The distributional properties of long nominal compounds in scientific articles: an investigation based on the uniform information density hypothesis

1 Introduction

All languages of the world allow for some form of compounding (Dressler 2006). In English, the result of this process can be a word (e.g., outstanding, snowman, strawberry, highlight), or a more complex structure composed of multiple words (e.g., blood moon, health care provider, dopamine production suppressor protein). This process of word formation is so pervasive that it has been referred to as the “universally fundamental word formation process” (itself a compound), offering “the easiest and most effective way to create and transfer new meanings” (Libben 2006). In English, the vast majority of compounds are nouns (Algeo and Algeo 1991: 7). These are ubiquitous in everyday language, comprising 2.6 % of the British National Corpus and 3.9 % of the Reuters corpus (Baldwin and Tanaka 2004).

In this paper, we are concerned with the sort of compounding that leads to the formation of complex structures composed of multiple words. In particular, we turn our focus to structures we refer to as nominal compounds (NC), multiword structures that, as a whole are categorized as nouns. For concreteness, we will focus on examples such as water waste or addictive substance (both of which, taken as a whole, form a noun) and not on resource poor (an adjective), or ambulance-chase (a verb).

On the surface, such NCs are typically composed of a head noun and one or more modifiers. For example, the NC health care is composed of the head noun care and the modifier health. On a deeper level, NCs present a hierarchical structure. For example, health care could be reused recursively as a single unit in a subsequent compounding process to construct more complicated structures such as health care provider (in which it is used as a modifier) or geriatric health care (in which it is used as a head noun).^[1] In addition, the way in which the NC words are linked can vary widely. For instance, olive oil can be interpreted as an “oil FROM olives”, but baby oil is typically an “oil FOR babies” and an olive tree is a “tree THAT HAS olives”. While the exact set of possible linking relations between the NC words has been a topic of debate, a number of studies have demonstrated that they have an effect on processing (see, e.g., Gagné and Shoben 1997; Levi 1978; Spalding et al. 2010).

NCs are particularly frequent in scientific texts (Bhatia 1992). In this register, between 9 % and 16 % percent of all words are part of an NC, and NCs tend to be longer than in everyday English (Salager 1984). Biber and Gray (2011), in an analysis of NC use since the 18th century in different English registers, found that their prevalence in the scientific register increased 10-fold between 1875 and 2005. They also found that their complexity increased, with three-word compounds being initially uncommon but becoming common by the 1950s. They attributed this increase in complexity to a “principle of economy” inherent to these registers. This conclusion echoes those of Montero (1996), who additionally suggested that NC use is the result of a “desire for novelty”; and of Salager (1984]: 142), for whom an NC is “a new concept for which the language code has no name … crystallized into a fixed expression owing [sic] a scientific meaning which the individual constituents do not have”. According to Salager’s analysis, this new concept, once used for the first time, can be reused or referred to as an entity that the writers know the readers can understand.

In this paper, we refer to longer NCs (composed of three or more words) as complex nominal compounds (CNC). CNCs are sometimes considered hard to understand, and are typically discouraged in “good writing” guidelines (e.g., Tobin 2002). Indeed, a number of studies have shown that they may lead to difficulty in some circumstances. For example, individuals have been shown to be unable to identify the head of a given CNC (Geer et al. 1972; Limaye and Pompian 1991), and the CNCs themselves are ambiguous sometimes (cf. Montero 1996): as noted by Kvam (1990), a kitchen towel rack may be a rack for kitchen towels or a towel rack in the kitchen. In addition, L2 speakers (who are common producers and consumers of scientific literature) often translate CNCs in inconsistent ways (Carrió Pastor 2008; Carrió Pastor and Candel Mora 2013), and are sometimes unable to recover the implicit semantic links between the NC words (Horsella and Pérez 1991).

2 Corpus construction

We formed a dataset containing research articles from nine high-impact journals in the fields of Biology, Economics and Linguistics, published either in 2016 or 2017. The choice of these fields was arbitrary, but we purposefully included texts from both the Natural Sciences (Biology), and from the Social Sciences (Linguistics and Economics). For each field, we collected exactly 54 articles, but the number of articles from each journal varied (see Table 1). The full list of articles can be found in the supplementary materials.

