistics and semantics are related. The meaning of the head words, in this case life cycles with respect to person X, influence which modifiers they tend to take.

  • 26 In Modern Hebrew the distinction between adjectives and adverbs is blurry. A similar dynamic appear (...)

34Summarising the three main groups, we can see that the model successfully distinguishes subtle semantic tendencies on a broad scale by identifying distinctive behaviours: modified versus non-modified, singular versus plural, and even life cycles with distinctive modifiers. The model may also show the shortcoming of traditional parts of speech. Diverse lexicon tags of noun (subs), interrogative particles (inrg), and adverbs (advb) are found grouped together in the ‘adverb’ cluster. The ‘noun’ side is also mixed, containing one adjective (ִ אר ׁ ןוש ‘first’).26 Secondly, several terms sit at intermediate positions along the x-axis, especially ֶ גר ְ ע ‘instant’, מ ָ ח ְֳָ תר ‘next day’, and ליל ְָ ה ‘night’. And terms like ְ לוע ָ ם ‘forever’, which some lexicons consider nouns, are grouped with the adverbs.

4.0. Analysis of Select Words

35In order to understand why the aforementioned words are classified in intermediate positions, we must look at their behaviour in context. For the intermediate words, text samples are retrieved and analysed below with an eye to their semantics.

36Table 10 contains the full sample set for ֶ עגר , a relatively small set (n=7). ֶ עגר is placed slightly on the ‘adverb’ side of the graph.

Table 10: Sampled sentences for ֶ גר ַ ע

Table 10: Sampled sentences for ֶ גר ַ ע
  • 27 As Croft (1990, 246) recognised with constructions like ‘day’ and ‘five times’.

37Three of the text examples are used without any kind of nominal modifier, three employ quantifiers, and one has a plural ending. Given the small sample size, we should withhold definitive judgments about this term. However, a number of observations are pertinent. Even though this word has a slight majority of nominal modifiers (4/7), the algorithm has situated it on the adverb side due to the stronger influence of the null feature. Another reason is the weaker influence of quantifiers. This is not just an accident of the PCA analysis; the feature weights are determined across the whole dataset by the degree, and thus predictive power, of their variability. Furthermore, an abstract word like ‘instant’ does not seem to fit the semantics of a prototypical noun.27 The data thus tentatively indicate a term that sits somewhere on the border between a noun and adverb.

  • 28 Exod. 18.13; 32.30; Num. 17.23; Judg. 9.42, 21.4; 1 Sam. 11.11; 18.10; 31.8; 2 Kgs 8.15; Jer. 20.3; (...)

38מ ֳ ח ָָ תר ‘next day’ appears 26 times in the sample. There are 20 cases of null modification and 6 cases of nominal modification. Note that 11 null cases not shown are identical to the first example with . היו ִ י28

Table 11: Sampled modifiers with מ ָּ ח ֳָּ תר

Table 11: Sampled modifiers with  מ ָּ ח ֳָּ תר

Table 13: Selected samples of מ ָ ח ֳָ תר (continued)

Table 13: Selected samples of  מ ָ ח ֳָ תר  (continued)
  • 29 The original status of some definite articles reflected only in the vocalisation is debated. The is (...)

39In 20 of 26 cases (77 percent), מ ָ ח ֳָ תר appears without modification, reflecting behaviour similar to its lexical cousin מ ָ ח ָ ר ‘tomorrow’ (100 percent Ø). There are six cases of explicit nominalisation: five construct relations and one (vocalic) definite article. Specifically, the construct cases construe מ ָ ח ֳָ תר as a specific ‘next day’ after the Passover, Sabbath (2x), second month, and ‘that day’ (ְ ה ְ םוי ְ אוהה ). The use with the (apparent29) definite article, מל ָ ח ֳָ תר ‘on the next day’, represents another possible nominalisation of the term.

  • 30 But cf. גָּלְיָּת ‘Goliath’ (1 Sam. 17.4); see GKC (§80g).

40Could it be that מ ָ ח ֳָ תר is simply a noun that happens to be used less frequently with noun modifiers? This is, of course, a possibility. The ָָּ ת - suffix is consistent with endings (typically feminine30) found on nouns such as חנ ֲ ל ָ ת ‘inheritance’ (Ps. 16.6), ז ִ מ ָ ר ְ ת ‘song/strength’ (Exod. 15.2; Isa. 12.2; Ps. 118.14), or י ִ ת ָ תר ‘remainder’ (Jer. 48.36) (GKC §80g; Bauer and Leander 1922, §62v). Yet, the behaviour of מ ָ ח ֳָ תר differs significantly from that of the other regular nouns in the PCA model, resulting in its placement.

  • 31 See also the case of לְמִן־הַיֹום ‘since the day of’ (Exod. 9.18), but here with the definite artic (...)

41Furthermore, the collocational preference for מ ִ ן ‘from/ since’ closely resembles that of other adverb-like words. The term occurs with ְ מ ִ ן in 23 of its 26 sampled cases. In Lev. 23.16 we find the telling phrase, דע ְ מ ִ מ ָ ח ֳ תר ‘until the day after’, where מ ְִ ן ‘from’ and מ ָ ח ֳָ תר ‘next day’ appear to function as a single lexical item. This pattern reflects a larger tendency in BH to fuse מ ְִ ן to adverb items that reflect orientation (van der Merwe, Naudé, and Kroeze 2017§ ,39.14.2). For example, the term ץוח ‘outside’ is found with an appended מ ִ ן in 64 of its 164 occurrences ( מ ִ ץוח ). Similar to מ ִ וח ְ ץ , מ ְִ מ ָ ח ֳָ תר is also occasionally found with another preposition: א ֶ מ־ל ִ ח ַּּ֨ ץו ְ ל ֶַ מ ַ ח ֲ נ ֶָּ֜ ה ‘to outside the camp’ (Lev. 4.12; see also Lev. 4.21; 6.4; 10.4, 5; 14.3, et al.; Num. 5.3–4; 15.36, et al.). This same pattern of preposition + מ ִ ן may also be found in the following constructions: ל ְ מ ִ ב ֵ תי ְ ל ַּ פ ָּ ר ֹּ כ ֶ ת ‘and within the veil’ (e.g., Num. 18.7), מל ִ צק ּ ה ְ שה ָ ימ ְ ם ‘from one end of heaven’ (e.g., Deut. 4.32), מל ָּ וחר ְ ק ‘for a great while’ (e.g., 2 Sam. 7.19), מל ִ תחת ְ מל ִ תורגס ‘under the borders’ (e.g., 1 Kgs 7.32), מל ִ מי ּ י ְ ק ֶ ד ֶ ם ‘from times of old’ (e.g., 2 Kgs 19.25).31 In these examples, מ ִ ן coordinates with the adverb item to convey orientation. The result is that even obvious nouns such as יב ִ ת ‘house’ and ק ָ צ ֶ ה ‘end’ become construed as adverbs with a prepositional sense.

  • 32 Thanks to Chip Hardy for pointing out these possibilities in relation to another similar case, לְעֹ (...)

