Semantic analysis of verb – noun zero derivation in Princeton WordNet

1 Introduction

Zero derivation is the morphological process of creating new words by means of what is called a zero suffix,^[1] which stands out among the other affixes as it lacks form. The process is also called conversion^[2] by many language specialists. The existence of this process, and, consequently, the legitimacy of the zero suffix are controversial (Borer 2013; Lieber 1992, 2004; Myers 1984; Pesetsky 1995; Plag 1999).

However, as shown by Bauer et al. (2013), the arguments against the existence of zero derivation are not completely convincing. Moreover, psychological experiments (Darby and Lahiri 2016; Fruchter et al. 2013; Pliatsikas et al. 2014) have brought arguments in favor of the existence of a derivational zero suffix. A further argument for its existence is that there are overt suffixes whose function is also shared by the zero suffix (Sanders 1988). Consider the pair drive _V – driver _N versus ally _V – ally _N : the noun denoting the person performing the activity expressed by the verb is created by adding the suffix -er to the verb in the former case, and by zero derivation, i.e., by adding the zero suffix to the verb, in the latter. Various meanings associated with zero derivation were observed and described (Bauer et al. 2013; Cetnarowska 1993; Clark and Clark 1979; Lieber 2004; Plag 1999).

The zero suffix is characterized by high ambiguity. Semantically, there is a diversity of possible meanings the zero suffix can add to the base word: when forming zero-derived nouns, it may denote the person who does the activity expressed by the base word (see ally _V – ally _N above), the means by which this activity is done (cover _V – cover _N), the action expressed by the verb (drain _V – drain _N), etc. Morphologically, different parts of speech can be created (Schönefeld 2005): noun from verb – see examples above, verb from noun: document _N – document _V, verb from adjective: clean _Adj – clean _V, noun from adjective: final _Adj – final _N. We focus here on the zero derivation involved in English verb-to-noun or noun-to-verb pairs of word senses as represented in our dataset. We adopt a semantic perspective, with the aim of answering the following research questions: what are the semantic regularities involved in zero derivation (i.e., meaning relations between the members of a pair) and what are the semantic conditions under which it occurs (i.e., the semantic classes the nouns and verbs belong to)?

Directionality in zero derivation is an ongoing research issue. While many researchers favor a directionality analysis, even if not all of them adopt a derivational approach (Arad 2005; Clark and Clark 1979; Kiparsky 1982; Lieber 1992, 2004), others support a non-derivational stance (Borer 2013; Farrell 2001). There are also those who adopt a bi-directional approach (Becker 1993; Tribout 2020). We adopt the view that the pairs related through zero derivation involve directionality. However, the direction of derivation is not encoded in the data we use. Moreover, it is not always straightforward to determine. This is why we cannot include directionality in our analysis but can only make judgments with respect to individual pairs of word senses and possibly in light of certain semantic patterns, such as the ones presented in Section 5.

We first present our dataset (Section 2) starting with an overview of the Princeton WordNet, its organizing principles and the morphosemantic relations used to label the pairs in its standoff file.^[3] We briefly outline the methodology, paying particular attention to the properties of the dataset, the general distribution of zero derivation and affixal derivation across the whole set of derivationally related pairs in the dataset which provides a background for the analysis of zero derivation, with no indication of the derivation direction. We also take into account the distribution of zero and affixal derivation per morphosemantic relation, which already offers some insight into the semantics of the two types of derivation. Section 3 contains an overview of the distribution of the five most frequent affixes for each morphosemantic relation with an eye on their semantic underspecification and their frequency for each possible sense (in terms of the verb and noun classes they connect and the relation expressed). Further semantic analysis of zero derivation and affixal derivation is presented in Section 4: for each morphosemantic relation we identify the most frequent combinations of noun senses and verb senses belonging to particular semantic classes (as expressed by their semantic primes, see Section 2.1) with zero derivation and/or with affixal derivation and remark on some tendencies that are observed in the data. In Section 5, a deeper look into the PWN structure helps us identify some semantic groupings (defined in the form of PWN subtrees) where pairs of morphosemantically related word senses tend to gather. Finally, we draw conclusions in Section 6 and envisage possible lines of research emerging from the observations of the data.

2 Dataset compilation and research methodology

Below we present the main resource employed in the study – Princeton WordNet, with a view to the data and features which we use for the dataset compilation. Further, we outline the methodology used in the analysis of the data.

2.2 Dataset compilation

The dataset that we subjected to analysis consisted of a total of 16,624 verb – noun sense pairs obtained from the initial dataset included in the PWN standoff file (containing 17,741 pairs), after removing duplicate entries and errors in the relation annotation process and changing some relations for the purposes of consistency (the exact procedures, involving correcting duplicate morphosemantic relations and inconsistencies in the morphosemantic relations assigned are presented in detail by Koeva et al. (2016).