We downloaded each article in PDF format and converted it to TXT using the AntFileConverter (Anthony 2017), which outputs text files in Unicode format. We then manually removed all abstracts, headers, references, appendices, tables, notes and pages numbers from the resulting text files. In addition, we manually replaced a number of mathematical formulas from the text files with a more natural textual continuation. For example, the sentence “To test this idea, we estimate the following regression: FORMULA where all variables have been defined previously” was changed to “To test this idea, we estimate the following regression, where all variables have been defined previously”.^[5] This latter manual text editing step was performed in an attempt to improve the output of the part of speech (POS) tagger, and did not affect results of the further analysis, because the changes were made on function words or punctuations.

We then manually inserted section tags in each of the files. In particular, we added the tags <Introduction>, <Middle> and <Conclusion> to the files as follows:

1. Introduction: added at the very beginning of each file. An Introduction extended from the beginning of the file until just before the Methods section. Since articles differ widely in their structure, this may include theoretical sections and descriptions of the related literature.

2. Middle: added just before the header corresponding to the Methods section. The Middle part generally contained the Methods and the Results sections. In articles reporting more than one experiment, the Middle part also included the Discussion sections associated with each single experiment.

3. Conclusion: added just before the header corresponding to the Discussion section of a given file. In articles reporting a single experiment, the “Conclusion” corresponded to the Discussion and Conclusion sections. In articles reporting several experiments, it contained the General Discussion and the Conclusion sections.

In the rest of this paper, we use the word section to refer to each of these article parts.

The TXT files also contained a number of spurious Unicode characters that needed to be cleaned before analysis. Hence, a Python (van Rossum and Drake 2009) script was used to remove all non-printable characters from the dataset, and replace all typographic ligatures with their constituent characters (e.g., the single character representing “ffi” was replaced by two letters “f” and one “i”). This did not remove special characters such as “é” or “ã”, which only appeared casually throughout the dataset.

The dataset was then tokenized, part-of-speech (POS) tagged and parsed using the spaCy library (Honnibal et al. 2019). The final output of this procedure was a table where each row contained a token,^[6] its lemma, its POS tag, its head, the relation between the word and its head, the file where the word came from, the field (Biology, Economics or Linguistics) of that file, the word’s position in the sentence, its position in the section, its position in the file, and a unique ID for the whole word in the dataset. This yielded a dataset containing 336,466 words originating from Biology papers, 510,685 words from Economics papers, and 550,992 words from Linguistics papers. For the sake of simplicity and ease of future reference, we call this the Sciper (SCIentific paPER) corpus.

3 Quantitative study of CNCs in research articles

In order to analyse the distributional properties of the CNCs in the Sciper corpus, we needed to first identify them. We start this section by describing the procedure used to identify CNCs. We then follow with an analysis of each of the aforementioned questions: We analyse the position of CNCs across the articles, the number of times CNCs are preceded by their subparts, and the number of times they repeat.

3.1 Identifying CNCs in Sciper

Recall that the corpus is a table containing the words of 162 articles such that each row represents a word. We used this table to count the number of CNCs by article and, inside the articles, by section. Since article length varied widely (see Figure 1) we also calculated the proportion of words pertaining to CNCs for each 1,000 words in the articles.

Figure 1:

Histogram of the length of the corpus files.

A sequence of words was deemed a CNC if it followed either the structure (Adjective)+ (Noun){2,} (at least one adjective followed by at least 2 nouns) or the structure Noun{3,} (three or more nouns in sequence).^[7] This yielded 24,519 sequences of 3 or more tokens. These sequences found were quite noisy. Many contained incorrectly POS-tagged tokens, tokens with spurious characters (letters in other languages, such as Greek, Chinese or Hebrew), or words from other languages that used the Latin alphabet (such as German, Finnish or French). Sometimes, the construction “ProperName et al.” was POS-tagged in a way that would yield a CNC. Sometimes, the tokens were single letter characters (such as “M E T H O D”), or the sequences contained words such as “A.” or “i.”. Sometimes, the PDF to TXT conversion led to incorrectly separated words (e.g., “inbothFrenchandZulu” or “Na ture”). Finally, many of the articles used cryptic acronyms (like “GABA”, “AgCl” or “trkA”) or unity values (like “ms” or “kg”) that were not of interest to us.

Therefore, we applied the following cleaning procedure to the data. First, we automatically identified and discarded all CNCs containing any non-letter symbols except for the “.” (period) character. Examples of non-letter symbols were numbers, Greek letters (mostly used in formulas), hyphens and mathematical operators. If a number appeared preceded by a “.” at the very end of the last CNC token (e.g., “earnings response coefficient.9”), we assumed that it corresponded to a footnote, and the CNC was not discarded. In addition, capitalized words containing up to three letters were also removed, as well as words containing 2 characters ended by a “.”, or words composed of a single character.