42It is telling, then, that in two of the six modified cases, מ ָ ח ֳָ תר appears without מל ָ ח ֳ תר םויה ְ אוהה  : מ ִ ן ‘on the day after that day’ (1 Chron. 29.2) and מל ָ ח ֳָ תר ‘on the following day’ (Jon. 4.7). These may show that the use with a nominal modifier represents a different conceptualisation of מ ָ ח ֳָ תר , one which is not incompatible with a combination of ְ מ ִ ן (as in Lev. 23.16), but which does not attract it. To put it another way, in view of its use with nominal modifiers, מ ָ ח ֳָ תר is arguably construed less like an adverb (with its habitually attached preposition) and more like a noun.32

43ל ַ י ְ ל ָּ ה ‘night’ is the most frequent intermediate word in the graph, with 65 percent of its 131 sampled forms collocating with nominal modifiers. This is compared with an average nominal modification of 93 percent on the noun-side of the graph.

Table 14: Frequency and proportion of ל ַ י ְ ל ָּ ה with modification type

Table 14: Frequency and proportion of  ל ַ י ְ ל ָּ ה  with modification type
  • 33 N.B. that this table contains a higher marginal total than the previous since modifiers can co-occu (...)

Table 15: Sampled modifiers with ל ַ י ְ ל ָּ ה33

Table 15: Sampled modifiers with  ל ַ י ְ ל ָּ ה33

Table 16: Selected samples of ל ַ י ְ ל ָּ ה

Table 16: Selected samples of  ל ַ י ְ ל ָּ ה
  • 34 The closest would be בֹּקֶר with 18/119 (15 percent) null modified uses, a substantial difference f (...)

44Again, there is the temptation to consider the null modifications of ליל ָ ה simply as unmodified uses of a noun. And, indeed, the PCA analysis does place ליל ָ ה on the noun side. Yet other terms, like םוי ‘day’, ב ּ ק ֶ ר ‘morning’, and ע ֶֶ בר ‘evening’, which are likewise well-represented in the dataset, do not have nearly the number of null-modified cases.34 ליל ָ ה clearly has a distinctive profile.

  • 35 Thanks to Chip Hardy for pointing this out.
  • 36 This is a common phenomenon in world languages (Haspelmath 1997).
  • 37 Note forms without the ending, e.g., כַלַָיִל ‘like the night’ (Isa. 16.3) and the construct לֵ֣יל (...)
  • 38 See Bybee, Perkins, and Pagliuca (1994, 15–16) on the comparison of English will, shall, and be goi (...)

45As noted above with מ ָ ח ֳָ תר , the habitual collocation with an adverbialising particle seems to discourage the use of noun-type modifiers. This may also be the case with ליל ָ ה . Some have proposed that the word’s ָָּ ה - ending originated in the heh locale found on locative adverbs, e.g., ַ א ְָ צר ָּ ה ‘to the ground’ (Gen. 24.52) (GKC §90f; Bauer and Leander 1922, 528).35 Similar to the way in which locative prepositions, such as -ְ ב ‘in’ are metaphorically extended to location-in-time function, e.g., ב ַ ב ֹּ ק ֶ ר ‘in the morning’ (van der Merwe, Naudé, and Kroeze 2017, §39.1.4),36 the heh locale might have been appended to an original proto-Semitic root lyl ‘night’ to indicate ‘at night’.37 This would explain the mixed behaviour of ליל ָ ה , a word which one might otherwise expect to share a symmetry with םוי ‘day’. Perhaps the 65: 35 ratio of noun-versus null-modified cases reflects a vague semantic memory of ליל ָ ה ’s original form—a memory that is increasingly forgotten, as happens in diachronic change.38

46The term לוע ָ ם ‘forever’ is often classified by Hebraists as a noun, but in our model it clusters with the apparent adverbs. לוע ְָ ם occurs with null modification in 95.5 percent of its samples (170/178) with 8 exceptions:

Table 17: Selected nominalised samples of לוע ָ ם

Table 17: Selected nominalised samples of  לוע ָ ם
  • 39 The others are Isa. 45.17; 51.9; Ps. 77.6; Eccl. 1.10; Dan. 9.24.
  • 40 Perhaps a clue for the origin of these plurals can be found in their use with physical/spatial noun (...)

47Five of the eight nominalisations in Time are plural endings. Besides Time phrases, plural endings with לוע ָ ם appear seven other times in the HB; for example, צ ָּ רו ְ ע ֹ לו ִָּ מ ֶ םי lit. ‘rock of eternities’ (Isa. 26.4), ִ כ ֣ מי ֵ י ְ ק ֶָ ד ֶ ם ְּ ד ָֹּ ר ֹ תו ְ ע ֹ לו ִָּ מ ָ םי ‘as in ancient days, generations of eternities’ (Isa. 51.9), מ ַ ל ְ כ ֶּּ֥ תו ְ כ ָּ ע־ל ֶֹּ ל ִָּ מ ָ םי ‘kingdom of all eternities’ (Ps. 145.13).39 Semantically, this use of the plural belongs to the “plural of extension” (GKC §124b), which refers to some “characteristic or condition” in abstract nouns (van der Merwe, Naudé, and Kroeze 2017, §24.3.3.6). In this case, the “characteristic” is the continuous nature of the duration.40 This same use of the plural was seen on the aforementioned life cycle terms, namely וענ ִ םיר ‘youth’, יח ִ םי ‘life’. Other examples include ְ ב ְ ח ִ םיר ‘youth’, ְ קז ֻ נ ִ םי ‘old age’, תב ּ לו ִ םי ‘virginhood’, and תולולכ ‘betrothal’ (GKC §124b–d).

  • 41 See the note in the methodology section about excluding complex phrases. This is done to ensure acc (...)

48Three times in the sample ְ לוע ָ ם appears with the definite article. Four other definite cases were excluded from the sample due to phrase complexity:41 מ ִ ה־ן ָ לוע ָ ם ְ דעו ְ ה ָ לוע ָ ם ‘from eternity to eternity’ (Ps. 41.14; 106.48; Neh. 9.5; 1 Chron. 16.36). לוע ָ ם is also used with the article at least three times outside the category of Time. For instance, Eccl. 3.11 makes the nominal construal clear: ג ַ֤ ם ְ א ֶ ה־ת ָּ ע ֹּ ל ָּ ם ְ נ ַָּ ת ֣ ן ְ ב ְ ל ִ ב ָָּ ם ‘Even the eternity he has placed in their hearts’. Dan. 12.7 has the genitive ב ְ ח ֵ֣ י ְ ה ָּ ע ֹ ו ָּ ל ָ ם ‘in life of the eternity’ְ. Another case in Ps. 133.3 uses דע within an adjectival expression: ח ַ י ִֶ םי ְ ע ַ ה־ד ָּ ע ֹ ו ָּ ל ֶ ם ‘life unto the eternity’.

