A statistical exploration of the effects of phoneme class contacts in German noun-noun compounds

Annelen Brunner Leibniz Institute for German Language, Mannheim, GERMANY Working paper: A statistical exploration of the effects of phoneme class contacts in German noun-noun compounds Version: 2025/05/23 This working paper describes a study that explores phoneme contact preferences in German noun-noun compounds with statistical methods on the basis of a large dataset (over 707,000 compounds). The following hypotheses are tested based on surprisal values: 1) vowel contacts are avoided  confirmed; 2) contacts of phonemes from the same phoneme class are avoided  not confirmed; 3) contacts where a phoneme with higher sonority is followed by a phoneme with lower sonority are avoided (syllable contact law)  largely confirmed. Acknowledgments: This working paper was developed in discussion with Alexander Koplenig, Katrin Hein and Stefan Engelberg, all from Leibniz Institute for German Language, Mannheim, GERMANY Supplemental material for this working paper is available via OSF: https://osf.io/e7afd Contact address: [email protected] 1 Research questions In this study, we examine the phoneme class contacts that occur in a large number of German noun-noun compounds comprising two simplex nouns and look for indications whether certain phonological contacts are preferred or avoided in compounding. We will focus on the following aspects: 1. Sonority: For syllable contacts within the same morphological word, phoneme contact where a more sonorous phoneme is followed by a less sonorous phoneme are preferred, while the opposite is avoided (‘syllable contact law’; cf. e.g. Hall 2011, p. 230-236; Vennemann 1988, p. 40). Do we find evidence that this rule is also relevant for contacts between the constituents in compounding? 2. Vowel contacts (hiatus): Direct contact between vowels is sometimes avoided, e.g. in the context of derivation (e.g. amerika-n-isch, ge-g-essen). Can we observe such a trend in compounding? 3. Contact of phonemes from the same phoneme class: Contacts of identical phonemes tend to be avoided. As we don’t have coding for individual phonemes in our data, we will address the more general case of contact between phoneme classes and check whether we can observe any tendencies that this is avoided. 2 Data set Our data is extracted from the “KoGra Untersuchungskorpus” 1 , which comprises roughly 7 billion tokens and is a subset of the German Reference Corpus (DeReKo, Release 2017-II 2 ). This corpus consists mainly of German newspaper texts (over 90%), but also contains some literary texts, and about 6% spoken language material (cf. Bubenhofer/Konopka/Schneider 2014). Detailed morphological information was added with a custom word analyzer based on the Canoo Language Tools 3 . This made it possible to automatically extract a large collection of over 489 million nominal compound tokens that serves as the basis for our studies on word formation. For this study, we used the same dataset as in Brunner/Engelberg/Hein 2021 (studies A, B and C). It consists only of compounds comprising of two simplex nouns, e.g. Stadthalle (‘town hall’). In addition to that, the constituents of the data set were annotated with categories from GermaNet (Hamp and Feldweg 1997; Henrich and Hinrichs 2010) and we kept only compounds for which a GermaNet match was found for each constituent. This had a cleanup effect, as it made is more likely that the compounds consisted of two plausible words. As there were still errors due to the automatic annotation, some additional manual cleanup was performed on the resulting list, removing compounds that contained constituents that were either not recognizable as words or were not simplex nouns. The resulting dataset comprises 707,910 compound types. 