In a final step, all items were manually inspected, producing a number of exceptions to the aforementioned rules (such as “AI”, “IQ”, “CEO”, “US”, “GDP” or “EFL”), and a number of words we considered unlikely to constitute CNCs despite conforming to the rules above (e.g., “vs.”, “Inc.”, “Ltd” or “Rev.”). We also manually listed many foreign words (from Dutch, German, Portuguese, French, Finnish, Japanese and other languages) that were found among the sequences, and whenever a CNC contained any of them it was discarded. Finally, this manual inspection also allowed us to find a number of items that conformed to all criteria but were not CNCs. This was typically the case when the item words were, for example, at the border of an adverbial clause (1) or constituted an apposition (2).

(1)

Table 2 shows that animacy significantly affected children’s performance: across all sentence types children performed better on sentences that contained inanimate head nouns. (Kirjavainen et al. 2017: 133)

(2)

We analyzed the data using mixed-effects logistic regression models to test the effect of four predictors on the continuation selected by participants: sentence type (cleft or canonical), grammatical function of the focus (subject or object), language group (English control, French control, or L2 French) and proficiency for the French data (native controls, low, intermediate, or advanced). Prior to analysis, the two predictors sentence type and grammatical function were effect coded (i.e., sum-coding with values -1 for cleft and +1 for canonical/-1 for object and +1 for subject). (Destruel and Donaldson 2017: 715)

This cleaning procedure led to the exclusion of 7,599 sequences. The remaining 16,920 CNCs were analysed as reported in the following sections.

3.2 How are CNCs distributed through scientific articles?

To analyse whether CNCs are more common toward the end than toward the beginning of scientific articles, we counted the number and length of CNCs in each section and paper. We analysed these counts using a Linear Mixed Effects Models (LMEM). Following Biber and Gray (2011), we calculated the sum of the lengths of the CNCs (i.e., the number of words of the CNCs) in a given article section divided by the section length and multiplied by 1,000. We refer to this measure as the CNC Proportion. For example, if a section contained exactly 1,000 words, 50 3-word CNCs, and 10 4-word CNCs, then:

CNC proportion = ( 50   CNCs × 3   words ) + ( 10   CNCs × 4   words )   1000   words   in   article × 1000

A visual inspection of the distribution of this measure made it clear that it is not normal. Therefore, we log-transformed the data (compare Figure 2a with Figure 2b). This led us to discard two data points (two article sections) because they had no CNCs.

Figure 2:

The CNC proportion by field and section. Each datapoint is a file section in the corpus. (a) Raw values; (b) log-transformed values.

In Figure 2 each data point represents the proportion of CNCs in a given section. Both plots show values calculated for the same data. Linguistics articles seem to have on average a lower proportion of CNCs than Biology, and Economics a higher proportion of CNCs than Biology; but there does not seem to be an effect of the article section.

These results were confirmed in the LMEM whose coefficients are shown in Table 2. The LMEM model was fit using the log-transformed data,^[8] with section and field of study as fixed effects, and random intercepts by file.^[9] In other words, the model used the formula:

DV ∼ section + field + ( 1   |  file )

Table 2:

Results of the linear mixed effects model.

CNC proportion (log-scaled)
	Estimate	Std. error	df	t Value	Pr(>|t|)
(Intercept)	3.498	0.063	160.573	55.776	0.000	***
Intro vs. middle	0.013	0.040	321.177	0.328	0.743
Middle vs. conclusion	0.012	0.040	321.177	0.305	0.761
Biology vs. economics	0.421	0.089	160.821	4.745	0.000	***
Biology vs. linguistics	-0.387	0.089	160.821	-4.359	0.000	***

1. ^∗∗∗ p < 0.001; ^∗∗ p < 0.01; ^∗ p < 0.05.

The Field factor was treatment coded with Biology as the reference factor. The Section factor was difference coded in the order Introduction, Middle, Conclusion.

When controlling for article length, we found no evidence that CNCs cluster around the Conclusion, or that the number of CNCs increases toward the end of the articles, as previously predicted based on the UIDh. We return to this point in the Discussion.

3.3 Are CNCs preceded by their subparts?

In this analysis, we take a first step in understanding the strategies authors use when introducing CNCs. A deeper discussion will be presented in the qualitative analysis.