  • 42 This also seems true of the corresponding Greek term αἰών ‘eon’, see Hebrews 1.2; 9.26; 11.3. Thank (...)
  • 43 Thanks again to Chip Hardy for this idea.

49Thus, in the vast majority of cases, לוע ָ ם appears without any nominal modifiers, a characteristic which is mathematically distinctive. Yet, in a number of cases, the word does appear with both definites and plurals. What is one to make of these cases with respect to parts of speech? First, the occasional deviation from the strong tendency for null modification (>95 percent in Time) shows that modification behaviours are not hard and fast rules. Semantically speaking, the deployment of a nominal modifier profiles with a more noun-like construal of the term. Second, the data above show that later books (Daniel, Chronicles, late Psalms) account for many, though not all, of the modified cases. Given that ע ֹ לו ָּ ם morphs into a noun meaning ‘world’ in Rabbinic Hebrew (Sáenz-Badillos 1997, 200),42 this may show a gradual shift in speakers’ perception of ע ֹ לו ָּ ם as increasingly noun-like.43

  • 44 Some of the other cases not mentioned here actually contain adjectival modification, e.g., יֹוםְטֹו (...)

50The PCA groupings also allow us to search for other word behaviour that contradicts intra-cluster tendencies. A query across the ‘noun-cluster’ terms for null modification reveals a pair of illuminating examples concerning םוי ‘day’, which otherwise co-occurs with nominal modifiers in 99 percent of the sample (1339/1351):44

Table 18: Select null-modified samples of םוי ‘day’

Table 18: Select null-modified samples of  םוי  ‘day’
  • 45 Josh. 24.2; Isa. 42.14; 57.11; 63.19; 64.3; Jer. 2.20; Prov. 8.23.

51These examples closely resemble in meaning and form the construction מ ּ לוע ָ ם ‘from eternity’, which likewise frequents the prophets.45 The term also bears resemblance in form to מ ִ מ ָ ח ֳָ תר ‘since the next day’. Here םוי is used within a pattern of null-modification that strongly resembles that of particle-like adverbs. These two cases reinforce the argument just made with respect to לוע ָ ם , namely, that the strong tendencies of a word do not constitute hard and fast rules. Rather, individual words are always subject to alternative construals depending on what kinds of contexts they are slotted into.

5.0. Cognitive Linguistic Analysis

52The quantitative data sheds light on the fact that word behaviour is just that: behaviour. Words do not woodenly follow abstract classifications; instead, they exhibit strong and weak tendencies, habitual collocations with other lexical items, and/or semantics dependent on functional context. These facts have a number of implications for linguistic theory and BH.

  • 46 A handful of terms possess adverb suffixes, such as ָָּם- as found in יֹומָּם ‘by day’; see van der (...)

53Classically, grammars have tended to treat time adverbials more rigidly under three discrete groups: adverbs, adverbial accusatives, and temporal prepositional phrases (e.g., GKC §100, 118; Waltke and O’Connor 1990, §§10.2.2c, 11.2, 39.3.1h; Joüon and Muraoka 1996, §102a–c, 126i, 133). Adverbial accusatives and prepositional phrases are typically noun-based. Yet, as seen already, the distinction between an adverb and noun in BH is not always clear (see also Waltke and O’Connor 1990, §39.3.1a; van der Merwe, Naudé, and Kroeze 2017, 380–81). BH lacks a productive morphology for adverbs, unlike languages such as English with morphemes like -ly.46 But even English and many languages do not always clearly distinguish nouns and adverbs (Cuzzolin, Putzu, and Ramat 2006). One reason for this may be that adverbial function is already an abstract, peripheral role compared with other sentence arguments (Ravid and Shlesinger 2000, 335–37; Croft 2001, 97). As a result, adverbs tend to be sourced from a variety of word classes, resulting in a heterogenous makeup (Ramat 2011). In BH, the most common sources for adverbs are nouns and adjectives (Ravid and Shlesinger 2000, 339–44; van der Merwe, Naudé, and Kroeze 2017, 380–81).

  • 47 Croft (2001, 65) notes Nootkan, Salishan, Iroquoian, Philippine, and Polynesian languages as oft-ci (...)
  • 48 Geeraerts and Cuyckens (2012, §2) write: “The perspectival nature of linguistic meaning implies tha (...)

54The absence of a neat and tidy word class is not unique to adverbs, but is endemic to parts of speech in languages around the world. Some languages, for instance, lack a clear morphological distinction between nouns and verbs.47 These problems have led cognitive linguists to abandon the notion of universal, language-inherent parts of speech (Croft 2001, 63–107). Instead, cognitivists propose that semantic meanings, not language categories, are the common elements across human language— though always subject to individual construal.48 Concepts are free to be mapped onto an endless range of constructions (Croft 2001, 86–87). Thus, cognitivists attempt to explain parts of speech as word tendencies that reflect underlying concepts. Furthermore, concepts as such can have a range of fuzzy behaviour, with prototypes (‘best examples’), family resemblance, and gradient tendencies (Lakoff 1987, 41).

55Croft (1990), in particular, analyses parts of speech in terms of their conceptual basis; his analysis is based on a typological approach that respects language differences. He acknowledges the strong, prototypical behaviour found among traditional nouns, verbs, and adjectives. Underlying these tendencies are two classes of concepts. The first constitutes semantic notions of objects (in space), properties (quantity), and actions (in time). Croft derives the second group from Searle’s ‘propositional acts’ of reference (identifying an entity), modification (enriching an entity), and predication (ascribing a property) (Croft 1990, 248–51, 253). These groups and their correlation with the three prototypical parts of speech are illustrated below:

Table 19: Croft’s Major Propositional Acts (modified from Croft 1990, 248)

Table 19: Croft’s Major Propositional Acts (modified from Croft 1990, 248)

56The chart can be interpreted spatially, so that constructions in language can point to any region or stretch within the space. For example, the English pattern -ing spans the events with reference, e.g., running, and events with modification regions, e.g., running man) (Croft 2001, 98– 99).

57But there are many constructions which do not fit prototypical noun, adjective, and verb behaviour. Notably for our purposes, Croft (1990, 246) gives the examples of day and time (e.g., five times) as non-prototypical nouns. Other drip-bucket word classes include adverbs, prepositions, numbers, articles, auxiliaries, demonstratives, etc. Croft (1990, 252) argues that humans intuitively interpret all such categories against the background of a real or metaphorical spatial dimension (‘background dimension’). Secondary word classes thus serve to locate and quantify real or abstract objects in said space (Croft 1990, 256–60). These dynamics are illustrated below (Croft 1990, 279–80):

Table 20: Croft’s Minor Propositional Acts (adapted and condensed)

Table 20: Croft’s Minor Propositional Acts (adapted and condensed)

58Croft places locative time adverbials within the group that situates in a physical dimension; this derives from the fact that locative time represents a metaphorical extension from space to time in all world languages (Croft 1990, 257; Haspelmath 1997; see also Fillmore 2002).

