Jump to ContentJump to Main Navigation
Show Summary Details
More options …

Folia Linguistica

Acta Societatis Linguisticae Europaeae

Editor-in-Chief: Fischer, Olga / Norde, Muriel

Folia Linguistica
IMPACT FACTOR 2018: 0.463
5-year IMPACT FACTOR: 0.647

CiteScore 2018: 0.59

SCImago Journal Rank (SJR) 2018: 0.284
Source Normalized Impact per Paper (SNIP) 2018: 0.971

See all formats and pricing
More options …
Volume 40, Issue 1


Word form shapes are selected to be morphotactically indicative

Andreas Baumann / Christina Prömer / Nikolaus Ritt
Published Online: 2019-07-28 | DOI: https://doi.org/10.1515/flih-2019-0007


This paper explores the hypothesis that morphotactically ambiguous segment sequences should be dispreferred and selected against in the evolution of languages. We define morphotactically ambiguous sequences as sequences that can occur both within morphemes and across boundaries, such as final /nd/ or /mz/ in ModE, which occur in simple forms like wind or alms and in complex ones like sinned or seems. We test the hypothesis in two diachronic corpus studies of Middle and Early Modern English word forms ending in clusters of sonorants followed by /d/ or /t/ and /s/ or /z/. These clusters became highly frequent after the loss of unstressed vowels in final syllables and were highly ambiguous when they emerged. Our data show that the ambiguity of these final clusters was indeed reduced so that the distribution of the final clusters became increasingly skewed: clusters ending in voiceless coronals became significantly clearly indicative of simple forms, while clusters ending in voiced ones came to signal inflectional complexity more reliably.

Keywords: coda clusters; corpus linguistics; inflection; morphonotactics

1 Introduction

The sound shapes of morphologically complex words normally differ from those of simple ones. For example, the plural dogs ends in a /z/ that the base form dog lacks, and the final /t/ in the past tense form kicked is absent in the base form kick. It is the raison d’être of such ‘morphotactic’ differences to indicate ‘morpho-semantic’ ones: because of the meanings they express, they are produced, learnt, shared among speakers, and transmitted across generations. Since some patterns indicate morpho-semantic complexity better than others, they are also transmitted more successfully, and become more frequent in the world’s languages.

Complex word forms work better, for example, when their shapes mirror their morpho-semantic complexity by being themselves complex (Dressler 1985): that played expresses both the ‘act of playing’ and that it ‘occurred in the past’, for example, is mirrored in the shape of played, which consists both of the stem play and the past tense morph –ed. In comparison, the irregular past tense came cannot easily be divided into separate bits and represents its semantic complexity less transparently than play+ed does (see Mayerthaler 1981; Dressler 1985; Wurzel 2001). Also suffixes are not all equally good either, but signal complexity better the more salient they are (Dressler 1985). Thus, the plural /ˈkɪsɪz/ kisses, with the syllabic variant of the suffix, is easier to identify than a hypothetical alternative /kɪsː/ would be, where the /s/-suffix more or less merges with the /s/ in the stem. It is well established, in other words, that the historical stability of complex word forms is affected by the relative ease with which they can be decomposed. 1

Although the ease with which a word form can be identified as complex obviously depends on its own shape, however, it also matters if the same shape occurs in simple word forms as well.

Consider word forms such as English eggs, for example. Its ending, i.e. the cluster /gz/, identifies eggs – unambiguously – as complex, because it can only arise when a {/z/} suffix is added to a stem that ends in /g/. Thus, eggs must be either pl, gen, or 3sgpres. Although English is morphologically poor, forms like eggs are common. Many coda clusters can only arise in suffixation, for example /bz/ (as in robes), /ðz/ (mouths), /mz/ (gyms), /θs/ (Keith’s), /gz/ (Meg’s), /vz/ (lives), /dz/ (bids), /md/ (seemed), /n(d)ʒd/ (ranged), /vd/ (lived), /fθ/ (fifth), /ksθ/ (sixth), /dθ/ (width), /pθ/ (depth), etc. All such clusters signal complexity (almost) unambiguously.

On the other hand, there are forms such as filled, banned, spilt, meant, or sins. Their final clusters are also created by suffixation, but they have homophones (‘sound-alikes’) in simple items such as hand, kilt, tent, or rinse, and this makes them ambiguous. As suggested in recent work by Dressler, Dziubalska-Kołaczyk and others (Dressler and Dziubalska-Kołaczyk 2006; Dressler et al. 2010), this kind of ambiguity can impede processing, bias speakers against the use of ambiguous word forms, and affect their transmission. If it does, the effects of such a bias should be detectable in language history, and this paper investigates if they are.

To test the prediction, we report two studies on corpus data from Middle and Early Modern English. The first of them focusses on word forms that end in sonorants followed by /t/ or /d/, and the second one investigates forms that end in sonorants followed by /z/ or /s/. The clusters still occur both in morphologically simple forms such as build, bird, hand, kilt, hunt, ounce, false, farce, or rinse, and in complex ones such as kill+ed, blurr+ed, sinn+ed, spil+t, burn+t, wall+s, pence (penny.pl), or bin+s. Most of them emerged during the Middle English period and were ambiguous from the start.

We count type and token frequencies of relevant word forms in Middle English and Early Modern English texts in order to identify diachronic changes in their ambiguity, i.e. how strongly different cluster types correlate with morphological simplicity on the one hand, or morphological complexity on the other. For that purpose, we compare and relate the developments of clusters ending in voiced /d/ or /z/ to those of clusters ending in voiceless /t/ or /s/. We show that the structural ambiguity of all clusters decreased significantly both on the type level and (albeit with some restrictions) also on the token level. We interpret our results as suggesting (a) that word forms are indeed preferred to have structurally unambiguous shapes, and (b) that this preference applies an evolutionary pressure on simple and complex forms to assume different phonotactic shapes.

In section two, we explain our hypothesis, specify the predictions it implies, justify our choice of data, and describe the methods by which we analyse them. Sections three and four report and discuss our two studies, and section five provides a summary and a brief outlook.