4 The automatic tool gave us a segmentation of each compound, but only lemmatized forms for the constituents. For a phonological study however, we need surface forms, so we inferred the segmentation of the compound surface automatically, using the available information about the lemmatized forms. This process was successful for 704,475 compounds, 99.5 percent of the original data set. 5 This is the data set this study is based on. A manual check of 200 randomly selected compounds revealed that 95% of the assigned surface segmentations were 1 Korpus des Projekts Korpusgrammatik. Leibniz-Institut für Deutsche Sprache: „Korpusgestützte Grammatik“. Grammatisches Informationssystem grammis. DOI: 10.14618/korpusgrammatik. URL: https://grammis.idsmannheim.de/korpusgrammatik/6615. 2 Deutsches Referenzkorpus / Archiv der Korpora geschriebener Gegenwartssprache 2017-II (Release: 01.10.2017). Mannheim: Leibniz-Institut für Deutsche Sprache. www.ids-mannheim.de/DeReKo. 3 http://www.canoonet.eu. Unfortunately, this website no longer exists. Parts of the content of canoonet were integrated into LEOdict (https://dict.leo.org/pages/about/ende/canoonet_de.html; accessed 23.05.2025) but the morphological analyzer is no longer accessible. 4 For a more detailed explanation on how this dataset was curated, cf. Brunner/Engelberg/Hein 2021 p. 9-12. 5 Cases where surface forms could not be assigned were due to errors in the automatic lemmatization of the constituents, especially the head word. correct. In the sample, all errors were due to problems in the original data, not the postprocessing step: Either the input data contained faulty segmentations or the entries were not valid compounds at all. Any linking elements were treated as part of the first constituent. So for example, Staatsanwalt will be segmented into staats (Staat plus linking element s) and anwalt. We have the information for which compounds a simple combination of the lemmatized forms of the constituents is identical to the surface form (e.g. Afrikabild – Afrika plus Bild). This subgroup contains no linking elements and was tested separately in our analyses (cf. section 3.2, end). To study phoneme contact in compounding we needed information about the last phoneme of the modifier (including linking element) and the first phoneme of the head for each compound. We devised a rule-based system to derive the phoneme classes based on the orthographical surfaces, identifying eight phoneme classes that are ranked by sonority. Table 1 shows the labels used for the phoneme classes. They are comprised of a letter as shorthand for the name of the class and a number that indicates the level of sonority with 1 being the least sonorous and 8 the most sonorous phoneme class. Appendix 1 documents the exact rules that were applied to assign labels. Note that in this study, the approximant/semivowel [j] (first phoneme of e.g. Jahr) is classified as a high vowel and [ɐ] (the last phoneme of e.g. Leber) is considered a low vowel. The glottal stop [ʔ] is not considered a phoneme but a phonetic feature because in Standard German it appears predictably in certain contexts (word-initial or stressed syllables that begin with a vowel). Table 1: Labels for the phoneme classes, ranked by sonority (examples are not exhaustive) label phoneme class examples 1_p plosive [t] [g] [k] [b] [d] [p] 2_f fricative [s] [f] [ʃ] [h] [x] 3_n nasal [n] [m] [ŋ] 4_l liquid (phoneme l) [l] [l] 5_l liquid (phoneme r) [ʁ] [ʀ] [r] 6_v high vowel [ɪ] [ʊ] [i] [ʏ] [i] [u] [y] [j] 7_v mid vowel [ə] [ɛ] [ɔ] [œ] [e] [o] 8_v low vowel [a] [aː] [ɐ] It must be noted that our data set contains a relatively high number of proper names as well as foreign language elements, especially English and French constituents. Though we tried to account for foreign languages in our rules to some extent, this led to some ambiguities. A manual check of a sample of 200 compounds revealed that the assignment of the phoneme class was correct in 97% percent of the cases. The errors were all due to foreign language constituents. Table 2 shows an excerpt from the dataset. Table 2: Excerpt from the data set (total size: 704,475 entries) frequency compound modifier surface head surface phoneme class end of modifier phoneme class beginning of head 2911504 Sonntag sonn tag 3_n 1_p 1758319 Donnerstag donners tag 2_f 1_p 1269620 Bürgermeister bürger