59Croft’s categories explain why the results obtained in the PCA experiment (, above) are not merely interesting side-effects of coincidental word behaviour. Following Croft (1990, 253), the noun concept represents objects in a real or imagined physical space. As such, nouns have a position, dimension, size, number, etc., all possibly metaphorical. These attributes are therefore available for situating and selecting by modifiers, which correspond with the minor propositional acts (above, Table 20). This interaction between the semantic-pragmatic features of nouns and modifiers produces statistical patterns that the PCA model identified. The ‘situating’ function of the definite article and demonstrative had the strongest separation effect. A subgroup of ‘countable’ nouns was separated by their collocation with plurals and quantifiers. The construct and suffix modifiers played a smaller role in the model, but, like the definite article, indicated situated ‘objects’.

  • 49 E.g., עַד־אָּן/אָּנָּה ‘how long? ’ (Exod. 16.28; Num. 14.11; Josh. 18.3, et al.); see BDB 33. This (...)

60While the ‘nouns’ in the PCA model are identified by their collocation with modifiers, the ‘adverbs’ in the model are identified by the lack of modification. Using Croft’s categories, we can now see that in fact adverbs themselves belong to a category of ‘modification’. Indeed, if the various modifiers had been treated as individual words with the potential for their own modification, we would likely see null-modified behaviour in a way similar to that of the ‘adverbs’ of the PCA graph. Breaking down the various items classified as adverbs, we see a group of durations, such as ת ָ מ ִ די ‘continuous’, נ ּ חצ ‘enduring’, לוע ָ ם ‘forever’, which correspond with Croft’s modifiers that select ‘quantity’ within events (similarly ֶ עגר ‘moment’); a group of irrealis indicators that situate events in a mental space, like א ָ ן ‘where’,49 מ ָ ית ‘when’; temporal deictic markers that situate events in space, like תע ָ ה ‘now’, א ְָ ז ‘then’, ְ מ ָ ח ָ ר ‘tomorrow’, ה ּ נ ָ ה ‘to here’, א ֶ לומת ‘yesterday’; and prepositional situators like ט ֶֶ םר ‘before’ and מ ָ ח ֳָ תר ‘next day’. It is likely that a model which includes variables for collocating verb tense, aspect, and modality would be able to subdivide these adverbs as the nouns were.

61The fuzzy behaviour observed with words like ֶ עגר ‘instant’, מ ָ ח ֳָ תר ‘next day’, ליל ָ ה ‘night’, לוע ָ ם ‘forever’, and even םוי ‘day’ might be attributed to the interaction between adverbial function, which itself is a modifier, and nominal construal, which is contained within the modifier. This containment reveals a recursive and hierarchical structure for part of speech concepts. But the levels of recursion should not be viewed as wholly self-contained. Indeed, as noted by Croft, words like םוי do not fit prototypical noun semantics. They are not physical objects. By this logic, the noun category within adverbials seems to represent a kind of analogical extension by which the very concept of an object is stretched to encompass phenomena such as the cycles of the sun. Perhaps this extension is accommodated by the metaphor of time as a line in space, and thus locative time as ‘places’ in that space (Haspelmath 1997, 23–42). If true, the fuzzy statistical behaviour is simply a product of the already abstract nature of time-as-object.

62What do these results mean for the broader theory of parts of speech in BH? These data ought to at least cast some doubt on generative descriptions of surprising word behaviour. For example, in a recent article on parts of speech, Miller-Naudé and Naudé (2017, 273–308) examine the behaviour of בוט ‘good’ in cases where it is modified by the definite article. For instance:

2 Sam. 19.28 ו ַ ע ֲ ש ֵֶּ֥ ה ְ ה ַ טּ ָֹ בו ְ ב ְ ע ֵ ני ֶֶ ךי Do the good in your sight.
2 Sam. 14.17 ל ִ ש ְַּ֨ מ ַֹּ ע ה ַ טּ ֹ֣ בו ְְ הו ָָּּ ר ָ ע to hear the good and the bad
Ps. 85.13 ג ַ י־ם ְ֭ וה ָּ ה י ִֵ ת ֣ ן ה ַ טּ ָֹ בו Also may Yhwh give the good

63The debate centres on whether these uses of בוט (and other similar cases) represent construals of an adjective as a noun (‘the good’). This possibility poses problems for the generative model, which maintains parts of speech as a category inherent to a word. Miller-Naudé and Naudé propose a ‘null-noun’ that בוט is actually modifying: ‘the good [thing]’ (2017, 293). Yet, while elision is no doubt a real and important part of language, this explanation will simply not suffice for the wide-sweeping tendencies reviewed in this article. The experimental data demonstrate a vast and systematic interaction of semantic-pragmatic factors which all cooperate to construe word class in an individual context. This, in general, casts doubt on the explanatory power of elision in cases where words ‘misbehave’.

6.0. Conclusions

64This article has applied a cognitive-statistical approach to examining parts of speech within time adverbial phrases in BH. The statistical model was able to separate words into broadly ‘noun’ items and ‘adverb’ items, with gradient behaviour in between. The linguistic analysis examined these tendencies against the background of cognitive theories on parts of speech. The data and analysis reveal an intricate interplay of contextual factors and indicate the need to classify parts of speech on a case-by-case basis. More generally, the success of the model demonstrates the fruitfulness and potential of applying statistics to the study of semantics. Co-occurrence behaviour of words is not random or coincidental, but arises from the interaction of the semantic-pragmatic values humans assign to them.

Bibliographie

References

Abdi, Hervé, and Lynne J. Williams. 2010. ‘Principal Component Analysis: Principal Component Analysis’. Wiley Interdisciplinary Reviews: Computational Statistics 2 (4): 433–59.

Barr, James. 2005. Biblical Words for Time. London: SCM Press.

— ——. 1989. ‘‘Determination’ and the Definite Article in Biblical Hebrew’. Journal of Semitic Studies 34 (2): 307–35.

Bauer, Hans and Pontus Leander. 1922. Historische Grammatik der Hebräischen Sprache des Alten Testamentes. Halle: M. Niemeyer.

BDB= Brown, Francis, Samuel R. Driver, and Charles A. Briggs. 1906. The Brown-Driver-Briggs Hebrew and English Lexicon. Boston: Houghton, Mifflin and Company.

Bekins, Peter. 2013. ‘Non-Prototypical Uses of the Definite Article in Biblical Hebrew’. Journal of Semitic Studies 58 (2): 225– 40.

— ——. 2016. ‘The Omission of the Definite Article in Biblical Poetry’. Paper delivered at Annual Meeting of the Society of Biblical Literature, San Antonio, TX. https://www.academia.edu/30145121/The_Omission_of_the_Definite_Article_in_Biblical_Poetry.

Brin, Gershon. 2001. The Concept of Time in the Bible and the Dead Sea Scrolls. Studies on the Texts of the Desert of Judah 39. Leiden: Brill.

Bybee, Joan L., Revere D. Perkins, and William Pagliuca. 1994. The Evolution of Grammar: Tense, Aspect, and Modality in the Languages of the World. Chicago: University of Chicago Press.