For each CNC, we count the number of bigram subparts that appeared in the whole text before the CNC. For example, given the CNC left feeder reward probability, we counted how many times left feeder, feeder reward and reward probability appeared in the whole portion of the article preceding it.^[10] We count bigram subparts because they are the simplest structure we could count that is more complex than a word.

Figure 3a shows the distribution of the CNCs by the number of subparts before them. Since CNCs at the end of a paper have much more text before them, we also calculated the distribution only considering the CNCs in the Conclusion section (all of which have, of course, a large portion of text preceding them; Figure 3b).

Figure 3:

The distribution of two-word subparts of the CNCs preceding the CNC. (a) Histogram of number of subparts found in the whole article text preceding the CNC. (b) Histogram of number of subparts found in the whole article text preceding the CNC (only considering CNCs in the Conclusion).

As can be seen, very few subparts appear before the vast majority of the CNCs in our corpus. A third of the CNCs in the corpus (34.1 %) are not preceded by any subpart at all. This percentage increases to 61.6 % if we add to this set the CNCs preceded by their subparts once (10 %), twice (8.16 %), thrice (5.05 %), four times (4.21 %) and five times (3.26 %). Analyzing it from the opposite perspective, less than 20 % of the CNCs (19.47 %) are preceded 15 or more times by their subparts.^[11] This was unexpected, but we return to this matter in the qualitative analysis below.

4 Qualitative analysis of CNC use

We then analysed CNC use from a qualitative perspective. We investigated the way in which CNCs are preset by their context. In particular, we looked for strategies authors might employ to introduce the CNCs in a way that makes their understanding easier. Since only 3.5 % of CNCs occur more than three times in a given paper, we decided to investigate what differentiates these CNCs from the others. Therefore, in the description below, we make a distinction between “recurrent” CNCs (those that appeared four or more times in a paper) and “disposable” CNCs (those that appeared up to three times). We randomly selected 50 disposable and 26 recurrent CNCs from our corpus for manual analysis. This produced a total of 259 CNC occurrences. We selected fewer recurrent CNCs because by definition they led to a much higher number of occurrences, that had to be analysed individually (see Table 6 for a list of all CNCs analysed, as well as the number of times they appeared in a given article). For each CNC occurrence, we examined the text surrounding the CNC, looking for similarities between the different CNC uses in our dataset. Three items (all of which were disposable CNCs) had to be discarded: Two of them were not CNCs upon closer analysis, and one of them was in a part of the corpus in which the PDF to TXT conversion did not work properly.

From this dataset, a number of strategies emerged that authors seem to commonly employ when introducing CNCs. Although some CNCs may not be perfectly categorized in one or the other strategy, we believe this is still a useful first step toward explaining the data. Table 5 shows the number of CNC occurrences that were categorized according to each strategy. As can be seen, the distribution of strategies was substantially similar for both recurrent and disposable CNCs, both in their first occurrence and whenever they were reused. In other words, we found no factors that explain what specifically distinguishes the CNCs that occur often in the articles from those that appear only a few times. In the following, we describe each of the strategies.

5 General discusion

In the analyses above, we investigated a number of distributional properties associated with complex nominal compounds made up of three or more words. In this investigation, we used the Uniform Information Density hypothesis to make predictions concerning these properties. In particular, we predicted that we would find the following results: CNCs would cluster toward the end of scientific articles, CNCs would be supported by the context in which they are used, and CNCs would be reused often once they are introduced. To answer these questions, we performed a quantitative analysis of the 16,920 CNCs present in 162 scientific articles collected from the fields of Biology, Linguistics and Economics, which in turn were divided into Introduction, Middle, and Conclusion sections, and a qualitative analysis of a small subset of them. The CNCs present in these sections were automatically identified and subsequently filtered using both an automatic and a manual method.

For the first prediction, we fit a linear mixed-effects model using as the dependent variable a log-scaled CNC Proportion (to control for section and CNC lengths). We found no differences in CNC Proportion between Introduction, Middle and Conclusion. In other words, we found no evidence in favor of our initial prediction. However, we found significant differences in the CNC use in the three fields of our corpus: Economics articles had a higher CNC Proportion than Biology articles, which in turn had a higher CNC proportion than Linguistics articles. Future research should examine the specific characteristics of CNC use in each of these research fields.

