Baseless derivation: the behavioural reality of derivational paradigms

2.1 The rooted tree view of morphology

The rooted tree view approaches derivational morphology with the assumption that every lexeme is either underived or relates to a single other lexeme, its base, that is both conceptually prior and formally simpler. As (Stump 2019) highlights, this leads to the view that derivational families are structured as rooted trees. As illustrated in Figure 1 for the family of derive, in such a view, relationships between lexemes are strictly hierarchical (some lexemes are higher up than others in the tree structure), and asymmetric (derived lexemes depend on their formal base and nothing else in the family). This view of derivational morphology was first formalised by Aronoff (1976), but is explicitly or implicitly assumed by the overwhelming majority of work in word formation, or work that uses word formation as a tool to answer other questions about language. Within Cognitive Linguistics, the two main theories staking claims about the structure of the derivational lexicon, Relational Morphology and Construction Morphology, assume that the rooted tree view is the default situation in derivational relationships and accounts for the majority of morphological relationships.

Figure 1:

A rooted tree representation of part of the derivational family of derive. The lexemes shown here are a subset of those documented in the Oxford English Dictionary.

Despite its popularity, the pure rooted tree model of word formation has a number of theoretical and empirical shortcomings, the most salient of which are outlined below.

There are numerous cases where a derived word has multiple potential bases. Rederivation is one such case, it could be the result of adding an agentive -ion suffix to rederive, or of adding a reiterative prefix re- to derivation. Both routes make equally accurate predictions about the form and meaning of the derived word. Ideally, one might like to posit both rederive and derivation as joint formal bases of the derived word,^[2] however this is not permitted by the rooted tree architecture. Decisions about which lexeme should be chosen as formal base in these situations are often somewhat arbitrary, and the constraint to only allow a single incoming edge per word cannot capture the multiple sources of support that the coinage is likely to have received.

Even more flagrantly, there are cases where a derived lexeme’s properties have clearly been taken from different lexemes in its derivational family (Hathout and Namer 2014b): take the triplet language, linguist, linguistic. Formally, linguistic is derived from linguist (not from language, cf. *languagic). Consider other denominal adjectifs in X(t)ic and their formal base: gene ∼ genetic, acid ∼ acidic, artist ∼ artistic show that the pattern is for the derived word in X(t)ic to mean ‘related to X’. Linguistic theory is not a theory about linguists, but about language. So while the word’s form is derived from linguist, its meaning is derived from language. The rooted tree’s property of allowing only one incoming edge per node does not allow for a representation of this situation.

A different but related issue stems from the monodirectional nature of the relations. There is extensive historical evidence for many instances of derivation by backformation.^[3] For instance, many English verbs have been backformed from latinate nouns, such as resurrect ∼ resurrection, project ∼ projection and insert ∼ insertion. Aside from the later attestation date compared to the noun, what gives away the backformed origin of the verb is its phonological form. These verb  ∼  noun pairs both share the Latin supine stem, identifiable by the final -t. Non-backformed latinate nouns (satisfy ∼ satisfaction, evoke ∼ evocation) instead use allomorphic stems, reflecting the different stems associated with verbs and deverbal nouns in Latin morphology. This shows that the first set of verbs was formed based on the noun, rather than the other way around.

The monodirectional nature of derivational trees leads to a further set of complications for situations where establishing directionality between two items is unmotivated or impossible – this is known as cross-formation (Becker 1993). A prime example are conversion pairs: march _V and march _N do not feature affixes that signal which one might have come from the other, and their lexical semantics are in a relationship of mutual predictibility. It would be desirable to note the existence of a bidirectional, non-hierarchical^[4] relationship between the two. Some phenomena are simultaneously troublesome for the requirements of directionality and hierarchy: English has many pairs of derived words with the pattern Xism ∼ Xist such as sexism ∼ sexist from the formal base sex, buddhism ∼ buddhist from the formal base Buddha. The two derived words are much more closely linked to each other than they are to their base in both form and meaning: buddhism is a set of beliefs and behaviours historically based on the teachings of the Buddha, but the concept denoted by the derived word is now rather far from that denoted by its base, encompassing specific beliefs and attitudes as well as having associated symbolism and behaviours that the Buddha is not responsible for and might not even endorse or be familiar with (e.g. the Gandharan Buddhism of the Kush empire, a synthesis of buddhist teachings and Greco-Roman, Iranian and Indian elements). A buddhist is someone who espouses buddhism, and buddhism is the total of ideas and practices espoused by buddhists: not only are the two concepts much closer than they are to their base, but the two also share the stem buddh-, further entrenching the contrast between the pair and their base (*buddhaism ∼ *buddhaist, where the absence of the vowel is not due to phonotactic constraints: the string aist/aism exists in dadaism ∼ dadaist). A particularly striking example of the same problem comes from cases of cross-formation where there is no base: optimist ∼ optimism closely resemble each other in form and meaning, and lack a base (*optime) that would enable them to fit into a rooted tree structure.