Clines, David J.A. (ed.). 2009. The Concise Dictionary of Classical Hebrew. Sheffield: Sheffield Phoenix Press.

Croft, William. 1990. ‘A Conceptual Framework for Grammatical Categories (or: A Taxonomy of Propositional Acts)’. Journal of Semantics 7: 245–79.

— ——. 2001. Radical Construction Grammar: Syntactic Theory in Typological Perspective. Oxford and New York: Oxford University Press.

Cuzzolin, Pierluigi, Ignazio Putzu, and Paolo Ramat. 2006. ‘The Indo-European Adverb in Diachronic and Typological Perspective’. Indogermanische Forschungen 111: 1–38.

Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. ‘BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding’. ArXiv: 1810.04805 [Cs], 24 May 2019. http://arxiv.org/abs/1810.04805.

DeVries, Simon J. 1975. Yesterday, Today and Tomorrow: Time and History in the Old Testament. Grand Rapids, MI: Eerdmans.

Ellis, Nick C., Matthew Brook O’Donnell, and Ute Römer. 2013. ‘Usage-Based Language: Investigating the Latent Structures That Underpin Acquisition: Usage-Based Language’. Language Learning 63: 25–51.

Fedorenko, Evelina, Idan Asher Blank, Matthew Siegelman, and Zachary Mineroff. 2020. ‘Lack of Selectivity for Syntax Relative to Word Meanings throughout the Language Network’. Cognition 203: 104348.doi.org/10.1016/j.cognition.2020.104348

Fillmore, Charles. 2002. ‘Mini-Grammars of Some Time-When Expressions in English’. In Complex Sentences in Grammar and Discourse: Essays in Honor of Sandra A. Thompson, edited by Joan L. Bybee, Sandra A. Thompson, and Michael Noonan, 31–59. Amsterdam: Benjamins.

Firth, J. R. 1962. ‘A Synopsis of Linguistic Theory, 1930–1955’. In Studies in Linguistic Analysis, 1–32. Oxford: Basil Blackwell.

Forbes, A. Dean. 2009. ‘Distributionally-Inferred Word and Form Classes in the Hebrew Lexicon: Known by the Company They Keep’. In Foundations for Syriac Lexicography II, edited by Peter J. Williams, 1–34. Piscataway, NJ: Gorgias Press.

— ——. 2005 ‘Squishes, Clines, and Fuzzy Signs: Mixed and Gradient Categories in the Biblical Hebrew Lexicon’. In Foundations for Syriac Lexicography: Colloquia of the International Syriac Language Project, edited by Beryl Turner, A. Dean Forbes, David G. K. Taylor, J. W. Dyk, and W. Th van Peursen, 105–39. Perspectives on Syriac Linguistics 1. Piscataway, NJ: Gorgias Press.

Geeraerts, Dirk. 2010. ‘The Doctor and the Semantician’. In Quantitative Methods in Cognitive Semantics: Corpus-Driven Approaches, edited by Dirk Geeraerts and John R. Taylor, 63– 78. Cognitive Linguistics Research 46. Berlin: De Gruyter Mouton.

— ——. 2010. Theories of Lexical Semantics. Oxford and New York: Oxford University Press.

Geeraerts, Dirk, and Hubert Cuyckens. 2012. ‘Introducing Cognitive Linguistics’. In The Oxford Handbook of Cognitive Linguistics, edited by Dirk Geeraerts and Hubert Cuyckens. Oxford: Oxford University Press.

GKC = Emil Kautsch (ed.). 1909. Gesenius’ Hebrew Grammar, translated by Arthur E. Cowley. Oxford: Clarendon Press.

Goldberg, Adele E., Devin M. Casenhiser, and Nitya Sethuraman. 2004. ‘Learning Argument Structure Generalizations’. Cognitive Linguistics 15 (3): 289–316.

Gries, Stefan Th. 2008. ‘Dispersions and Adjusted Frequencies in Corpora’. International Journal of Corpus Linguistics 13 (4): 403–37.

Haspelmath, Martin. 1997. From Space to Time: Temporal Adverbials in the World’s Languages. LINCOM Studies in Theoretical Linguistics 2. Munich: Lincom Europa.

— ——. 2007. ‘Pre-Established Categories Don’t Exist: Consequences for Language Description and Typology’. Linguistic Typology 11: 119–32.

Joüon, Paul, and T. Muraoka. 1996. A Grammar of Biblical Hebrew. Subsidia Biblica, 14/1–2. Rome: Pontifical Biblical Institute Press.

Kingham, Cody. 2020. CambridgeSemiticsLab/BH_time_colloca- tions: Time Collocations with Article Data (version 4.1). Zenodo. doi.org/10.5281/ZENODO.3931953.

Kingham, Cody, and Wido Van Peursen. 2018. ‘The ETCBC Database of the Hebrew Bible’. Journal for Semitics 27 (1): 1–13.

Köhler, Ludwig, and Walter Baumgartner. 1958. Supplementum ad Lexicon in Veteris Testamenti Libros. Leiden: Brill.

Kuhn, Thomas S. 2000. ‘The Natural and the Human Sciences’. In The Road Since Structure: Philosophical Essays, 1970–1993, with an Autobiographical Interview, edited by James Conant and John Haugeland, 216–23. Chicago: University of Chicago Press.

Lakoff, George. 1987. Women, Fire, and Dangerous Things: What Categories Reveal about the Mind. Chicago: University of Chicago Press.

Levshina, Natalia. 2015. How to Do Linguistics with R: Data Exploration and Statistical Analysis. Amsterdam: John Benjamins.

Locatell, Christian. 2017. ‘Grammatical Polysemy in the Hebrew Bible: A Cognitive Linguistic Approach to כי’. PhD dissertation, University of Stellenbosch.

Lyavdansky, Alexey. 2010. ‘Temporal Deictic Adverbs as Discourse Markers in Hebrew, Aramaic and Akkadian’. Journal of Language Relationship 3: 22–42.

Miller-Naudé, Cynthia L., and Jacobus A. Naudé. 2017. ‘A Reexamination of Grammatical Categorization in Biblical Hebrew’. In From Ancient Manuscripts to Modern Dictionaries: Select Studies in Aramaic, Hebrew and Greek, edited by Tarsee Li and Keith D. Dyer, 273–308. Perspectives on Linguistics and Ancient Languages 9. Piscataway, NJ: Gorgias Press.

Niccacci, Alviero. 2013. ‘Temporal Clause: Biblical Hebrew’. In Encyclopedia of Hebrew Language and Linguistics, edited by Geoffrey Khan et al. Leiden: Brill.

Oosting, Reinoud. 2016. ‘Computer-Assisted Analysis of Old Testament Texts: The Contribution of the WIVU to Old Testament Scholarship’. In The Present State of Old Testament Studies in the Low Countries: A Collection of Old Testament Studies Published on the Occasion of the Seventy-Fifth Anniversary of the Oudtestamentisch Werkgezelschap, edited by Klaas Spronk, 192–209. Oudtestamentische Studiën 69. Leiden: Brill.