2 Hypotheses, methods, and data

We assume that speakers find it difficult to process morphotactically ambiguous word forms such as kill+ed and build, or grin+s and rinse. This requires justification. After all, English admits find and fined, build and billed, sand and tanned, pact and packed, Max and Macs, lift and miffed, so speakers do not seem to mind morphotactic ambiguity, and resolve it easily in most contexts. 2 Nevertheless, ambiguous word shapes do create problems. When a form could potentially result from affixation, it seems that speakers automatically attempt to decompose it, both when it actually is complex and when it is not. Thus, in experiments (Post et al. 2008), word forms like eggs, packed, pact, rose, and even act, took slightly (milliseconds) longer to process than forms like lamp, bee, bid, etc. This means that the morphotactic ambiguity of word form shapes incurs unwarranted processing costs, and although they may be low, speakers might nevertheless be sensitive to them and develop a bias against ambiguous word shapes.

We hypothesize that such a bias against ambiguous word shapes should reduce their ambiguity over time. Language history results from massively parallel and iterated transmission processes. To acquire and maintain their historical stability, words are used and learnt repeatedly by many individuals. Each time, they need to pass the same articulatory, auditory and cognitive bottlenecks. This can amplify the effects of even slight biases (Fehér et al. 2016; Kirby et al. 2014; Reali and Griffiths 2009; Smith et al. 2017), so that they accumulate over longer time-spans. Thus, the effects of a potential bias against morphotactically ambiguous word shapes should show up in language histories, although they are certainly weak. If they do, then what we expect to see is that ambiguous word forms should become less ambiguous over time.

To check if that is indeed the case, we operationalize ‘ambiguity’ as follows: from the perspective of perception, it is clearly smallest in words that are either unambiguously complex (such seem+s, which ends in /mz/), or unambiguously simple (such as ink, which ends in /ŋk/). Conversely, ambiguity is highest in word forms that are as often complex as they are simple. Forms that end in /nd/ like hand or tann+ed come to mind. – Ambiguity is a two-way relation, however. One cannot only ask how often a word shape represents morphologically simple forms and how often complex ones, but also how often simple or complex forms are represented by one or more types of word shape. This is the perspective of production, where one needs to choose a phonotactic shape for representing a specific morphotactic structure. For past tense forms of verbs in /l/, for example, there is the choice between /ld/ (as in filled or spoiled), and /lt/ (as in spilt or spoilt). Here, the best case is when there is only a single candidate. In the worst case, all candidates occur equally often. Taking the two-faced nature of ambiguity into account, our measure of ambiguity incorporates both perspectives.

Technically speaking, we operationalize it by means of the phi-coefficient φ (Cumming 2014), which measures the correlation between two binary variables. In our case, these are (a) ‘morphological complexity’ (complex vs. simple), and (b) ‘phonotactic shape’ (e.g. sonorant+/d/ vs. sonorant+/t/). The phi-coefficient is close to 1 if there is a tight correlation between phonotactic shape and morphological complexity. This is the case, for example, if simple items are almost always associated with final /t/, while complex items are associated with final /t/, and vice versa. If there is no clear relationship between morphological complexity and phonotactic shape, on the other hand, the phi-coefficient is close to 0. This means that one can neither infer the morphological structure of a word from its phonotactic shape (in perception), nor predict its phonotactic shape form its morphological structure (in production).

Another thing we had to consider was that the effects of cognitively grounded, potentially universal biases, will not be equally identifiable in all languages or in all periods. Instead, the periods in which they are likely to reveal themselves are bound to be rare, because universal biases are by definition equally strong all the time. Thus, if there is a cognitive bias against ambiguous word shapes at all, any language stage should in principle (have come to) reflect it as much as any other, and significant changes should not be observable. 3 To detect any diachronic effects of a ‘universal’ bias, it is therefore necessary to identify a period in the history of a language in which, for whatever reason, it has come to be reflected suboptimally. Only then can one expect a language to (re-)adapt to the bias in its further evolution.

A language stage that meets this criterion is the Middle English period. At its beginning a sound change started to unfold that altered the shapes of English word endings radically and throughout the lexicon. It was a system-wide loss of vowels in unstressed final syllables, which is known as ‘schwa loss’. It first affected open syllables, and produced forms such as /ma:k/ ‘make’ from earlier /ˈmakə/, or /ˈkli:mb/ ‘climb’ from earlier /ˈkli:mbə/. Later, it affected checked syllables (i.e. syllables with codas, e.g. CVC), producing forms like /fa:md/ ‘famed’ from /fa:məd/, or /faild/ ‘failed’ from /failəd/. Schwa loss unfolded gradually over centuries. In open syllables, it started in the late tenth century, and was completed by the middle of the fifteenth. Checked syllables started to display it from the thirteenth century onwards, but it was not before the end of the sixteenth century that the change was completed (Minkova 1991, Minkova 2009).

In our context, the most relevant effects of schwa loss are the following. First, it created many morphologically simple word forms from previously complex ones. Such simple words resulted from forms with inflectional schwa suffixes that lost their morphological function. Previously inflected forms like end+e or wild+e, for example, were reduced to simple end and wild. Second, in forms that remained complex, schwa loss reduced previously syllabic suffixes to mere consonants. In forms like seemed, or sins, for example, the suffixes {/əd}/ (past tense or past participle) and {/əs/} (plural, third person, or genitive) were reduced to {/d/} and {/z/}. Third, schwa loss greatly increased both the number of word forms that ended in consonant clusters and the diversity of these clusters. Finally, the new endings that schwa loss produced in simple words on the one hand, and complex ones on the other, were frequently homophonous, i.e. morphotactically ambiguous in the sense that is relevant for our purposes.

Most importantly, however, schwa loss was insensitive to its effects on morphological transparency. Being a system-wide phonological change with hardly any exceptions, it reduced the transparency of most inflections and contributed to their eventual disappearance. Thus, schwa loss is highly unlikely to have respected any bias against morphotactically ambiguous word forms either, and many of the word forms that it produced were indeed morphotactically ambiguous. Therefore, they represent a promising dataset for testing our hypothesis: if there exists a bias against morphotactic word form ambiguity at all, then the subsequent evolution of word forms created by schwa loss created ought to reveal its effects.