2.2 Paradigmatic links in derivational morphology

These phenomena put into question the usefulness and accuracy of the rooted tree view. However, their well-known and extensively documented existence has not been met with a desire to reshape morphological theories of word formation: a large part of the morphological community sees them as marginal phenomena that should not drastically change the way we conceive of word formation.

The seminal study of van Marle (1984) was influential in suggesting that these phenomena should not be ignored by theorists, though he still claimed that they were fundamentally different in quality from garden-variety derivation, encoding this in the terminology he chooses to discuss paradigmatic relationships: they give rise to secondary or analogical coinings. This view is currently held by Construction Morphology (Booij 2010), which makes a distinction between first-order and second-order relationships (classical base-centric monodirectional derivation versus paradigmatic relationships) and Relational Morphology (Jackendoff and Audring 2020) which makes the same distinction under the name of daughter versus sister schemas, reflecting a view of the world that is akin to the rooted tree with extra links between leaf nodes. While this tradition essentially accepts the rooted tree view, it allows for mechanisms of connectivity between leaves in order to account for the problematic phenomena outlined above.

Nevertheless, a subset of morphologists has argued for decades that the existence of said problematic phenomena motivates a reconceptualization of derivational morphology as the general study of relatedness between lexemes (Bauer 1997; Becker 1993; Bochner 1993; Bonami and Strnadová 2019; Hathout and Namer 2022; Robins 1959; Štekauer 2014). This amounts to extending to derivation the central tenets of the Word and Paradigm (WP) approach to inflectional morphology (Anderson 1992; Blevins 2016; Hockett 1954; Matthews 1972; Robins 1959; Stump 2001). Matthews (1991, p. 216) cogently summarises these as holding that words are “not wholes composed of simple parts, but are themselves the parts within a complex whole,” namely a paradigm. The paradigmatic approach focuses on systematic relationships between words, and holds these relationships to be bidirectional and non-hierarchical, making them suitable for formalising previously problematic phenomena.

In this paper, we adopt the formalisation of the derivational paradigm proposed by Bonami and Strnadová (2019). Their process for inducing paradigmatic structure is rooted in patterns of language use: they begin by finding pairs of words that are morphologically related, that is, have a relationship of both form and meaning. Then relationships are filtered for systematicity: only recurrent morphological relationships may form the basis of a paradigm. Groups of words that are morphologically related form a morphological family. Multiple morphological families are then aligned on the basis of the semantic contrasts they instantiate (see Bybee (2001) for evidence on the primacy of meaning in lexical relatedness and Štekauer (2014) on its role in paradigmatic organization), leading to a multidimensional paradidmatic structure. The output of this process is illustrated in Figure 2.

Figure 2:

A visualisation of three partial derivational paradigms of English, presented horizontally, with parallelism across families presented vertically.

This conception of derivational paradigms is an instantiation of a broader trend of seeing morphology specifically, but also language in general, as a complex adaptive system (Blevins et al. 2016). The idea that language can be fruitfully modelled as a complex adaptive system or as a dense dynamic network, a rich structure in which constructions enter relationships of multiple kinds with each other based on patterns of usage, is a position that converges with that generally held by cognitive linguistics when applied to domains other than morphology. To give a few examples, Diessel (2019) outlines a general program for a nested model of grammar. In a chapter that offers an overview of concepts and methodologies for usage-based linguistics Gries and Ellis (2015) highlight that much recent work at the intersection of cognition and language is based on associative learning and discuss tools for approaching corpus studies through this lens. Ellis et al. (2016)’s book discusses this in the context of language acquisition and how humans develop flexible representations of their language over time by exploiting relationships of usage between constructions. Sommerer and Smirnova (2020) discuss the same dynamic network approach for modelling diachronic language change.