Pedregosa, F., G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, et al. 2011. ‘Scikit-Learn: Machine Learning in Python’. Journal of Machine Learning Research 12: 2825–2830.

Perek, Florent, and Adele E. Goldberg. 2017. ‘Linguistic Generalization on the Basis of Function and Constraints on the Basis of Statistical Preemption’. Cognition 168: 276–93.

Piantadosi, Steven T. 2014. ‘Zipf’s Word Frequency Law in Natural Language: A Critical Review and Future Directions’. Psychonomic Bulletin & Review 21 (5): 1112–30.

Ramat, Paolo. 2011. ‘Adverbial Grammaticalization’. In The Oxford Handbook of Grammaticalization, edited by Bernd Heine and Heiko Narrog. Oxford: Oxford University Press.

Ravid, Dorit, and Yitzhak Shlesinger. 2000. ‘Modern Hebrew Adverbials: Between Syntactic Class and Lexical Category’. In Between Grammar and Lexicon, edited by Ellen Contini-Morava and Y. Tobin, 333–51. Amsterdam Studies in the Theory and History of Linguistic Science 183. Amsterdam: John Benjamins.

Rokhlin, Vladimir, Arthur Szlam, and Mark Tygert. 2010. ‘A Randomized Algorithm for Principal Component Analysis’. SIAM Journal on Matrix Analysis and Applications 31 (3): 1100–24. doi.org/10.1137/080736417.

Roorda, Dirk. 2017. ‘The Hebrew Bible as Data: Laboratory–Sharing–Experiences’. In Clarin in the Low Countries, edited by J. Odijk and A. Van Hessen. London: Ubiquity Press.

Roorda, Dirk, Constantijn Sikkel, Wido Van Peursen, Martijn Naaijer, Christiaan Erwich, and Cody Kingham. 2019. ‘Biblia Hebraica Stuttgartensia (Amstelodamensis) Documentation’. Documentation. ETCBC Github. https://etcbc.github.io/bhsa/.

Sáenz-Badillos, Angel. 1997. A History of the Hebrew Language. Reprinted. Cambridge: Cambridge University Press.

Scholz, Barbara C., Francis Jeffry Pelletier, and Geoffrey K. Pullum. 2011. ‘Philosophy of Linguistics’. In The Stanford Encyclopedia of Philosophy, edited by Edward N. Zalta.plato.stanford.edu/archives/win2016/entries/linguistics/.

Schütze, Carson T. 2016. The Empirical Base of Linguistics: Grammaticality Judgments and Linguistic Methodology. Classics in Linguistics 2. Berlin: Language Science Press.

Sloutsky, Vladimir M., Wei (Sophia) Deng, Anna V. Fisher, and Heidi Kloos. 2015. ‘Conceptual Influences on Induction: A Case for a Late Onset’. Cognitive Psychology 82: 1– 31.doi.org/10.1016/j.cogpsych.2015.08.005.

Stefanowitsch, Anatol. 2020. Corpus Linguistics: A Guide to the Methodology. Textbooks in Language Sciences 7. Berlin: Language Science Press.

— ——. 2010. ‘Empirical Cognitive Semantics: Some Thoughts’. In Quantitative Methods in Cognitive Semantics: Corpus-Driven Approaches, edited by Dylan Glynn and Kertin Fischer, 355– 80. Cognitive Linguistics Research 46. Berlin: De Gruyter Mouton.

Stefanowitsch, Anatol, and Stefan Th. Gries. 2003. ‘Collostructions: Investigating the Interaction of Words and Constructions’. International Journal of Corpus Linguistics 8 (2): 209– 43.

ttnphns (https://stats.stackexchange.com/users/3277/ttnphns). ‘Loadings vs Eigenvectors in PCA: When to Use One or Another?’ Stack Exchange, 29 March 2015. https://stats.stackexchange.com/q/143949.

van der Merwe, Christo H. J., J. A. Naudé, and Jan Kroeze. 2017. A Biblical Hebrew Reference Grammar. Second edition. New York: Bloomsbury T & T Clark.

van der Merwe, Christo H. J. 1997. ‘Reconsidering Biblical Hebrew Temporal Expressions’. Zeitschrift für Althebraistik 10 (1): 42–62.

Waltke, Bruce, and M. O’Connor. 1990. An Introduction to Biblical Hebrew Syntax. Winona Lake, IN: Eisenbrauns.

Xiao, R. ‘Collocation’. 2015. In The Cambridge Handbook of English Corpus Linguistics, edited by D. Biber and R. Reppen, 106– 24. Cambridge: Cambridge University.

Notes

1 See, though, the important work of van der Merwe (1997). There have also been numerous studies dedicated to individual lexemes or to temporal clauses (כִי ‘for/when’, עַתָּה ‘now’). There have also been two large exegetical studies on time phrases: DeVries (1975) and Brin (2001). This study focuses instead on time adverbials as a broader linguistic phenomenon. See also Lyavdansky (2010); Barr (2005); Niccacci (2013); Locatell (2017).

2 van der Merwe (1997, 49) asks similar questions:
Is there any difference in the syntax of non-temporal adjuncts and temporal adjuncts? Which BH constructions can function as temporal adjuncts? Apart from the above-mentioned semantic classes of temporal position, duration and frequency, are there other semantic classes or subclasses to be identified among temporal adjuncts?

3 I agree with Ellis, O’Donnell, and Römer (2013, 30): “We take the Saussurian (1916) view that the units of language are constructions— form-meaning mappings, conventionalized in the speech community, and entrenched as language knowledge in the learner’s mind.” For experiential categories see Lakoff (1987).

4 On similarity-based categorisation see Sloutsky et al. (2015).

5 Indeed, as noted above, these mental associations are learned from relationships observed in the world. Language thus represents a re-externalisation of semantic input.

6 Famously summarised by Firth (1962, 11) as “knowing a word by the company it keeps.”

7 Based on an ongoing analysis by this author: https://github.com/codykingham/noun_semantics/blob/master/analysis.ipynb.

8 These links may not be one-to-one (e.g., concept B > construction Z and Y).

9 This is the process known as ‘operationalisation’ and is the ‘secret’ to the success of the natural sciences; see Stefanowitsch (2010, 358–61).

10 This is done using a custom Python parser which works within the existing phrase segmentation of the BHSA. Where the phrase segmentations are deficient, they are modified. For the parser, see the project repository at https://github.com/CambridgeSemiticsLab/BH_time_collocations/blob/master/data/cxs/phrase_grammar.py.

11 The whole project is archived at the following reference. For this article, data is stored in the various directories under the heading ‘advb_article’. The datasets and production script can be found at data/advb_article (see function_data.csv and generate_function_data. ipynb). The parsers can be found under /tools/cx_analysis and data/cxs. The principal analysis code is under results/notebooks/advb_article. Figures are under results/figures. The directory results/spreadsheets/ advb_article includes the PCA values for the part of speech analysis. The whole directory requires a virtual python environment with /tools appended to the Python namespace. The file requirements.txt provides the necessary prerequisites; see Kingham (2020).