The specific forms we singled out for our study end in sonorants followed (a) by /d/ or /t/ (i.e. words such as fill+ed, mild, spil+t, hilt, seem+ed, dream+t, sinn+ed, wind, mean+t, tent), and (b) by /s/ or /z/ (i.e. words like sin+s, prince, wall+s, false, bar+s, farce, etc.). Both of them figure centrally in English inflectional morphology and have been strongly involved in morphotactic ambiguities. Furthermore, the final obstruents have been involved in voicings (the Modern English {/z/} suffixes all derive from {/əs/}), and devoicings (as in /bɜrnd/ from /bʊrnəd/ ‘burnt’, and others), so that there has been variation and competition between clusters ending in voiced obstruents and clusters ending in voiced ones.

To obtain our data, we used a diachronic database of word final consonants clusters in English, 4 which we had derived from the Penn-Helsinki Parsed Corpora of Middle and Early Modern English (Kroch and Taylor 2000; Kroch et al. 2004). From these corpora, we extracted all word forms that had the potential of ending in clusters at any time between the eleventh and the eighteenth century. That is to say, we gathered (a) word forms that ended in consonant clusters at the time of their attestation (e.g. EME ald /ɑld/ ‘old’, or EModE seemed), (b) word forms that had the potential of ending in clusters after schwa loss (e.g. EME sinnes /sinəs/>/sinz/ ‘sins’), and (c) word forms that had lost a cluster present in their predecessors (e.g. EModE climb /kli:m/</kli:mb/ after final consonant deletion).

We then entered the retrieved items into a database, lemmatized them, and analysed them with regard to a number of phonological, morphological, and syntactic criteria. For each word form that had the potential of ending in a cluster, we added a measure of the probability that it was actually realized with one at the time of its attestation. That measure represented an estimate derived from a number of factors, namely (a) spelling, (b) time of attestation, and (c) phonological context. For occurrences spelt like <spilt> or <wind>, for example, we assumed that the probability of cluster realization was 100%, while for forms spelt like <luved> or <ende> the probability we assigned reflected the probability that the vowels potentially represented by the <e>s were actually realized. That depended, in turn, on the progress of schwa loss and the phonological context. We assumed that the spread of schwa loss followed two s-curves, one for schwa loss in open final syllables (as in ende), and the other for schwa loss in checked ones (as in luved). We derived these s-curves on the basis of consensus views (see Minkova 1991) on the beginning and the completion of schwa loss 5 and calibrated them against evidence from verse texts from different periods. The other parameter we took into account was the beginning of following words: before words beginning with vowels schwa loss in open syllables was implemented earlier than before words starting with consonants.

Importantly, our database allows one to ignore probability estimates or not. If one takes them into account, one sees how many words that were actually realized with final clusters, given the progression of schwa loss at the time of their attestation. If one ignores them, one sees how many words would end in clusters if schwa loss had been completed. If one interprets data from a language stage in which schwa loss had only just started as if it had been fully completed, one constructs a ‘virtual post-schwa-loss stage’ in which no changes had occurred except schwa loss.

In our study of sonorant+/d/ or /t/ clusters, we exploited this possibility to generate baselines against which we could test the prediction that the morphotactic ambiguities brought about by schwa loss were reduced afterwards. In other words, we compared ‘virtual post-schwa loss English’ to ‘actual post-schwa loss English’, and predicted that the clusters should be less ambiguous in the latter.

In our second case study, which focusses on final clusters ending in /s/ or /z/, we applied a similar method. However, in that study our aim was not just to find out whether word forms became less ambiguous over time, but, more specifically, to what extent their disambiguation resulted from a particular change, namely the voicing of the originally voiceless {/əs/} suffix to ModE {/z/}. Therefore, we compared actual EModE data to a virtual data set on which we simulated the absence of {/s/} voicing. That is to say, we read actual Early Modern English corpus data, as if all plural, genitives, and 3-present forms were pronounced voiceless. Once again, or prediction was that ambiguity should be lower in the actual data set than in the virtual one.

3 Sonorant+/d|t/ clusters

3.1 Overview

During the Middle English period, sonorant+/d|t/ clusters became frequent in past tense forms after schwa loss had reduced the originally syllabic suffix {əd} to {d}. The morpheme was pronounced [d] or [t], but did not quite have the allophonic distribution that it has today. Particularly after sonorants both variants could occur. The voiceless ones reflected a rather sporadic sound change of final devoicing that affected mostly weak verbs and has left traces in modern irregulars such as spoilt, spelt, sent, or learnt (Weɫna 2009). Examples of Middle English complex word forms ending in sonorant+[t/d] clusters are

fill+ed ‘fill.pt’deal+t ‘deal.pt’
rear+ed ‘rear.pthar+t ‘hear.pt
joyn+ed ‘join.pptbren+t ‘burn.pt
assham+ed ‘ashame. pptidem+et ‘sentence.ppt
forr-lang+edd ‘long.ppt

At the same time, many simple words came to end in the same codas. They typically inherited them from their Old English ancestors, which had often taken syllabic inflections, so that their stem final clusters had surfaced as medial rather than final in many forms (as in [end]stemə]word). During the Middle English period most of these syllabic endings were lost, so that the stem final clusters surfaced as word final (as in [end]stem]word). Also in simple items final /d/ was sometimes devoiced. The change was rather sporadic, but forms like <ant> and, <chilt> child, or healent for heland ‘saviour’ are frequently attested. Examples of simple words in sonorant+(d/t) are:

fremd ‘foreign’excemt ‘exempt’

Some final sonorant-(d/t) clusters have remained ambiguous until today. Examples are /ld/ as in build or billed, /rd/ as in bird or reared, /nd/ as in band or banned, /lt/ as in kilt or spilt, and /nt/ as in hunt or meant. Others have lost their ambiguity. The cluster /md/, for example, occurs only in Modern English past tense forms like seemed, although it did also occur in simple forms such as fremd ‘strange, foreign’ or (lexicalised) samod ‘together’ in Early Middle English. The cluster /rt/, on the other hand, survives only in simple items, although in Middle English forms such as iheret or hart ‘heard’ were still attested.