A verb – noun pair can occur several times in the PWN standoff file, each time at least one of the words having a different sense number. For example, the relation between the noun sense net:2 and the verb sense net:4 (presented above) is Instrument.^[5] Other senses of the two words are also considered derivationally related and an appropriate semantic relation is attached to them, as shown in Table 3.

Table 3:

Semantic relations attached to derivational relations.

Verb gloss	Verb	Semantic relation	Noun	Noun gloss
“catch with a net”	net:4	Instrument	net:2	“a trap made of netting to catch fish or birds or insects”
“yield as a net profit”	net:2	Result	net:3	“the excess of revenues over outlays in a given period of time (including depreciation and other non-cash expenses)”
“construct or form a web, as if by weaving”	net:3	Result	net:6	“an open fabric of string or rope or wire woven together at regular intervals”

Figure 1 shows the distribution of each relation in the dataset. We notice that Event accounts for a little less than half of the total number of pairs and Agent for almost a fifth of the pairs. The remaining relations are less frequent.

Figure 1:

Distribution of semantic relations in the PWN standoff file.

Overall, the proportion of zero derivation in our dataset is 46.5%, which shows that it is extremely productive in English (Plag 1999). The distribution of zero derivation and affixal derivation with respect to each of the 14 morphosemantic relations is presented in Figure 2, which confirms that zero derivation expresses a variety of meanings and that no specific one can be defined for it (Clark and Clark 1979; Plag 1999): all morphosemantic relations have cases of zero derivation.

Figure 2:

Distribution of derivation across the 14 morphosemantic relations.

The distribution of zero and overt affixal derivation per morphosemantic relation shows three types of situations:

1. a higher proportion of zero derivation – for Uses, Location, Undergoer, Property, and Result,

2. a higher proportion of affixal derivation – for Agent, Destination, Material, State, and

3. almost equal distribution – for Vehicle, By-means-of, Body-part, Event, and Instrument.

3 The distribution of zero and overt affixes

In this section we look at the semantics of the zero suffix, namely its meaning contribution to the sense of the derived word: this is reflected in the semantic relation that labels the pair of word senses that are morphologically analyzed as cases of zero derivation in our dataset.

For the purpose of this paper, for each pair in the standoff file we made explicit the affix(es) by means of which one word in the pair is obtained from the other. The identification of affixes was done automatically, using a manually compiled list of English prefixes and suffixes and a set of rules for attaching these affixes to the stem (e.g., transformation of final y preceded by a consonant into i before attaching the suffix, doubling the final consonant preceded by an unstressed vowel, etc.). The results of the automatic process were manually corrected to ensure an exact representation of the data. A first remark on this is that the nominal suffixes are much better represented than the verbal ones in out dataset. For each semantic relation a clear picture of the affixes involved and of their frequency emerges. Table 4 shows the most frequent suffixes for each relation.

The proportional contribution of each suffix to rendering the morphosemantic relations is represented in Figure 3. Each relation is dominated by one or two affixes at most and one of these is always the zero suffix (Ø). It is the most frequent suffix with the relations By-means-of, Body-part, Event, Location, Property, Result, Undergoer, Uses and Vehicle and the second most frequent with the relations Agent, Instrument, and State. For Destination it competes with the verb suffix -ify for the second place, while for Material it takes third place.

Figure 3:

Distribution of suffixes within morphosemantic relations.

The zero suffix is the only one that can be found with all the relations, thus being the most underspecified among the suffixes in the dataset. As shown in Table 5, it has Event as default reading (45.41% of all its occurrences), but Result (with 11.34%), By-means-of (with 8.69%), Undergoer (with 8.52%) and Uses (with 8.40%) are also important readings. However, it is clearly not the only underspecified suffix, as other (noun) suffixes are also found with most of the relations, having various default and prevailing readings:

1. -ion (including cases with -(a)tion/-sion) favors the Event relation (76.58% out of the 12 relations with which it is found),

2. -er favors the Agent relation (76.34% out of the 12 relations^[6] with which it is found),

3. -ment is less underspecified (covering 10 relations) and has Event as its default reading (65.83%), while State (with 10.50%) and By-means-of (with 8.93%) are also available,

4. -ance, -al and -ee are even less underspecified (covering only 8, 6 and 4 relations, respectively) with clear preference for particular relations – Event for -ance and -al, Undergoer and Destination for -ee.