12 I used Croft’s (2001, 257–59) notion of semantic headship as a general guide. Objects of prepositions are the semantic heads of prepositional phrases. Quantifiers including כֹּל ‘all’ are excluded as heads if they modify a word.

13 A measure called deviation of proportions (Gries 2008, 415–19) was used to compare the sampled distribution by book against the whole dataset. The most affected books are Deut. (3 percent total deviation), Ps. (3 percent), 2 Chron. (2 percent), Isa. (2 percent), and Job (2 percent).

14 Throughout the study, I refer to proportions out of 1, which can be converted to percentages by multiplying them by 100. This is standard practice in statistical reporting; see Levshina (2015, 70–71).

15 This includes articles reflected only in the vocalised text of BHS.

16 In this case the modifier term is the nomen rectum, or trailing term in a construct chain, rather than the head which is morphologically marked as construct. For this reason, the term ‘genitive’ is used. The BHSA dataset includes constructs indicated via the accents or context.

17 The attributive adjunct construction involving the definite article, i.e., - ַה + word + -ַה + adjunct, is also tagged separately. But since that construction itself involves definite modification, it is excluded from this study. The simple adjunct construction of word + adjunct is not yet tagged due to existing limitations in the dataset. In the BHSA, these relations are tagged as so-called ‘subphrase relations’; but subphrase relations suffer from a number of problems and inadequacies making them unreliable. In the case of time adverbials, it seems (based on previous analyses) that adjectival elements play a lesser role in the semantics of time. Nevertheless, in the future it would be better to include these modifiers in the dataset.

18 For a mathematical description, see Abdi and Williams (2010).

19 Normalisation involves adjusting the counts so that they are comparable. For instance, some words occur very frequently and, therefore, have high co-occurrence counts, while others are comparatively rare. A common method is to take a percentage of a sample’s overall frequency, here the decimal proportion, i.e., the percentage divided by 100.

20 The lexicon parts of speech values come from the BHSA’s ‘sp’ feature, which is derived from Köhler and Baumgartner (1958). The respective abbreviated values are substantive, interrogative, adjective, and adverb.

21 These are the loading scores, which are the eigen vectors × √eigen values following the discussion of ttnphns, 29 March 2015.

22 This is a crude description of PCA, but gets the point across. Abdi and Williams (2010) present a more precise mathematical description.

23 The loading score includes direction from the eigen vector and magnitude from the eigen value. The Ø modifier has a PC1 (x-axis) of −4.8 and PC2 (y-axis) of 3.9. The line is drawn from origin to this point.

24 E.g., Gen. 7.11; Deut. 3.14; Josh. 20.6; 1 Sam. 7.6; 2 Sam. 3.38; 2 Kgs 20.1; Isa. 28.5; Ezek. 39.11; Zech. 13.4; 2 Chron. 35.16.

25 E.g., Exod. 12.14, 17, 42; 40.15; Lev. 18.18; Judg. 16.30; 1 Sam. 17.33; 2 Sam. 18.18; 1 Kgs 18.12; Isa. 47.12, 15; Ezek. 23.3, 8; Ps. 49.19; 104.33; 146.2.

26 In Modern Hebrew the distinction between adjectives and adverbs is blurry. A similar dynamic appears in BH. Ravid and Shlesinger (2000, 346).

27 As Croft (1990, 246) recognised with constructions like ‘day’ and ‘five times’.

28 Exod. 18.13; 32.30; Num. 17.23; Judg. 9.42, 21.4; 1 Sam. 11.11; 18.10; 31.8; 2 Kgs 8.15; Jer. 20.3; 1 Chron. 10.8.

29 The original status of some definite articles reflected only in the vocalisation is debated. The issue at hand is whether the vocalisation tradition preserves an original article, or whether it can be attributed to the tradition itself. As part of the PhD project within which this article is embedded, I have found preliminary data that time adverbials as a semantic function are statistically associated with definite articles, even when merely vocalic articles are removed from consideration. So, at bare minimum, definiteness plays a key role in time semantics. For the wider discussion see GKC (§126w); Barr (1989); Bekins (2013; 2016).

30 But cf. גָּלְיָּת ‘Goliath’ (1 Sam. 17.4); see GKC (§80g).

31 See also the case of לְמִן־הַיֹום ‘since the day of’ (Exod. 9.18), but here with the definite article.

32 Thanks to Chip Hardy for pointing out these possibilities in relation to another similar case, לְעֹולָּם.

33 N.B. that this table contains a higher marginal total than the previous since modifiers can co-occur.

34 The closest would be בֹּקֶר with 18/119 (15 percent) null modified uses, a substantial difference from the 35 percent of יֹום .לַיְלָּה has 0.08 percent Ø; עֶֶרב has 3.5 percent Ø.

35 Thanks to Chip Hardy for pointing this out.

36 This is a common phenomenon in world languages (Haspelmath 1997).

37 Note forms without the ending, e.g., כַלַָיִל ‘like the night’ (Isa. 16.3) and the construct לֵ֣יל ‘night of’ (Exod. 12.42).

38 See Bybee, Perkins, and Pagliuca (1994, 15–16) on the comparison of English will, shall, and be going to.

39 The others are Isa. 45.17; 51.9; Ps. 77.6; Eccl. 1.10; Dan. 9.24.

40 Perhaps a clue for the origin of these plurals can be found in their use with physical/spatial nouns. There the plural “denote [s] localities in general,” with examples such as שָּמַיִם ‘sky’, מַיִם ‘water’, ּפָּנִים ‘face, front’, אֲחֹוִרים ‘back’, מְַרגְלֹות ‘place of feet’, מִשְכָּבִים ‘bed place’, etc. (GKC §124b– d). It seems possible that a location-in-space construction might have been semantically extended to include abstract concepts which are spread out over a metaphorical sphere such as a timeline.

41 See the note in the methodology section about excluding complex phrases. This is done to ensure accurate selection of phrase heads and modifiers.

42 This also seems true of the corresponding Greek term αἰών ‘eon’, see Hebrews 1.2; 9.26; 11.3. Thanks to the anonymous reviewer who pointed this out.

43 Thanks again to Chip Hardy for this idea.

44 Some of the other cases not mentioned here actually contain adjectival modification, e.g., יֹוםְטֹוב ‘good day’ (1 Sam. 25.8), or adverbial modification, e.g., בְיֹום ְמָּחָּר lit. ‘on day of tomorrow’ (Gen. 30.33) (though this item might also be construed as a construct with some nominal sense), which were excluded from consideration for the purposes of this analysis. These cases should not have been tagged as Ø. A future model could include these considerations. However, in the case of time adverbials they are comparatively quite rare. Other interesting cases of unmodified יֹום include מִיֶֹּ֥ום ְעַד־לַָיְלָּה ‘from day to night’ (Isa. 38.12יֹום ,) ‘at day’ (Jer. 7.25; Ps. 88.2), and מִיֶֹום־לְיֶֹום ‘from day to day’ (Ps. 96.2).