meister 5_l 3_n 1160962 Wochenende wochen ende 3_n 7_v 781104 Geburtstag geburts tag 2_f 1_p 696116 Fußball fuß ball 2_f 1_p 467966 Landtag land tag 1_p 1_p 465068 Landkreis land kreis 1_p 1_p 449355 Kindergarten kinder garten 5_l 1_p 428906 Jahrzehnt jahr zehnt 5_l 1_p 3 Empirical exploration There are two possibilities to use the quantitative data: Either we only look at the types in our data set or we also incorporate the frequency information we have for these types. Both approaches have merit: Considering only the types gives us an idea which distinct compounds occur in real life data. However, as each compound type has the same impact on the result, very infrequent compounds and errors in the data gain strong influence. If we weigh the types according to their frequencies, we get information about which distinct compounds are common, which is arguably more useful to tell us something about productivity and entrenchment in German compounding. However, as the frequency curve in our data set follows a Zipf-like distribution with few very frequent entries and a huge tail end of hapax legomena, the most frequent entries gain a strong influence on the results (c.f. figure 1). As both aspects are interesting, we will look at the distributions on both ways and discuss the results. Figure 1: Frequency curve for the compounds in our data set (logarithmic scale) 3.1 Data distribution matrices In the following matrices (figures 2-5), the phoneme classes of the last phoneme of the modifiers are plotted on the y-axis and the phoneme classes of the first phoneme of the head words on the x-axis. Each scale is ordered according to sonority (least to most sonorous). The cells show the values for contacts between those phoneme classes. For example, figure 1 shows that a contact between a fricative as last phoneme of the modifier and a plosive as first phoneme of the head (shorthand: [2_f + 1_p]) occurs in 47,751 compounds types. The cells forming the diagonal line from upper left to lower right represent the cases where phonemes of the same class come into contact. If the hypothesis holds true that contacts where the second element is less sonorous than the first are preferred, the cells below this diagonal line should have generally higher values than the cells above it. The 9 cells in the lower right quadrant of the matrix show the contacts between vowel phoneme classes. If the hypothesis that vowel contacts are avoided is true, those should have lower values. Figure 2 shows a matrix of type counts for all combinations of phoneme class contacts. It is evident that some phoneme classes, e.g. plosives, are much more common than others. To get an impression whether there are distributional preferences besides simple frequency of the phoneme classes, we calculated the number of cases that would appear in each matrix cell, if the distribution were random – the expected distribution. 6 For figure 3, we calculated which percentage of the expected values the real values represent. If a real number corresponded exactly to the expected number, the cell value would be 100. Looking at the contact [2_f + 1_p] again, we see that the value is 102.83. This means that it is 102.83% of the expected value, i.e. only slightly larger than expected. Figure 3 indicates that the deviations from the expected numbers are not large, the cell values range between 73 and 111. The distribution of contacts between phoneme classes in compounds seems quite uniform on type level. There might be a slight trend that contacts with decreasing sonority are favored, which would support our hypothesis, but the picture is not clear. We can, however, observe a trend that contacts between vowels are avoided, especially those where vowels of same phoneme class come into contact and those where the second vowel is more sonorous. When looking at contacts between identical phoneme classes of any kind, we see that these are a little less frequent than expected as well. 6 There are statistical tests to determine whether the distribution in a matrix deviates from the expected distribution