One could suggest that the lack of differences in CNC use between the sections reflects the audience that writers had in mind when preparing their text: authors could try to make the Introduction and Conclusion friendlier to broader audiences, but to concentrate the study’s technical aspects in the Methods and Results sections. Figure 2, however, does not support this idea: the proportion of CNCs in the Middle section is not higher than in the other sections. Indeed, in this case we would still expect more CNCs in the Conclusion than in the Introduction, but that is not what we found.

For the second prediction, we performed both a quantitative analysis investigating whether CNCs are preceded by their subparts in the article text, and a qualitative analysis identifying the strategies used by authors to introduce them. From the quantitative analysis, we found that the majority of CNCs are preceded by very few bigram subparts, and a third of them are not preceded by bigram subparts at all, suggesting that CNCs were not supported by their context.

A better picture of the contextual support provided to CNCs was then acquired through the qualitative analysis. It showed that CNCs are supported by the context, but in ways that are more sophisticated than simple word-pattern repetitions, which therefore could not be captured by our quantitative measure. Indeed, the majority of the CNCs examined were gradually preset over the course of several paragraphs, with strategies that included embedding in more complex structures single words (rather than bigrams) that were ultimately part of the CNC, or using semantically similar words to hint at the CNC meaning. Among those that were not introduced, there was a number of CNCs that constituted either field or scientific jargon, so that (we assume) the author(s) considered that explicit introduction was not necessary. This pattern is consistent with the predictions of the UIDh. Once the simpler structures are introduced, we assume that the mental model kept by the reader about the linguistic characteristics of the text is updated, making them more ‘available’, increasing the probability that these structures will recur, and therefore reducing their informativeness. With the simpler structures conveying less information, the channel capacity would be underused if newer and more complex structures are not introduced that carry more information.

Note that the scientific articles we analysed were collected from high impact journals. While these may be representative of what ‘good’ scientific articles look like, they may not be representative of the scientific register as a whole. That is, it is possible that the fact that the CNCs are normally gradually preset in the analysed articles is an artifact of their higher impact, and that scientific articles published in lower impact journals may contain a higher proportion of CNCs used in contexts that are not helpful for their understanding. Indeed, we believe that articles that are judged as ‘hard to read’ may be ‘hard’ precisely because they contain frequent peaks of information, many of which may involve CNCs.

Throughout this study, we made assumptions about the difficulty associated with CNCs, claiming that, based on the UIDh, a CNC should only appear in a scientific paper if its context introduces it enough. We made no distinction between CNCs such as “heart rate variability” (quite familiar to a lay reader) and “start arm barrier” (not common at all). Indeed, we had no way to assess the real difficulty associated with any given CNC. Therefore, in the future it may be useful to approach this question experimentally, comparing the difficulty participants experience when confronted with familiar versus unfamiliar CNCs, presented in context versus in isolation. In addition, given the difference in familiarity between the aforementioned examples, it is likely that the first would require much less contextual introduction (and would cause a lower density peak) than the latter. This may have been one reason why so many CNCs were not preceded by bigram subparts.^[18] In order to consider this possibility, we looked up the CNCs of our corpus on Wikipedia, assuming that, if a sequence has its own Wikipedia article, then it is familiar. Denote by inW(w _{1, …,n} ) a function that is 1 if the sequence of words w _{1, …,n} has its own article in Wikipedia, and 0 otherwise. We calculated a familiarity score f for each CNC composed of words w ₁, w ₂ and w ₃ using the formula below, producing scores ranging from 0 to 2.75, composed by the sum of three terms: one in which the entire CNC is looked up in Wikipedia (which will be either 0 or 1); a second term in which the CNC’s bigram subparts are looked up in Wikipedia (resulting in 0, 0.5 or 1, depending on how many bigram subparts have their own pages); and a final term corresponding to whether each of the CNC words have their individual Wikipedia articles (producing 0, 0.25, 0.5 or 0.75).^[19]

Figure 4 shows how the number of preceding subparts relates to CNC familiarity. In all familiarity levels (except for f = 2.5, for which we only have a single CNC repeating 14 times), the median number of preceding subparts is between 1 and 4, i.e., very close to zero. While there is no clear trend indicating that more familiar CNCs are preceded by fewer subparts, we cannot rule out this possibility either. Future studies should consider in more detail the effect of familiarity on the degree to which CNCs are preset. Since familiarity may change over time (a CNC like “database management system”, familiar today, barely existed 60 years ago), it might be useful to include date of publication as an additional covariate in such an analysis.^[20]

Figure 4:

Points represent CNCs with a given familiarity score and a given number of preceding subparts. While in (a) we consider all 3-word CNCs (14,357 data points), in (b) we consider only the first occurrence of a CNC in a given article (9,956 data points).