45 Josh. 24.2; Isa. 42.14; 57.11; 63.19; 64.3; Jer. 2.20; Prov. 8.23.

46 A handful of terms possess adverb suffixes, such as ָָּם- as found in יֹומָּם ‘by day’; see van der Merwe, Naudé, and Kroeze (2017, 380).

47 Croft (2001, 65) notes Nootkan, Salishan, Iroquoian, Philippine, and Polynesian languages as oft-cited examples.

48 Geeraerts and Cuyckens (2012, §2) write: “The perspectival nature of linguistic meaning implies that the world is not objectively reflected in the language: the categorization function of the language imposes a structure on the world rather than just mirroring objective reality.”

49 E.g., עַד־אָּן/אָּנָּה ‘how long? ’ (Exod. 16.28; Num. 14.11; Josh. 18.3, et al.); see BDB 33. This is similar to הֵנָּה.

Table des illustrations

Titre Figure 1: Mutual relationship between linguistic, cognitive, and statistical data
URL http://books.openedition.org/obp/docannexe/image/20280/img-1.jpg
Fichier image/jpeg, 16k
Titre Table 1: Time Adverbial phrase function frequencies and proportion of selected samples versus BHSA14
URL http://books.openedition.org/obp/docannexe/image/20280/img-2.jpg
Fichier image/jpeg, 7,4k
Titre Table 2: Collocational frequencies between Time head and nominal modifiers
URL http://books.openedition.org/obp/docannexe/image/20280/img-3.jpg
Fichier image/jpeg, 10k
Titre Table 3: Excerpt of Time head x modifier co-occurrence matrix (original: 41x9) Using the Pandas Python package and a spreadsheet of individual observations, a tabulation is compiled which counts how frequently a given head word lexeme is used with a given modifier type. The result is a table with 41 rows and 9 columns (not including the ‘head’ column), containing a total of 369 data points. The excerpt below shows the first three rows of that table.
URL http://books.openedition.org/obp/docannexe/image/20280/img-4.jpg
Fichier image/jpeg, 15k
Titre Table 4: Excerpt of Time head x modifier proportion matrix (original: 41x9) The data from Table 3, containing the raw co-occurrence frequencies, are converted to a table containing proportions using a Pandas sum (across rows) and divide operation (across columns). The resulting decimal values across each row sum to 1, which corresponds with 100 percent of a given lexeme’s attested modifiers. For example, the ‘cardinal’ feature with י ֹ םו comprises 0.07 in 1 of all the modifiers observed with י ֹ םו , or 7 percent.
URL http://books.openedition.org/obp/docannexe/image/20280/img-5.jpg
Fichier image/jpeg, 17k
Titre Figure 2: PCA analysis of head word tendencies The PCA analysis is performed using the scikit-learn Python package. The function takes in the modification proportions per head lexeme, which contains nine dimensions, and reduces the dimensions to two vectors, each of which can then be assigned as the X and Y coordinates of the respective head lexeme. The first vector (X-axis) accounts for the most variation in the dataset, and the second (Y-axis), the second-most. Together the two components account for about 80 percent of variance in this sample.
URL http://books.openedition.org/obp/docannexe/image/20280/img-6.jpg
Fichier image/jpeg, 31k
Titre Table 5: Mean proportions for null (Ø) versus nominal modifiers by placement along x-axis
URL http://books.openedition.org/obp/docannexe/image/20280/img-7.jpg
Fichier image/jpeg, 4,7k
Titre Table 6: PCA Loading Scores (rounded to two decimal places)
URL http://books.openedition.org/obp/docannexe/image/20280/img-8.jpg
Fichier image/jpeg, 14k
Titre Table 7: Adverb-like lexeme cluster (PC1 < 0) (with modification proportions per lexeme shown across rows; each row sums to 1, i.e., 100 percent)
URL http://books.openedition.org/obp/docannexe/image/20280/img-9.jpg
Fichier image/jpeg, 60k
Titre Table 8: Noun-like lexeme cluster with strong definite/demonstrative influence (PC1 > 0, PC2 > 0) (with modification proportions per lexeme shown across rows; each row sums to 1, i.e., 100 percent)
URL http://books.openedition.org/obp/docannexe/image/20280/img-10.jpg
Fichier image/jpeg, 43k
Titre Table 9: Noun-like lexeme cluster with plural, genitive, suffix (PC1 > 0, PC2 < 0) (shown with modification proportions across rows; each row sums to 1, i.e., 100 percent)
URL http://books.openedition.org/obp/docannexe/image/20280/img-11.jpg
Fichier image/jpeg, 38k
Titre Table 10: Sampled sentences for ֶ גר ַ ע
URL http://books.openedition.org/obp/docannexe/image/20280/img-12.jpg
Fichier image/jpeg, 40k
Titre Table 11: Sampled modifiers with מ ָּ ח ֳָּ תר
URL http://books.openedition.org/obp/docannexe/image/20280/img-13.jpg
Fichier image/jpeg, 81k
Titre Table 13: Selected samples of מ ָ ח ֳָ תר (continued)
URL http://books.openedition.org/obp/docannexe/image/20280/img-14.jpg
Fichier image/jpeg, 25k
Titre Table 14: Frequency and proportion of ל ַ י ְ ל ָּ ה with modification type
URL http://books.openedition.org/obp/docannexe/image/20280/img-15.jpg
Fichier image/jpeg, 5,0k
Titre Table 15: Sampled modifiers with ל ַ י ְ ל ָּ ה33
URL http://books.openedition.org/obp/docannexe/image/20280/img-16.jpg
Fichier image/jpeg, 9,8k
Titre Table 16: Selected samples of ל ַ י ְ ל ָּ ה
URL http://books.openedition.org/obp/docannexe/image/20280/img-17.jpg
Fichier image/jpeg, 104k
Titre Table 17: Selected nominalised samples of לוע ָ ם
URL http://books.openedition.org/obp/docannexe/image/20280/img-18.jpg
Fichier image/jpeg, 60k
Titre Table 18: Select null-modified samples of םוי ‘day’
URL http://books.openedition.org/obp/docannexe/image/20280/img-19.jpg
Fichier image/jpeg, 17k
Titre Table 19: Croft’s Major Propositional Acts (modified from Croft 1990, 248)
URL http://books.openedition.org/obp/docannexe/image/20280/img-20.jpg
Fichier image/jpeg, 20k
Titre Table 20: Croft’s Minor Propositional Acts (adapted and condensed)
URL http://books.openedition.org/obp/docannexe/image/20280/img-21.jpg
Fichier image/jpeg, 19k

Auteur

Cody Kingham (PhD candidate, University of Cambridge) applies Machine Learning and data-driven methods to the study of Biblical Hebrew. He has co-authored articles on Hebrew databases and probabilistic models of Hebrew grammar. He is currently working on his doctoral thesis which looks at the collocational behaviour of time adverbials using cognitive linguistics and advanced statistical approaches.

Acheter

Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search