3.2 Results

As explained, we carried out two searches. In one we searched for words potentially ending in clusters in texts from the period between 1150 and 1250 and analysed them as if schwa loss had been completed, i.e. as if they did actually end in clusters. These virtual post-schwa-loss words could obviously not reflect any effects of a bias against the ambiguities that schwa loss created. The other search was carried out on data from the period between 1640 and 1710, and returned words from actual post-schwa-loss English. We then compared the ambiguities displayed by the clusters in the two datasets, and checked if the one in the actual data was lower than the one in the virtual data, as we predicted.

Tables 1 and 2, and Figure 1 show the results we obtained when we counted word form tokens:

Table 1:

Virtual post-schwa-loss in SON{d/t}# clusters; token frequencies. Correlation: φ= 0.11; 95%-CI = (0.08,0.14).

Table 2:

Actual post-schwa-loss forms in SON{d/t}# clusters; token frequencies. Correlation: φ= 0.07; 95%-CI = (0.06,0.08).

Mosaic plots of Tables 1 and 2 and development of correlation coefficients in SON{d/t}# clusters (token frequencies). Error bars denote 95% confidence intervals.
Figure 1:

Mosaic plots of Tables 1 and 2 and development of correlation coefficients in SON{d/t}# clusters (token frequencies). Error bars denote 95% confidence intervals.

On the whole, the morphotactic ambiguity of the two clusters did not decrease but increased slightly, although not significantly. In fact, however, the developments do not show a clear direction. Reading the mosaic plots from top to bottom captures, roughly, the point-of-view of perception. From that perspective, both son+/t/ and son+/d/ clusters remained more indicative of simplicity, but while son+/t/ clusters came to indicate it more clearly (93% simple in actual data vs. 85% in virtual data), son+/d/ clusters actually became more ambiguous (87% simple in actual data vs. 93% in virtual data). From the point-of-view of production (read the mosaic plots from left to right), on the other hand, signalling complexity by /d/ became clearly more preferred (93% simple in actual data vs. 85% in virtual data). However, it remained equally preferred as a strategy for signalling simplicity (85% in both actual and virtual data). Thus, token counts provide no evidence for any decrease in the morphotactic ambiguity of sonorant+(d/t) clusters.

A rather different picture emerges on the type level, however. The results are presented in Tables 3 and 4, as well as in Figure 2.

Table 3:

Virtual post-schwa-loss forms in SON{d/t}# clusters; type frequencies. Correlation: φ= 0.2; 95%-CI = (0.12,0.22).

Table 4:

Actual post-schwa-loss in SON{d/t}# clusters; type frequencies. Correlation: φ= 0.53; 95%-CI = (0.49,0.57).

Mosaic plots of Tables 3 and 4 and development of correlation coefficients in SON{d/t}# clusters (type frequencies). Error bars denote 95% confidence intervals.
Figure 2:

Mosaic plots of Tables 3 and 4 and development of correlation coefficients in SON{d/t}# clusters (type frequencies). Error bars denote 95% confidence intervals.

Among word types, a highly significant decrease in word form ambiguity is observable. Sonorant+/d/ clusters became indicative of complexity while before they had been more common in simple types (59% complex in actual data vs. 36% in virtual data). Sonorant-/t/ clusters, on the other hand, which had already been associated with simplicity before, became much more indicative of it (96% simple in actual data vs. 75% in virtual data). From the point-of view of production, a similar development is observable: simple words, which had preferred final son+/d/ now came to prefer son+/t/ (56% of all items in actual data vs. 31% in virtual data). Complex ones came to prefer son+/d/ much more strongly (96% of all items in actual data vs. 79% in virtual data).

Thus, type-level data do seem to confirm our prediction: the development of words ending in sonorant+/t/ and sonorant+/d/ clusters went in the direction of a complementary distribution. The former came to be predominantly simple, and the latter predominantly complex, and the overall ambiguity of their distribution decreased significantly. 6

3.3 Discussion

Our results are ambivalent and raise two questions. The first is whether token level results or type level results are more relevant for our hypothesis. We think that there are two arguments in favour of the type level. One is that the token level results include the conjunction and, which is far more frequent than any other word. It accounts for 17,862 of 25,634 simple tokens (or about 70%) ending in sonorant+/d/. Being extremely frequent, the occurrence of and has always been highly predictable, it has hardly ever been stressed, and has been frequently targeted by phonetic reductions. So it may actually have been realised without a final obstruent in a large number of occurrences. In Present Day English, it is frequently realized simply as [ən] or even just [n]. Also, its high frequency makes it so expectable and learnable that it would have hardly ever been mistaken for a complex form. If one subtracts and from the token count, however, token level results get much more similar to type level results. They are represented in Tables 5 and 6, and in Figure 3:

Table 5:

Virtual post-schwa-loss forms in SON{d/t}# clusters; token frequencies without and. Correlation: φ = 0.00; 95%-CI = (0.00,0.05).

Table 6:

Actual post-schwa-loss forms in SON{d/t}# clusters; token frequencies without and. Correlation: φ= 0.28; 95%-CI = (0.26,0.29).

Mosaic plots of Tables 5 and 6 and development of correlation coefficients in SON{d/t}# clusters (token frequencies without and). Error bars denote 95% confidence intervals.
Figure 3:

Mosaic plots of Tables 5 and 6 and development of correlation coefficients in SON{d/t}# clusters (token frequencies without and). Error bars denote 95% confidence intervals.

Although not as clearly as type-level results, measurements disregarding and also show a significant decrease in ambiguity, even though it remains rather high.