In a morphological system with these characteristics, the lexicon is a densely connected graph organised along two dimensions: the morphological family and the paradigmatic cell. Within a single morphological family, each lexeme is in a bidirectional relationship with every other member, acknowledging that influences of form and meaning operate in a much more holistic fashion.

Section 2.1 discussed the phenomena that motivated a change of framework for word formation. The next question is whether this shift accurately captures essential aspects of lexical derivation that were disregarded by a rooted tree methodology. A key difference in prediction between the two concerns the way in which information flows. In a rooted tree model, information may only flow from the root outwards, never in the reverse direction, and never between nodes which are not in a base → derivative relationship. In a fully paradigmatic model, information may flow between any two connected nodes, in any direction.

Much recent literature on inflectional morphology focuses on the implicative structure of paradigms (Ackerman and Malouf 2013; Ackerman et al. 2009; Stump and Finkel 2013; Wurzel 1989). Paradigmatic relationships are implicative in nature since they constrain possible relations of form and meaning between paradigmatically related words in a way that can be captured in terms of (probabilistic) conditional statements. Using English inflection as an example, if we know the past form of a verb has the shape Xed, the we can be certain that its present form has the shape X. On the other hand, if we know the present of a verb to be X, then it is highly probable that the past has the shape Xed, but it is not certain. Under a non-paradigmatic view, it is surprising that the past tense is a better predictor of the present (thought to be the unmarked/neutral/base form of the English verbal system) than the present is of the past tense – information should flow from one cell only (the formal base) towards the rest. Under a paradigmatic view, this is expected: a form in any cell is supposed to provide at least some information about the form in any other cell (Ackerman and Malouf 2013) and combining information from more cells will improve prediction towards other cells in the system (Bonami and Beniamine 2016).

Bonami and Strnadová (2019) extend this methodology to derivational morphology. Table 1 documents predictability relations between French verbs and corresponding action nouns and agent nouns; importantly, the verb is the formal base of the triplet in the overwhelming majority of cases in French. The table shows that the form of the action noun is just as easy to predict from the verb as it is from the agent noun, and the action noun is a better predictor of the form of the agent noun than the verb is. This is unexpected under a rooted tree model: the verb should be the best predictor of the other two forms, and it should be hard to predict the verb. The authors additionally show that using multiple predictors is always an improvement over using a single one: this means that the action and agent noun have information about forms of other members of the paradigm that are not available based on the verb alone. It is noteworthy that the easiest prediction in this dataset involved predicting verb from the agent and action nouns – under a rooted tree model, this would be expected to be the hardest prediction. This work shows that, from an information-theoretical perspective, the word forms of derivational families are characterised by implicative relationships of predictability form each form to any other, as would be expected under a paradigmatic perspective.

2.3 Evaluating the claim for monodirectedness in derivational families

Many theories of morphology assume directionality in the relationships that link together the lexicon. Theories that fall under the categories Hockett (1954) terms item and arrangement (IA) and item and process (IP) require a root to which material is added or to which a process is applied, making all relationships in this framework inherently directional. Word-based approaches in the tradition of Aronoff (1976) substitute a concrete word for the abstract root, but do so while keeping the directionality assumptions inherent in IA and IP. The recent increase of popularity for word and paradigm theories has afforded morphologists the possibility of questioning the concept of the base, and of the directionality of relationships.

The theoretical literature has begun to grapple with this matter in both inflection and derivation, as discussed in the previous few sections. However, evidence bearing on this matter is often reliant on specific data points and their relationships of form and meaning, rather than on more direct evidence about how humans store and process language. Within morphology, behaviourally-oriented work on the status of the base in inflection is almost equally scarce as work in derivation. A notable exception is to be found in the work of Adam Albright. Albright (2002) proposes explicitly that the base of a paradigm has special primitive status (the Single Base Hypothesis), and that this should be borne out by behavioural studies. Jun and Albright (2017) then tested this prediction explicitly by asking speakers of Korean to judge the acceptability of the base form of a pseudolexeme, using a non-base as conditional information. They calculate two sets of predictability scores between the base and non-base form using the Minimal Generalisation Learner (Albright 2002; Albright and Hayes 2003) (MGL). The MGL is a method for assigning likelihoods to form relationships between two cells, outputing a score for form F1 of word family WF in cell C1, conditional on knowing that WF has form F2 in cell C2. The method relies on the type frequency of morphological patterns, which is derived in a way that takes into account the phonology of the word forms. Jun and Albright (2017) attempt to predict participants’ behavioural responses to non-base ∼ base pairs of words in Korean with forward scores (predicting the non-base form from the base) and backward scores (predicting the base form from the non-base), named after whether the direction of prediction in the experimental task matched the input-output pair presented to the MGL. They find that backward predictability scores explained more variability in participant judgements than forward scores: despite participants predicting the base form from a non-base form, scores that assumed participants were using the base form as a starting point provided a better fit for the data, taking this as evidence in favour of the Single Base Hypothesis. However, evidence showing that speakers are aware of paradigmatic implicative relationships between inflected word forms also exists: in a behavioural study with the goal of investigating speakers’ awareness of and reliance upon paradigmatic relationships, Copot and Bonami (2024) perform an experiment on the French verbal system that parallels Jun and Albright’s in methodology, and find that the cell deemed to be the base appears to have no inherently special status in the system, once the frequency of forms within the cell is accounted for.