in a significant way; a common one is the Chi square test. If we run it for the matrices in figures 2 and 4, we do get significant results – so we know that the distribution is not random. However, when applied to a very large dataset such as ours, the statistical test can be misleading, as it picks up on small effects. In addition to that, the assumption of independence underlying the statistical tests does not hold for language data in general, which makes it even more problematic to rely on the test result. We therefore opted for a purely descriptive approach here. Figure 2: Distribution of compound types according to their phoneme contacts (type-based) Figure 3: Percentage of real frequencies in relation to expected frequencies (type-based) We will now look at the token-based distribution, taking into account the frequency for each compound type (figure 4 and 5). The deviations from the expected distribution are much larger – while in figure 2 the values ranged only between 73 and 111, in figure 5 they range from ca. 38% (less than half the expected value) to nearly 300 (three times more than expected). We again observe a slight preference for contacts where the second phoneme is less sonorous and a dispreference for most vowel combinations, but there are outliers. As mentioned above, the frequency curve for our data is Zipf-like with a very steep slope (cf. figure 1). If we take a closer look at a few cells that stand out in particular with much higher values than expected, we see that compounds from the highest frequency bands have indeed an impact: The combination [3_n + 7_v] (nasal + mid vowel) has the highest value with 294.55%. Compounds with this combination include Wochenende which is one of the most frequent compounds in our data. Similarly, the combination [3_l + 6_v] (liquid + high vowel), which stands out with a value of 274.44 %, includes the very frequent word Schuljahr. Just removing the words Wochenende and Schuljahr already leads to a matrix where the respective cell values are much less extreme, but in order to have a more objective picture of the influence of very frequent compounds, we recalculated the matrix without the 100 most frequent entries (figure 6). The distribution becomes notably more even. The matrix now looks more similar to the type-based matrix (figure 3), albeit with stronger deviations from the expected distribution in both directions. Only the cell [6_v + 4_l] (high vowel + liquid) still has the value 234,72, i.e. more than twice the expected items. This cell contains frequent compounds as well (e.g. EU-Land, Industrieland), but is not dominated by a single entry. Figure 4: Distribution of compound tokens according to their phoneme contacts Figure 5: Percentage of the real numbers in relation to the expected numbers (tokens) Figure 6: Percentage of the real numbers in relation to the expected numbers (tokens without the 100 most frequent compounds) As mentioned at the beginning of the section, the plots shown here are based on the token counts in our data. We recalculated the same plots omitting the 100 most frequent compounds to control for the effect of very frequent compounds (cf. figure 6). Though the boxplots look slightly different, the permutation tests gave the same results. We also calculated the plots based on type counts (cf. figures 2 and 3). Again, the permutation tests were significant for the same group contrasts. In section 2 we reported that linking elements were counted as part of the modifier. This implies that in cases where a linking element is present, the last phoneme of this linking element was used in our study. As linking elements are a special (though common) case in compounding, we repeated our calculations excluding all compounds where linking elements were detected. Still, we could confirm the same significant differences on the basis of tokens, tokens without the 100 most frequent compounds and on basis of types. To sum up, the following results proved stable for different configurations:  Contacts between vowel classes have significantly higher surprisal values.  