The second argument in favour of the type level is that the signalling relations that we are dealing with are highly abstract. The signifiers /x-sonorant-(d/t)/ represent a whole set of different cluster types and their signifieds are the abstract concepts of morphological complexity vs. simplicity. Learning to associate two highly abstract categories requires high-level generalisations, and these require evidence that is diverse in terms of types rather than frequent in terms of tokens. This has been shown, for example, about the acquisition of regular morphological processes (see e.g. Bybee and Hopper 2001; Bybee 2007). What makes abstract patterns learnable is not the frequency of individual tokens that display a specific pattern but rather the number of different types that share it. Thus, type level diversity might indeed be more relevant to our concerns than token level frequency, and this incline us to take the overall results as providing at least tentative support for the hypothesis that word form ambiguity was indeed selected against and reduced in the evolution of English after schwa loss.

The other question that our comparison raises is what specific developments actually caused the observed changes in the distribution of son+/d/ and son+/t/clusters. The question is relevant, because our hypothesis implies that ambiguity was not reduced by accident, but that the changes that reduced it were favoured because they did. The idea is that a bias against ambiguity should be strong enough to facilitate the learning, use, and transmission of word form variants that keep it low or even reduce it, while impeding the learning, use, and transmission of those that would raise it. It should ‘select’ for the former and against the latter, in other words, but cannot have an active role in producing the variants.

We have tentatively interpreted our findings as evidence that the ambiguity of sonorant+/d|t/ forms produced by schwa loss was indeed reduced. The observed differences between the global ambiguity measures thus seem to suggest that word forms were indeed sensitive to a bias against morphotactic ambiguity. It is still relevant to know, however, what the specific changes were that amounted to the observable disambiguation. The more easily they can be motivated on different grounds, the less one needs a bias against ambiguity to explain them.

A crude look at our data suggests that the distributional changes of son+/t/ and son+/d/ clusters had no single most important cause. Instead, we could identify at least three different types of factors. The first was that voiceless variants of past tense and participle forms, such as dreint ‘drained’, (con-)joint, ordeynt ‘ordained’, constreynt ‘constrained’, brant ‘burned’, openet ‘opened’, drownt ‘drowned’, stenet ‘stoned’, hart, iheret ‘heard’, spilt, etc. became less frequent, and have been survived only in a few irregular (and lexicalised!) variants such as burnt, learnt, or spilt. Also, inflected forms such as wilt ‘you will’, or shalt ‘you shall’ disappeared more or less completely. – The second factor was that the present participle ending -Vnd, as in lykend ‘liking’, folowand ‘following’, lyuind ‘living’, coruinde ‘carving’, sendend ‘sending’, etc. was replaced by -ing. – And the third factor was that the diversity of words ending in simple son+/t/ clusters increased more strongly over time than that of words ending in simple son+/d/ clusters. Many of them were Latinate loans such as transcendent (first attestation in OED 1598), ancient (1498), comment (1450), strident (1656), etc., or words ending in -ment such as commandment (1250), instrument (1300), etc. Simple words in son+/d/, such as abounde (1325) blonde (1481), command (1325), depend (1413), horrend (1420), etc. were also imported, but their number was much smaller: the OED reports 4386 Latinate loans in son+/t/ against 593 ending in son+/d/.

As far as the disproportional increase in word forms ending in simple son+/t/ clusters is concerned, our study can merely acknowledge it. To what extent it was also motivated by the reduction of word form ambiguity that it contributed to, is difficult to assess and would require a comparison with the frequency of potential inputs from French and Latin. That word form ambiguity may have played a role can merely not be ruled out. The case is only slightly different for the disappearance of voiceless past tense suffixes in forms that once had them. It might have resulted quite simply from the generalisation of /d/ as the default past tense suffix, but the fact that son+/t/ became increasingly frequent at the end of simple morphemes may in turn have supported the generalisation of the {/d/} suffix. So, the evidence is also ambivalent. The strongest case in favour of our hypothesis can probably be made for the replacement of Vnd by -ing as the participle ending. Although the change is well studied (see Zehentner 2012, and the references therein), its motivations are not completely understood. In particular, the change of final /d/ to /g/ is difficult to account for on phonological grounds and remains puzzling. Thus, the reduction of morphotactic ambiguity (especially in forms such as christened, reckoned, beckoned, etc.) may indeed have contributed to bringing the change about and may have motivated speakers to favour –ing when it was still competing with –Vnd. All in all, however, the strongest argument for counting the changes as evidence of a bias against morphotactically ambiguous word forms is that their combined effect was indeed to reduce ambiguity.

As we shall see, this is different in the case of sonorant+(z/s) clusters, which are the focus of our second study.

4 Sonorant+/z|s/ clusters

4.1 Overview

In contrast to the developments that reduced the ambiguity of sonorant+/d|t/ clusters, changes in the morphotactic ambiguity of sonorant+/s|z/ clusters were mostly brought about by a single rather puzzling change. It involved the –(e)s suffix that has come to signal nominal plurals and genitives as well as verbal third person singular present forms in Modern English. In Old English, the suffix was syllabic, and fricatives had two complementarily distributed allophones. In word final position they were realised without voicing, so that the suffix was –V[s]. In Late Old English the vowel was reduced to schwa, and during the Middle English period it got lost except after sibilants. Parallel to that, two developments took place that amounted to the lexical phonological representations and the phonetic variants of the pl.gen.3sg suffixes in Modern English. On the one hand, voicing contrasts became phonologically contrastive among fricatives, and on the other hand, the voiceless [s] allophone of the inherited fricative came to be re-interpreted, in the suffixes, as voiced rather than voiceless. Thus, their lexical representations have become /z/, which surfaces in all contexts that allow it, i.e. after vowels, sonorants and voiced obstruents – as in bees, sins, and dogs. After voiceless obstruents – as in cats or laughs –, it is devoiced reflecting a constraint against obstruent sequences that differ in voicing. After sibilants it is still mostly expressed as [ɪz]. Devoiced traces of the suffix survive in adverbs such as once, since, hence, etc. These stem from derivations with adverbial -Vs suffixes (i.e. on-es, sin-es, hen-es, see OED s.v), but have become fully lexicalised.