There is no parallel explicit claim for the processing of derivational relationships, but the assumption can still be found in countless publications that assume the formally simpler of two morphologically related words must somehow be more basic, and thus in a privileged position with regards to informativity. Consider for example the literature on whether morphologically complex words are parsed as wholes or segmented into morphemes. In addition to being historically IA/IP-informed, this literature assumes that suitable evidence for decomposition is yielded by showing priming for a morphologically complex word from its formal base. Prominent examples are Levelt et al. (1999), Rastle et al. (2004) and Christiansen and Chater (2016), but most work in this tradition shows the same implicit belief that there is such a thing as the base of a derived word.

The cognitive linguistics literature is also aware of directionality as a parameter: the move towards associative learning has been one away from directional relationships, as discussed explicitly in e.g. Gries (2013), and towards symmetrical ones, defined by patterns of occurrence and co-occurrence. Despite this, theories of morphology aligned with cognitive linguistics have largely borrowed the assumption that relationships between constructions are directional, unless otherwise specified.

In this paper, we present an experiment that addresses the status of the base within derivational paradigms for language speakers.

2.4 Research questions

This paper wishes to test some of the main diverging assumptions made about the structure of the derivational lexicon through a behavioural experiment. In the background, we outlined three main views: the rooted tree view (the lexicon is a series of star graphs made up of monodirected relationships), the paradigmatic view (the lexicon is a densely connected network of bidirectional associations) and the cognitive morphology view (the rooted tree is the default situation, though some leaf nodes might be connected with each other). These three sample positions make different claims about how humans store and process derivationally related lexemes. All three approaches expect speakers to be aware and keep track of repeated patterns in the lexicon and their likelihood. However, they differ in which relationships they suggest speakers are keeping track of.

We test this with an acceptability judgement task asking speakers to judge a novel word form conditional on knowledge of another member of its paradigm, inspired by the methodology of Jun and Albright (2017) for inflectional morphology. Jun and Albright observed that in inflection, if the form of the word to be judged was very predictable conditional on the form of another member of its paradigm, the word would be rated higher than if it was unpredictable. The first claim we wish to test is whether form predictability between derivational forms correlates with speaker behaviour, in the same way it does for inflectionally related forms. A positive correlation between the predictability of this form from its paradigm mate (i.e. a word belonging to the same paradigm) and speaker judgement of the form would constitute evidence that speakers are aware of and exploit the implicative structure between derivational forms.

Paradigmatic approaches make two predictions that diverge from those of the rooted tree view. Firstly, they predict that speakers will show awareness of implicative relations even when making predictions between two non-base forms. In this setting, a rooted tree approach would predict that speakers are making the prediction in multiple steps, which must necessarily pass through the prediction of a base form. If speaker judgement correlates positively with form predictability even when two non-base forms are involved, this constitutes evidence for the paradigmatic view. Secondly, rooted tree approaches claim that all predictions are made from the base. Jun and Albright (2017) make this claim from a cognitive perspective for inflection explicitly. A paradigmatic approach does not give the base any inherently special status. Therefore the different views make differing hypotheses about how speakers perform predictions from a non-base form towards a base form: the paradigmatic view predicts that this will be a prediction like any other, while the rooted tree view predicts that what speakers are doing when processing a word form is to evaluate the predictability of the non-base form from the base, even when they are asked to perform the prediction in the opposite direction. As Relational Morphology and Construction Morphology decide on a case by case basis whether two series stand in a paradigmatic relationships or not, this family of approaches predicts that speakers have awareness of implicative relationships between series at least some of the time, but do not make a priori predictions about which series this applies to. Therefore while observing speaker behaviour that matches paradigmatic approaches to morphology is compatible with Relational Morphology and Construction Morphology, so is observing speaker behaviour that matches the rooted tree predictions.^[5]

3.1 Items

The experiment is based on French data. For each trial, participants were presented with a sentence containing two morphologically related pseudowords. For crucial items, the two pseudowords were in a derivational relationship, while for distractors, the two pseudowords were in an inflectional relationship.