Contacts that are ‘bad’ according to the syllable contact law have significantly higher surprisal values than those that are regarded as ‘good’.  There were no significant results for the other tests. 4 Appendix 4.1 Rules for assigning phoneme classes The following rules were applied to assign a phoneme class to the orthographical form of the last element of the modifier and the first element of the head word. The strategy was the same in both tasks: We looked at the end of the word (for modifiers) and at the beginning of the word (for heads) and checked longer letter combinations first. Each table lists the rules in the order in which they were applied to the orthographical surface. If a rule matched, execution was stopped and the corresponding phoneme class label was assigned. Note that some rules in these tables are quite idiosyncratic for our dataset as we had to account for some foreign, especially English and French spellings and decided to include explicit rules for special cases to deal with ambiguity. Rules for the phoneme class of the last phoneme of the compound modifier execution order letter combinations label class name 1 ["aille", "view"] 6_v high vowel 2 ["sch", "löw"] 2_f fricative 3 ["ieh"] 6_v high vowel 4 ["th", "qu", "ig"] 1_p plosive 5 ["ch", "sh", "ph", "pf"] 2_f fricative 6 ["ng"] 3_n nasal 7 ["rh"] 5_l liquid (r) 8 ["ie", "ih", "uh", "ou", "üh", "ei", "ai", "eu", "äu", "au", "ew"] 6_v high vowel 9 ["eh", "oh", "äh", "öh", "ow", "aw"] 7_v mid vowel 10 ["ah"] 8_v low vowel 11 ["t", "d", "g", "k", "b", "c", "p"] 1_p plosive 12 ["ß", "z", "s", "f", "v", "h", "x"] 2_f fricative 13 ["n", "m"] 3_n nasal 14 ["l"] 4_l liquid (l) 15 ["r"] 5_l liquid (r) 16 ["i", "u", "y", "ü", "j", "q10"] 6_v high vowel 17 ["e", "o", "ä", "ö", "é"] 7_v mid vowel 18 ["a", "à"] 8_v low vowel Rules for the phoneme class of the first phoneme of the compound head execution order letter combinations label class name 1 ["sch"] 2_f fricative 2 ["ch", "sh", "ph"] 2_f fricative 3 ["ei"] 8_v high vowel 4 ["t", "d", "g", "k", "b", "x", "p", "c", "q", "z"] 1_p plosive 5 ["s", "f", "v", "w", "h"] 2_f fricative 6 ["m", "n"] 3_n nasal 7 ["l"] 4_l liquid (l) 10 “q” triggers an annotation as 6_v (high vowel) because in our data set the only modifier that ended in the letter “q” was “IQ”. As an abbreviation this is pronounced with an “u” sound at the end. 8 ["r"] 5_l liquid (r ) 9 ["i", "u", "ü", "j", "y"] 6_v high vowel 10 ["e", "o", "ä", "ö"] 7_v mid vowel 11 ["a"] 8_v low vowel 4.2 Table of surprisal values, based on compound tokens Additional tables for other data configurations are available at the OSF repository. Modifier (last phoneme) Head (first phoneme) Frequency of combination Total frequency of modifier class p(head class| modifier class) surprisal same class contact vowel class contact sonority sonority difference 1_p 1_p 9782346 26229355 0,372954 1,42293 yes no same 0 1_p 2_f 10066856 26229355 0,383801 1,381569 no no more -1 1_p 3_n 2236945 26229355 0,085284 3,551581 no no more -2 1_p 4_l 529642 26229355 0,020193 5,630021 no no more -3 1_p 5_l 1230246 26229355 0,046903 4,414164 no no more -4 1_p 6_v 334875 26229355 0,012767 6,291416 no no more -5 1_p 7_v 696491 26229355 0,026554 5,234934 no no more -6 1_p 8_v 1351954 26229355 0,051544 4,278064 no no more -7 2_f 1_p 9713135 17242357 0,56333 0,827948 no no less 1 2_f 2_f 3886129 17242357 0,225383 2,149551 yes no same 0 2_f 3_n 1657100 17242357 0,096106 3,379224 no no more -1 2_f 4_l 430253 17242357 0,024953 5,324628 no no more -2 2_f 5_l 195860 17242357 0,011359 6,45999 no no more -3 2_f 6_v 122602 17242357 0,007111 7,135831 no no more -4 2_f 7_v 212057 17242357 0,012299 6,345361 no no more -5 2_f 8_v 1025221 17242357 0,059459 4,07195 no no more -6 3_n 1_p 8074539 16724684 0,482792 1,050527 no no less 2 3_n 2_f 4626159 16724684 0,276607 1,854092 no no less 1 3_n 3_n 674675 16724684 0,04034 4,631642 yes no same 0 3_n 4_l 463551 16724684 0,027717 5,173107 no no more -1 3_n 5_l 373353 16724684 0,022323 5,485295 no no more -2 3_n 6_v 176727 16724684 0,010567 6,564313 no no more -3 3_n 