The eventual (re-)interpretation of the inherited [s] allophone in the suffix as /z/ (rather than the expected /s/) has traditionally been described as a sporadic sound change by which voiceless fricatives were voiced in weakly stressed items. Apart from the suffix, the change is supposed to account for was from earlier /wæs/, these /ðiːz/ and those /ðəʊz/ from earlier /θeːs/ and /θaːs/, etc. (see e. g. Luick 1914–21; Lass 1992; Pinsker 1974; Ringe 2003).

Extant accounts offer little in terms of explanation. Indeed, the change is phonologically unexpected: voiced obstruents count as highly dispreferred in final position, and even in Present day English word final /z/ phonemes are often expressed with reduced voicing and might better be described as voiceless lenis. So maybe the change that produced the Modern English {/z/} allomorph should not be thought of as a voicing at all, but rather as a lenition? This would indeed appear to make it less puzzling, since lenition, i.e. weakening, is not unnatural in weakly stressed positions. Yet, phonologically, there is little doubt that they do behave like voiced fricatives and unlike their voiceless counterparts. For instance, they have the expected effect of lengthening the duration of preceding vowels, and they cannot appear in clusters next to voiceless plosives (so we have /dɔgz/ ‘dogs’. but /kæts/ ‘cats’). Also, they can be pronounced with measurable voicing if they come to stand between two vowels, as in Bees are insects, while voiceless /s/ phonemes – as in Grass is green – cannot. Therefore, the final sibilants that surface in ModE these, sins, or graze cannot be expressions of phonologically voiceless /s/, as the ones in ModE house, since, or grass, which reflect the same Old English sibilant phoneme. Thus – and even though this might be not quite accurate considering their usual phonetic realizations –, we shall keep referring to final /z/ phonemes as phonologically ‘voiced’, and regard the question why they came to emerge in the plural morpheme {/z/} as still open.

In morphologically simple items, sonorant+/z/ endings have remained extremely rare up to the present, since final /s/ never got voiced in any of them except sporadically. Apart from frequent proper names, such as Thames, James or Evans, they include items like cleanse, summons (which has become only simple by lexicalisation) or alms (which goes back to the singular form almesse, but may be analysed as a plural by many native speakers).

In order to test the effects of final {/z/} voicing on the overall ambiguity of sonorant+(s/z) clusters, we proceeded basically in the same way as with sonorant+(d/t) clusters, except that we created our virtual baseline language not by simulating schwa loss on pre-schwa loss data, but by taking Early Modern Data (from periods E2 and E3 of the PPCEME, ranging from 1570 to 1720) and simulating the absence of suffix voicing on them. That is to say, we interpreted forms like sins, feels, and stars, as if they were pronounced /sins/, /feːls/, and /stɑrs/ rather than /sinz/, /feːlz/, and /stɑrs/. Then we checked for differences in word form ambiguity in the same way as before.

4.2 Results

As Tables 7, 8, 9, and 10 as well as Figures 4 and 5 show, the effects of /z/-voicing on ambiguity reduction were strong both on the token level and on the type level. Without it, the vast majority of both simple and complex forms would have come to be realised as voiceless. While this would have created few problems in production, it would have made perceptual processing much more difficult. Although the majority of word forms ending in sonorant+/s/ would have been complex, attempts to decompose them would have been unwarranted in more than a quarter of all possible cases, and /z/-voicing changed that radically. Interpreting sonorant+/z/ as a signal of complexity, now yielded only a negligible number of false positives, and most sonorant+/s/ forms could be safely interpreted as simple and processed in the lexicon. In fact, this might even apply to the 506 /ns/-tokens in actual Early Modern English that we analysed as morphotactically complex. They include items such as once, since, hence, etc., which had probably lost their transparency in Middle English already and might have been analysed as lexically simple as well. We decided to analyse them as complex because we did not want our analytic decisions to work in favour of our hypothesis.

Table 7:

Virtual post-schwa-loss in SON{z/s}# clusters; token frequencies. Correlation: φ= 0.06; 95%-CI = (0.04,0.08).

Table 8:

Actual post-schwa-loss in SON{z/s}# clusters; token frequencies. Correlation: φ = 0.88; 95%-CI = (0.87,0.89).

Table 9:

Virtual post-schwa-loss in SON{z/s}# clusters; type frequencies. Correlation: φ = 0.01; 95%-CI = (0.06,0.11).

Table 10:

Actual post-schwa-loss in SON{z/s}# clusters; type frequencies. Correlation: φ = 0.97; 95%-CI = (0.97,0.99).

Mosaic plots of Tables 7 and 8 and development of correlation coefficients in SON{z/s}# clusters (type frequencies). Error bars denote 95% confidence intervals.
Figure 4:

Mosaic plots of Tables 7 and 8 and development of correlation coefficients in SON{z/s}# clusters (type frequencies). Error bars denote 95% confidence intervals.

Mosaic plots of Tables 7 and 8 and development of correlation coefficients in SON{z/s}# clusters (type frequencies). Error bars denote 95% confidence intervals.
Figure 5:

Mosaic plots of Tables 7 and 8 and development of correlation coefficients in SON{z/s}# clusters (type frequencies). Error bars denote 95% confidence intervals.

4.3 Discussion

In contrast to sonorant+/d|t/ clusters, the sonorant+/s|z/ data provide a much clearer picture: the voicing of /s/ in inflectional suffixes led to a spectacular disambiguation. At the same time, there does not seem to be a plausible independent motivation of the voicing. The sequences it created were not easier to pronounce but more difficult to distinguish than the resident types they came to oust. Thus, the only purpose that the voicing clearly serves is the disambiguation of complex and simple word forms. Therefore, we think there is no better explanation of the change than assuming that it actually occurred for that purpose. We therefore take the case of sonorant+/s|z/ as strong evidence of a preference for word forms to indicate their morphological structure, or for a cognitive bias against ambiguous ones.