As illustrated in Figure 3, items consisted of the combination of a sentence frame (j’adore le monde de la ___. Je veux être ___quand je serai grand) and a pair of pseudowords filling that frame (catonisation-catonisateur/catonisiteur/catoniseur). The same sentence could be filled by pseudowords differing in the level of predictability of the second form conditional on the first – this was quantified thanks to the Minimal Generalisation Learner, outlined below. Only one of the choices in the figure would be shown to any given participant.

Figure 3:

Sample experimental item, followed by an English translation. Only one of the forms filling the second slot is presented to each participant. The two crucial words are pseudolexemes. The three possible forms in the second sentence have different levels of predictability conditional on the knowledge that their action noun is catonisation – catonisateur would be the most expected agent noun based on the frequency of form patterns between action nouns ending in -ation and their corresponding agent noun in French, followed by catonisiteur, and lastly by catoniseur.

To operationalise the predictability of one word form conditional on another, we chose Minimal Generalisation Learner (MGL) scores (Albright 2002), due to their wide adoption in the morphological literature (and especially in Jun and Albright (2017), a study that parallels the current one in many respects, bearing on the structure of the inflectional lexicon). The MGL outputs how likely a word form is in cell C1 given a related form of the same lexeme or morphological family in cell C2. In the case above, this would be the probability that e.g. catonisateur is the agent nouns of the action noun catonisation). In order to produce this score, the MGL is trained on several pairs of phonologically transcribed word forms instantiating two different paradigm cells (for example, all agent noun ∼ action noun pairs in the French language). From this, the MGL extracts all possible form mappings between C1 and C2, using the phonological representations to both specify conditions in which each mapping can apply, and to create more generalised versions of sets of mappings whenever possible (for example, it might extract that one possible pattern between action and agent nouns in French is Xation ∼ Xateur). Once trained, form pairs can be supplied to the MGL, which will output a score for how likely form F1 is in cell C1, given that the lexeme or morphological family has form F2 in cell C2. This score is the adjusted probability that the mapping that holds between F1 and F2 applies to an input with the phonology of F2. At its core, the MGL score represents the type frequency of a morphological pattern (a variable whose impact on human linguistic behaviour is well documented, e.g. Berko Gleason (1958) and all subsequent literature on wug tasks) adjusted for the phonology of the input.

Six directed pairs of cells were chosen for the experiment (Table 2). The choice of cells is based on previous work on identifying derivational paradigms in French by Bonami and Strnadová (2019). The verb, present in the items only in the infinitive form, is assumed to be the base of both action nouns, and masculine agent deverbal nouns by traditional accounts (Figure 4). For readers who might not speak French, the French triplets {verb, action noun, masculine agent noun} have fairly close parallels in English triplets such as {banish, banishment, banisher} or {abolish, abolition, abolitioner}, in which similar complex and variable formal dependencies can be seen.

Figure 4:

The rooted tree (a) versus the paradigmatic (b) view of the links between the three chosen cells.

Within each of the six item conditions, we identified three morphological patterns of differing levels of predictability, illustrated in Table 3, and which we now describe briefly.

• verb  →  action noun . For prediction of action nouns for verbs, we focus on second conjugation verbs (infinitives in -ir) and their compatibility with different nominal suffixes. These verbs are most often associated with an action noun in -ment, instantiating the pattern Xir ∼ Xissement. In a sizeable minority of cases where -age is used instead and the pattern is Xir ∼ Xissage. Finally, a handful of morphological families match a second conjugation verb with a suffixed noun in -ion, leading to the pattern Xir ∼ Xition.

• action noun  →  verb . We focus on items in -issement. Most of these match a second conjugation verb in -ir (pattern Xissement ∼ Xir), but some match a first conjugation verb (pattern Xissement ∼ Xisser). Finally, although this is not attested, it is conceivable to have a first conjugation verb in -er, dropping the -iss- sequence (pattern Xissement ∼ Xer).