7_v 1444712 16724684 0,086382 3,533125 no no more -4 3_n 8_v 890968 16724684 0,053273 4,230462 no no more -5 4_l 1_p 3251965 7325810 0,443905 1,171677 no no less 3 4_l 2_f 2355677 7325810 0,321559 1,636847 no no less 2 4_l 3_n 516733 7325810 0,070536 3,825497 no no less 1 4_l 4_l 181935 7325810 0,024835 5,331493 yes no same 0 4_l 5_l 211360 7325810 0,028851 5,115214 no no more -1 4_l 6_v 290147 7325810 0,039606 4,658132 no no more -2 4_l 7_v 164437 7325810 0,022446 5,477381 no no more -3 4_l 8_v 353556 7325810 0,048262 4,372978 no no more -4 5_l 1_p 5547747 13783101 0,402504 1,312927 no no less 4 5_l 2_f 4131515 13783101 0,299752 1,738158 no no less 3 5_l 3_n 2476063 13783101 0,179645 2,476781 no no less 2 5_l 4_l 486991 13783101 0,035332 4,822862 no no less 1 5_l 5_l 300124 13783101 0,021775 5,521198 yes no same 0 5_l 6_v 168566 13783101 0,01223 6,353443 no no more -1 5_l 7_v 203811 13783101 0,014787 6,079525 no no more -2 5_l 8_v 468284 13783101 0,033975 4,879373 no no more -3 6_v 1_p 1159659 2641025 0,439094 1,187397 no no less 5 6_v 2_f 880392 2641025 0,333352 1,58488 no no less 4 6_v 3_n 188100 2641025 0,071222 3,811526 no no less 3 6_v 4_l 169247 2641025 0,064084 3,963896 no no less 2 6_v 5_l 69348 2641025 0,026258 5,2511 no no less 1 6_v 6_v 30441 2641025 0,011526 6,438938 yes yes same 0 6_v 7_v 31913 2641025 0,012084 6,37081 no yes more -1 6_v 8_v 111925 2641025 0,042379 4,560494 no yes more -2 7_v 1_p 1724384 4290094 0,401946 1,314928 no no less 6 7_v 2_f 1399232 4290094 0,326154 1,616374 no no less 5 7_v 3_n 640368 4290094 0,149267 2,744036 no no less 4 7_v 4_l 148193 4290094 0,034543 4,85546 no no less 3 7_v 5_l 119316 4290094 0,027812 5,16815 no no less 2 7_v 6_v 61574 4290094 0,014353 6,122544 no yes less 1 7_v 7_v 43610 4290094 0,010165 6,620206 yes yes same 0 7_v 8_v 153417 4290094 0,035761 4,805479 no yes more -1 8_v 1_p 491683 1028146 0,478223 1,064245 no no less 7 8_v 2_f 273897 1028146 0,266399 1,90834 no no less 6 8_v 3_n 137226 1028146 0,133469 2,905419 no no less 5 8_v 4_l 32282 1028146 0,031398 4,993171 no no less 4 8_v 5_l 39380 1028146 0,038302 4,706438 no no less 3 8_v 6_v 15351 1028146 0,014931 6,065569 no yes less 2 8_v 7_v 12613 1028146 0,012268 6,34899 no yes less 1 8_v 8_v 25714 1028146 0,02501 5,321347 yes yes same 0 Bibliography Brunner, A., Engelberg, S. & K. Hein. 2021. The distribution of constituent words in nominal compounds and its impact on semantic interpretation: an empirical study. In: Journal of Word Formation 1: 7-36. Bubenhofer, N., Konopka, M. & R. Schneider (Eds.). 2014. Präliminarien einer Korpusgrammatik. Tübingen: Narr. Gibson, E., Futrell, R., Piantadosi S. P., Dautriche, I., Mahowald, K., Bergen, L. & R. Levy. 2019. How Efficiency Shapes Human Language. Trends in cognitive sciences 23 (5): 389407. Hall, T. Alan. 2011. Phonologie. Eine Einführung. Berlin / New York: De Gruyter. Hamp, B. & Feldweg, H. 1997. GermaNet – a Lexical-Semantic Net for German. In Proceedings of the ACL workshop Automatic Information Extraction and Building of Lexical Semantic Resources for NLP Applications. Madrid 1997. 9–15. Henrich, V. & Hinrichs, E. 2010. GernEdiT – The GermaNet Editing Tool. In Proceedings of the Seventh Conference on International Language Resources and Evaluation (LREC 2010), 2228–2235. Valletta, Malta. Koplenig, A. 2019. A non-parametric significance test to compare corpora. PLoS ONE 14(9). [online: https://doi.org/10.1371/journal.pone.0222703] Koplenig, A., Wolfer, S., Rüdiger, J. O., & Meyer, P. 2024. Human languages trade off complexity against efficiency. In: Charlottesville: OSF Preprints. [online: https://osf.io/preprints/osf/8xgqz_v1] Shannon, C. E. 1948. A Mathematical Theory of Communication. Bell System Technical Journal 27: 379-423. Tu, N. D. T.. 2024. Eine korpuslinguistische Untersuchung zur lexikalischen Vielfalt von direkten und indirekten Redeeinleitern. Mannheim: IDS-Verlag. [online: https://pub.idsmannheim.de/laufend/idsopen/idsopen06.html] Vennemann, T. 1988. Preference Laws for Syllable Structure and the Explanation of Sound Change. Berlin / New York: De Gruyter.