That bias seems to have positively selected for unambiguous word-form variants, for whatever reasons they emerged. In the case of /s/-voicing, for example, they may have emerged indirectly. The purely consonantal {/s/}-suffix that resulted from {/əs/} through schwa loss was probably first voiced when it occurred after voiced obstruents as in egg+s, field+s, or rib+s, because the constraint for obstruent sequences to agree in voicing has long ranked high in English (see e.g. Pinker 1999). Confronted with phonetic surface variation between [z] and [s], learners had to decide which of the two reflected the lexically underlying target. Both variants may have been tried out, generating variation among {/z/} and {/s/} variants of the suffix. Since the {/z/} variant then had the positive side-effect of indicating complexity unambiguously, it came to be selected and stabilized, while the inherited {/s/} variant disappeared. Thus, the cognitive bias in favour of morphotactically indicative word shapes exerted its selective effects just when historical contingencies provided the variation on which it could act. This is exactly what one would expect of a cognitive constraint in the cultural evolution of language constituents.

5 Conclusion

As we have seen, our two studies have produced different kinds of evidence regarding the hypothesis they were meant to test. While the sonorant+/s|z/ data seem to support it strongly, the data on sonorant+/d|t/ clusters appear to do so to a lesser degree. They do not contradict it either, however, but appear to point in the same direction. In their combination, our studies therefore suggest that a preference for word form shapes that signal morphological structure may indeed represent a constant pressure in the evolution of word form shapes, and may be causally involved in developments that seem to have little to do with it at first sight. Since the preference for morphologically indicative word forms is weak, its effects are difficult to detect except in specific phases in the histories of specific languages. Under appropriate circumstances, however, they can be detected, at least statistically. At the same time, the fact that they unfold only under specific circumstances, i.e. when the opportunity arises, qualifies them as potential explanations of developments that occur only in specific languages and during specific phases. The voicing of English /s/-suffixes might not have occurred, for example, had English not lost its final schwa sounds, and had it not happened to contain many simple items ending in sonorant+/s/. Likewise, the replacement of -Vnd by -ing as the present participle suffix might not have occurred, had the regular past tense suffix not produced an increasing number of complex verb forms ending in Vn+d.

We think that it would be interesting to see if similar results can be obtained if one studies other developments from the perspective we have sketched. In fact, some of the data reported in other papers in this issue appear to lend themselves for the purpose. Take Minkova’s account of word-final /n/ in the history of English, for example: in Old English, it occurred most frequently in inflectional suffixes, and marked word forms as complex. In Modern English, however, its distribution has come to be reversed. Word final /n/ is still very frequent, but occurs mostly within stems, and signals simplicity. Thus, its loss in inflectional endings may have in fact increased its stability at the end of simple forms. Another potential case is in Adamczyk and Versloot’s contribution to this issue. They show that monosyllabic root nouns such as OE stud ‘post’ or gōs ‘goose’ were significantly more likely to acquire syllabic genitive suffixes by analogical transfer than disyllabic nouns such as fæder ‘father’. This would have increased the proportion of polysyllabic word forms among morphologically complex ones, and would have made it easier to infer their morphological structure from their syllabic shape. These are just two examples, of course. They suggest, however, that further potential effects of a preference for morphotactically indicative word forms should be detectable if one looks for them.


This paper was funded by the Austrian Science Fund (FWF, Grant No. P27592–G18). For valuable feedback on previous versions, we owe gratitude to the participants on a 2017 workshop on diachronic phonotactics, many of whom are also contributors to this issue. Remaining errors are of course ours.


  • Anderson, Roy & Robert May. 1991. Infectious diseases of humans. Dynamics and control. New York: Oxford University Press. Google Scholar

  • Bybee, Joan. 2007. Frequency of use and the organization of language. Oxford: Oxford University Press. Google Scholar

  • Bybee, Joan L. & Paul J. Hopper. 2001. Frequency and the emergence of linguistic structure. Amsterdam: Benjamins. Google Scholar

  • Cumming, Geoff. 2014. The new statistics: Why and how. Psychological science 25(1). 7–29.Google Scholar

  • Dressler, Wolfgang U. 1985. Morphonology: The dynamics of derivation. Ann Arbor: Karoma. Google Scholar

  • Dressler, Wolfgang U. 2003. Naturalness and morphological change. In Brian D. Joseph & Richard D. Janda (eds.), The handbook of historical linguistics (Blackwell handbooks in linguistics), 461–471. Malden, MA: Blackwell Pub. Google Scholar

  • Dressler, Wolfgang U. & Katarzyna Dziubalska-Kołaczyk. 2006. Proposing morphonotactics. Wiener Linguistische Gazette 73. 69–87. Google Scholar

  • Dressler, Wolfgang U., Katarzyna Dziubalska-Kołaczyk & Lina Pestal. 2010. Change and variation in morphonotactics. Folia Linguistica Historica 31. 51–68. Web of ScienceGoogle Scholar

  • Fehér, Olga, Elizabeth Wonnacott & Kenny Smith. 2016. Structural priming in artificial languages and the regularisation of unpredictable variation. Journal of Memory and Language 91. 158–180. CrossrefWeb of ScienceGoogle Scholar

  • Giegerich, Heinz J. 1999. Lexical strata in English: Morphological causes, phonological effects (Cambridge studies in linguistics 89). Cambridge: Cambridge University Press. Google Scholar

  • Kaye, Jonathan. 1995. Derivations and interfaces. In Jacques Durand & Francis Katamba (eds.), Frontiers of phonology, 289–332. London & New York: Longman. Google Scholar

  • Kiparsky, Paul. 1982a. From cyclic phonology to lexical phonology. In Harry Van Der Hulst & Norval Smith (eds.), The structure of phonological representations, 131–175. Dordrecht: Foris Publications. Google Scholar

  • Kiparsky, Paul. 1982b. Lexical phonology and morphology. In Ik-Hwan Lee (ed.), Linguistics in the morning calm, 1st edn, 3–91. Seoul: Hanshin Publ. Google Scholar

  • Kirby, Simon, Thomas L. Griffiths & Kenny Smith. 2014. Iterated learning and the evolution of language. Current Opinion in Neurobiology 28. 108–114. Web of ScienceCrossrefGoogle Scholar