• verb  →  agent noun . For prediction of agent nouns from verbs, we look to regular first conjugation verbs in -er, and build on the fact that an agent noun in -eur can either be formed directly on the basic stem (pattern Xer ∼ Xeur) or on a learnèd stem, that will most often end in -at- (pattern Xer ∼ Xateur); and that the former situation is more prevalent than the latter. Additionally, a handful of agent nouns use instead the suffix -aire (pattern Xer ∼ Xataire), which mostly forms nouns of other semantic types, and is generally specialized to legal vocabulary unused by ordinary speakers.^[6]

• agent noun  →  verb . We look to nouns ending in -ateur. Most of these are build from a learnèd stem in -at-, and hence the expectation is for the corresponding verb to not contain the -at- sequence (pattern Xateur ∼ Xer). In a sizable minority of cases however, the -at- sequence is part of the basic stem, and hence present in both noun and verb (pattern Xateur ∼ Xater). For a most unexpected situation, we look at the unattested but clearly possible situation of a noun with a basic stem ending in -at- and a corresponding second conjugation verb in -ir (pattern Xateur ∼ Xatir).

• action noun  →  agent noun . We focus on predictors in -ation, and build on the strong tendency for morphological families with an action noun based on a learnèd stem to use that same learnèd stem in the agent noun (pattern Xation∼Xateur). A less common but attested situation is for the agent noun to differ from the action noun in being formed on the basic stem, leading to the pattern Xation∼Xeur. Finally, as our example of a most unexpected pattern, we designed items where the action noun uses a learnèd stem in -at-, like e.g. tentation ‘temptation’ but the agent noun instead a learnt stem in -it, like e.g. répétiteur ‘tutor’. This leads to the pattern Xation∼Xiteur.

• agent noun  →  action noun . We again build on the ambiguity of stem final -at-. Most agent nouns in -ateur are instances of a learnèd formation. In such a situation, the overwhelming majority of derivational families contain a matching action noun in -ion built on the same learnèd stem (pattern Xateur ∼ Xation). In some agent nouns however, the -at sequence is just the end of the basic stem. In that case the most prevalent way of forming an action noun is to use suffix -age to the basic stem leading to the pattern Xateur ∼ Xatage. Finally, in a handful of cases, the agent noun is a formed on a learnèd stem in -at-, but the action noun is formed by suffixation of -age to the basic stem, leading to the pattern Xateur ∼ Xage.

Pseudoword pairs were chosen as follows. First, we collected all pairs of lexemes that instantiated the patterns of interest in the Démonette database (Hathout and Namer 2014a) and also had phonemic transcriptions in the Flexique database (Bonami et al. 2014), a machine-readable inflected lexicon of French. Pseudolexeme predictor forms were then contructed with the help of Wuggy (Keuleers and Brysbaert 2010), by choosing a pseudolexeme fitting one of the schematic shapes of interest (e.g. Xasiɔ̴̃). A regex constraint was placed on the output form to ensure that the pseudoword had the right derivational marker. The generated input pseudoword forms were then transposed into the corresponding output pseudowords by applying appropriate patterns (leading to e.g. catonisation). Note that some of the pairs chosen do not correspond to alternations instantiated by any French lexeme pair, all the exponents featured in the items are found in the French morphological system, just not necessarily ever attested together within a single morphological family.

Table 4 provides some statistics on the distribution of MGL scores across conditions, showing that these are far from being consistent. This situation is inherent in the data, but should be kept in mind when interpreting results.

3.2 Procedure

The data was collected using an acceptability judgement task lasting an average of 20 minutes. In each trial, participants would see a video of a person producing a single utterance, containing two morphologically related forms.

The video format of the prompt was chosen for two reasons. First, the French spelling of a word gives clues different from phonological form on the rest of the morphological family. Second, the video format made it easier to signal to participants which word form they should rate. The person in the video was instructed to use nonverbal cues such as head tilts to signal the two forms of interest. The participant was asked to rate how well the second form sounded by manipulating a slider of which only the extremes were labelled, as indicated in Figure 5. As an attention check, participants were asked to provide a synonym of the pseudolexeme every fourth item.

Figure 5:

The scale employed in the experiment. Translation: does the second word sound good, as an invented word in the context of this sentence? Extreme labels: sounds bad, sounds good.