Given the failure of the quantitative measure we used to capture the kind of preset used by authors when introducing CNCs, and given the complex characteristics we found during the qualitative analysis, we believe that a better quantitative measure might have been the number of semantically related words preceding the CNC. Consider (8), where the target CNC is high reward probability side, the previous occurrences of the CNC words are bold, and the semantically related words are underscored.^[21] In addition, each underscored word is tagged with a number indicating the CNC word that it is related to (for example, the word low is tagged with a 1 because it is related to the first word of the CNC, high). It is clear that the CNC high reward probability side is preset not only by having high reward, reward probability and probability side appear before it, but also by having contextual cues that imply the possibility that these word combinations are likely to occur. For example, the words choice, left and right indicate that there is a side, even if the word side never appears in the article before the CNC. Similarly, upon reading about the probability of receiving a reward, the reader is naturally able to expect the wording reward probability.

Of course, this measure also has shortcomings. For example, it does not take into account sentence complexity, and does not consider how the words are combined in the context. In addition, counting the number of semantically related words is subjective, requires a large amount of resources, and is beset by problems. For example, how can we best define what constitutes the “preceding text”? Should we consider only the section where the CNC appears, or should we consider the whole article? If the whole article, how can we fairly compare longer versus shorter articles? How can we compare a CNC that appears around the beginning of an article with another one that appears near the end? Some of these problems may be solved with computational approaches that can tag large amounts of data and require fewer resources; however, whether these approaches would lead to results that are similar to those produced by humans is an open question. We intend to explore this issue in future work.

As for the last prediction investigated in this paper, that CNCs would be reused often after their first use, we also found no evidence that this is the case. Indeed, the vast majority of CNCs (83.7 %) were never reused. CNCs do not seem to become ‘ad hoc names’ for new concepts as suggested by Salager (1984), and instead have a very local use, being immediately discarded. This held true both for CNCs composed of three words, and for those composed of four or more words.

One should be careful before treating this as evidence against the UIDh for two reasons. First, while we did not investigate this possibility in this paper, it is possible that very dense CNCs become acronyms after their first use. Like CNCs, acronym use has substantially increased in the last decades (e.g., Barnett and Doubleday 2020). Note that this was precisely the case with the two most commonly used CNCs in this very paper (i.e., complex nominal compound and Uniform Information Density hypothesis). Second, it is possible that repeating CNCs actually conveys too little information, and that, instead of repetition, CNCs would actually undergo deletion of some of their words after their first use. For example, a CNC such as high reward probability side could be reused as high reward side or even high side in contexts where the missing words are obvious. If that is the case, then we should find very different results if we use a relaxed version of the quantitative measure used in this paper. We intend to explore this idea in future work.

The three predictions discussed above were made taking into account the fact that the transmitter does not have full knowledge about the expectations of the receiver, and is therefore not able to calculate perfectly the amount of information of each word from the perspective of the receiver. As discussed earlier, we suggested that each communication partner keeps a probabilistic model of the communication process and updates this model as new words are transmitted through the channel. As the reader progresses through the text, the two models would slowly converge toward similar probabilities for most words. But is this really the case? The results of our second prediction point in that direction. As Table 5 suggests, authors seem to organize their articles so that, in most cases, CNCs are only used when they can be understood. This makes it surprising that we did not find more CNCs toward the end of the articles. Further research is necessary to investigate why this was the case.

In summary, in this paper, we found no evidence that CNCs are more frequent toward the end of scientific articles, nor that CNCs become ‘ad hoc names’ after their first use in an article. These possibilities, however, could not be completely discarded, and thus our findings constitute no evidence in favor of or against the UIDh. On the other hand, we found that CNCs are preset in their context, but that the way in which this presetting occurs is complex, and could not be captured by the strict quantitative measure we used. This latter finding constitutes evidence in favor of the UIDh. In addition, this finding suggests that speakers do keep a probabilistic estimation of the communication process, updating this model as the communication unfolds.

6 Conclusions

In this study, we investigated the distributional properties of complex nominal compounds composed of at least three words. To do so, we performed a quantitative and a qualitative analysis of the CNCs identified in our corpus, the Sciper Corpus. Some of our results constitute evidence in favor of the Uniform Information Density hypothesis. From the qualitative analysis, a number of improvements arose that can be made on the quantitative measures we used. In the future, we intend to investigate how the quantitative results differ if performed considering these improvements.