  • Kroch, Anthony. 1989. Reflexes of grammar in patterns of language change. Language Variation and Change 1. 199–244. CrossrefGoogle Scholar

  • Kroch, Anthony, Beatrice Santorini & Lauren Delfs. 2004. Penn-Helsinki Parsed corpus of early modern English. http://www.ling.upenn.edu/hist-corpora/

  • Kroch, Anthony & Ann Taylor. 2000. Penn-Helsinki Parsed corpus of middle English. http://www.ling.upenn.edu/hist-corpora/

  • Lass, Roger. 1992. Phonology and morphology. In Norman Blake (ed.), Cambridge history of the English language, vol. 2: 10661476, 23–155. Cambridge: Cambridge University Press. Google Scholar

  • Luick, Karl. 1914–21. Historische Grammatik der englischen Sprache. Leipzig: Tauchnitz. Google Scholar

  • Mayerthaler, Willi. 1981. Morphologische Natürlichkeit (Linguistische Forschungen 28). Wiesbaden: Akademische Verlagsgesellschaft Atnenaion. Google Scholar

  • McCarthy, John J. & Alan Prince. 2004. Generalized alignment: Introduction and theory. In John J. McCarthy (ed.), Optimality theory in phonology, 72–76. Oxford: Blackwell. Google Scholar

  • McMahon, April M. 2009. Lexical phonology and the history of English (Cambridge studies in linguistics 91). Cambridge: Cambridge University Press. Google Scholar

  • Minkova, Donka. 1991. The history of final vowels in English: The sound of muting (Topics in English linguistics 4). Berlin, New York: M. de Gruyter. Google Scholar

  • Minkova, Donka. 2009. Phonological weakness in English: From old to present-day English (Palgrave studies in language history and language change). Basingstoke, England, New York: Palgrave Macmillan. Google Scholar

  • Pinker, Steven. 1999. Words and rules: The ingredients of language. 1st edn. New York: Basic Books. Google Scholar

  • Pinsker, Hans E. 1974. Historische Englische Grammatik. München: Hueber. Google Scholar

  • Post, Brechtje, William D. Marslen-Wilson, Billi Randall & Lorraine K. Tyler. 2008. The processing of English regular inflections: Phonological cues to morphological structure. Cognition 109(1). 1–17. Web of ScienceCrossrefGoogle Scholar

  • Reali, Florencia & Thomas L. Griffiths. 2009. The evolution of frequency distributions: Relating regularization to inductive biases through iterated learning. Cognition 111(3). 317–328. CrossrefWeb of ScienceGoogle Scholar

  • Ringe, Don. 2003. Internal reconstruction. In Brian D. Joseph & Richard D. Janda (eds.), The handbook of historical linguistics (Blackwell handbooks in linguistics), 244–261. Malden, MA: Blackwell Pub. Google Scholar

  • Smith, Kenny, Amy Perfors, Olga Fehér, Anna Samara, Kate Swoboda & Elizabeth Wonnacott. 2017. Language learning, language use and the evolution of linguistic variation. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 372. 1711. Web of ScienceGoogle Scholar

  • Solé, Ricard. 2011. Phase transitions. Princeton: Princeton University Press. Google Scholar

  • Weɫna, Jerzy. 2009. The post-sonorant devocing of [d] in the past/past participle forms of weak verbs (sent, spend, etc.). In Marcin Krygier & Liliana Sikorska (eds.), Þe laurer of oure Englische tonge, 21–34. Frankfurt: Peter Lang. Google Scholar

  • Wurzel, Wolfgang U. 2001. Flexionsmorphologie und Natürlichkeit (Studia grammatica v. 21). Berlin: Akademie Verlag. Google Scholar

  • Zehentner, Eva. 2012. -AND vs. -ING. Vienna: University of Vienna Master’s Thesis. Google Scholar


  • 1

    Already Neogrammarian research showed that the transparency of complex word forms comes to be (re-)established, or increased, by processes of analogical transfer. In naturalist approaches to phonology-morphology interaction, preferences for complex word forms to be morphologically transparent also play a central role (e.g. Dressler (1985, 2003), Wurzel (2001)); in Lexical Phonology, differences between the phonotactics of simple and complex items represent a core area of investigation (e.g. Kiparsky (1982a), Kiparsky (1982b), McMahon (2009), Giegerich (1999)); and Optimality Theory posits constraints by which prosodic boundaries and morphosyntactic boundaries come to be aligned (e.g. McCarthy and Prince (2004)), so that the compositionality of complex forms is reflected in their phonotactic structure (see also Kaye 1995). 

  • 2

    In the sentence We need to act, for example, it is clear that act can neither be a past tense nor a participle, and in A rose is a rose is a rose, rose must be a singular noun, and not the third person singular of row. 

  • 3

    The histories of all languages are long enough for them to be sufficiently well adapted to universal cognitive constraints. Modern languages are not any better than their predecessors. 

  • 4

    The creation of that database was funded by the Austrian Science Fund. It is accessible online under <ecce.acdh.oeaw.ac.at>. 

  • 5

    Technically, we derived the s-curve by means of two interacting logistic-spread models (Solé 2011; Anderson and May 1991; Kroch 1989) 

  • 6

    At the same time, type level data also corroborate the more general prediction that cluster types which are phonologically less preferred should be more likely to become signals of complexity (see Dziubalska-Koɫaczyk (this issue)). In phonological terms final sonorant-/d/ is less preferred than sonorant-/t/ – both because final obstruents are generally preferred to be voiceless, and because they contrast more strongly with the preceding sonorants. 

About the article

Received: 2018-06-15

Revised: 2018-11-30

Accepted: 2019-01-23

Published Online: 2019-07-28

Published in Print: 2019-07-26

Citation Information: Folia Linguistica, Volume 40, Issue 1, Pages 129–151, ISSN (Online) 1614-7308, ISSN (Print) 0165-4004, DOI: https://doi.org/10.1515/flih-2019-0007.

Export Citation

© 2019 Walter de Gruyter GmbH, Berlin/Boston.Get Permission

Comments (0)

Please log in or register to comment.
Log in