The experiment began with a description of what we meant to capture by sounds good/bad: the naturalness of the word in French, assuming the existence of a given related word form. We gave examples of neologisms based on existing words that would be more or less likely, highlighting to participants that they should make this judgement conditional on the first form provided. There were three practice trials before the experiment began.

The participants were presented with 54 crucial items, 9 from each directed cell pair above, and 24 distractors, 6 from each of four directed inflectional cell pairs (inf → prs.1pl, prs.1pl → inf, pst.ptcp.m.sg → prs.2pl, prs.2pl → pst.ptcp.m.sg). The same sentence frame could appear with three different pseudolexeme pairs, of different levels of predictability – the level of predictability of the pseudowords that each sentence appeared with was randomised. Within items for crucial cell pairs, the three levels of predictability were uniformly distributed.

The experiment was administered on a local installation of PCIBEX (Zehr and Florian 2018), allowing us to handle all data locally, and conform to GDPR requirements.

3.3 Subjects

Sixty French native speakers were recruited on Prolific.co, and compensated 9 euro an hour for their time. Further demographic data was not collected as not crucial to the predictions of the experiment.

3.4 Analysis

We fit a mixed effects zero-and-one inflated Bayesian beta regression with the brms package in R (Bürkner 2017).

The scale was coded as having 100 points (though the numbers were not visible to the speakers, who only saw the labels on the extremes), responses were divided by 100 so they would be of the appropriate scale. The model contains the following fixed effects.

• Paradigmatic predictability. To operationalize form predictability, we trained the Minimal Generalization Learner on each relevant pair of cells, using as training data all relevant pairs of forms from Démonette (Hathout and Namer 2014a) that also had a transcription in Flexique. We then asked the MGL model to predict and score all possible target forms for our predictor forms of interest. The MGL produces multiple rules of narrower or broader scope for each of the target forms it predicts from a given predictor. The score we retained is the adjusted reliability of the most reliable rule producing the target form of interest, which ranges between 0 and 1. If no rule produces the form of interest, the score is 0.

• Cell Pair. The six conditions correspond to each of the (predictor cell, target cell) combinations illustrated in Table 2. This variable was treated as a factor, and was sum-coded (contrasts are applied so that the mean of each level is compared to the overall mean of the variable, as shown in Table 5).

• Paradigmatic predictability: cell pair. The interaction of the two previous variables.

• Phonological Well-formedness. One confound that may influence participants’ judgements is the extent to which the pseudowords created are plausible French words. To take this into account, we ran a separate experiment to gather phonological well-formedness scores for each pseudoword output form. Audio recordings were presented preceded by minimal context,^[7] and participants were asked to rate the extent to which the form sounded like a plausible French word, using the scale in Figure 5. Each word was rated by 20 participants, and no participant saw two words belonging to the same pseudolexeme. The experiment took an average of 10 min, and native French speakers recruited on Prolific.co were paid 9 euro an hour. Ratings were transformed to z-scores within each participant: a z-score of 1 therefore signifies that the item was rated one standard deviation above the average by a given participant. The phonological well-formedness score for a word corresponds to the average of its standardised ratings.

We use all these variables to predict the judgement the participants made regarding the second word in the sentence. The predicted variable can be modelled with a beta distribution, since participant judgements could range between 0 and 100 (or, rescaled, 0 and 1). Since extreme values of this scale were often used by participants and some participants only used extreme values, we chose to augment the model with zero- and one- inflation, meaning that the model treats ratings of precisely 0 and 1 as potentially having been the result of a different generation process than in-between values. As is customary in analysis of experimental data, the model has random intercept for participants and sentence frames, so that the model has awareness of which observations are correlated because they belong to the same participant or to the same sentence frame. Paradigmatic predictability and well-formedness were set as random slopes over sentence frames, and paradigmatic predictability, cell, the interaction and well-formedness were set as random slopes over participants, making this a maximal model (Barr et al. 2013) and resulting in the following formula:

• judgment ∼ predictability * cell + wellformedness +

• (predictability * cell + wellformedness | participant) +

• (predictability + wellformedness | sentence_frame)

The second part of the analysis involves testing whether claims that the verb is treated as the base of deverbal nouns have a cognitive reality. If this is the case, we expect that when predicting towards the verb, the predictability information from the predictor cell will not be used. To test this, we single out predictions towards the verb and fit two models with the same formula as above, with a single difference: one uses MGL scores predicting towards the verb conditional on the predictor, and the other uses scores predicting towards the predictor conditional on the verb.