As far as the verbal suffixes are concerned, Table 5 reveals that for both -ify and -ise/-ize the Result meaning prevails, although with the former this reading is more evident (55.55%) than with the latter (39.13%). Other notable meanings for -ify are By-means-of (18.05%) and Uses (8.33%). For -ise/-ize Event (13.77%), as well as Instrument (9.42%) and Uses (7.25%) are also important readings. For -ate the predominant reading is Result (26.67%), but there are also examples of Agent, By-means-of and Property (13.33% each).

4 Semantic primes: between zero derivation and affixal derivation

The semantic conditions under which zero derivation (Ø) and affixal derivation occur in terms of the semantic classes of the base and the derived word senses are presented below. In PWN, these semantic classes are defined by the semantic primes presented in Table 1 above. We thus use ‘prime’ as an abstract reference to the relevant semantic class when we study the regularities between the semantics of the verb and noun senses entering a particular morphosemantic relation.

Below we sum up the characteristics of each relation: these were identified by looking at the most frequent verb – noun prime pairs within the relation estimated in terms of the number of verb – noun pairs for each prime pair, and trying to further specify the relation’s properties on the basis of the observed data. In addition, we looked at some clear-cut patterns that were identified in the data, and especially at those that were elucidating with respect to the semantics of the relation. We represent them by means of generalized definitions abstracted from the synsets’ glosses, such as ‘provide (with an) N’; these correspond to semantic patterns in the sense intended by Marchand (1964) and to Plag’s semantic representations (Plag 1999). Where possible in our analysis, we inspected all the relevant pairs, but where instances were too many (e.g. Event, with more than 8,000 verb – noun pairs), we looked at the overall tendencies.

Below, the relations are presented according to the prevalent type of derivation – zero derivation or affixation, with a third category reserved for those relations for which the distribution of these two types are rather similar.

5 Semantic groupings

In this section we touch upon a deeper level of analysis of verb – noun morphosemantics that emerges from the PWN data and builds upon prime combinations;^[8] it may, however, be found implicitly in previous research as well. In addition to and within verb – noun prime pairs we identified more specific semantic patterns that often correspond to hierarchical substructures in the lexicon (organized roughly as PWN hypernymic subtrees). Below we provide a glimpse into the potential of this kind of analysis although to paint a comprehensive picture would involve further analysis of the data.

Researchers have aligned derivational semantics (zero derivation in particular) with the semantics of the classes of verbs. Two such accounts are probably best known. Clark and Clark (1979), focusing on concrete nouns as bases for creating verbs by zero suffixation, describe five types of verbs that are further subdivided: locatum verbs (blanket the bed), location and duration verbs (kennel the dog, summer in Paris), agent and experiencer verbs (butcher the cow, witness an accident), goal and source verbs (powder the aspirin, word a sentence) and instrument verbs (bicycle into town). Different perspectives are assumed for semantically classifying nouns, too: for example, for agentive verbs the semantic classes of the parent nouns are occupations (butcher the cow), special roles (referee a game) and animals (parrot every word). Plag (1999) distinguishes among the following types of zero derived verbs, also suggesting a general semantic interpretation for each type: locative “put (in)to X” (jail), ornative “provide with X” (staff), causative “make (more) X” (yellow), resultative “make into X” (bundle), inchoative “become X” (cool), performative “perform X” (counterattack), similative “act like X” (chauffeur), instrumental “use X” (hammer), privative “remove X” (bark), stative “be X” (hostess).

The predefined 14 PWN morphosemantic relations determine similar or more specific semantic classes than the ones already described in the literature; some of the following non-exhaustive trends are observed. Clark and Clark’s agentive and Plag’s similative, which in essence is a kind of agentive, both correspond to the relation Agent; the same goes for the relation between instrumental verbs and their base nouns, on the one hand, and the Instrument morphosemantic relation, on the other. Locatum verbs correspond to the PWN subtree of cover:1 “provide with a covering” or coat:1 “cover the surface of” and possibly other subtrees; Clark and Clark’s location/Plag’s locative verbs are in line with the subtree of put:1. Ornative verbs may be aligned with the subtree of supply:1 “give something useful or necessary to”. Locatum, location and ornative verbs, as described by the authors, are linked to their base nouns by means of the relation Uses in the account provided here.

As these examples show, given the hierarchical structure of nouns and verbs in PWN, the grouping of verbs and/or nouns with similar semantics correspond to hypernymic substructures in the PWN hierarchy. Thus, the morphosemantic relations may be viewed not just as connecting individual noun and verb senses, but as linking semantically coherent parts of the lexicon, i.e. subtrees. The PWN hierarchical structure allows us to observe more (often narrower) morphosemantically related formations in addition to the ones proposed by Clark and Clark (1979) and Plag (1999).

To illustrate this, we consider several such formations with the relation Undergoer. One is found between verbs of consumption and nouns denoting food and drink such as drink:1 – drink:1, wine:1 – wine:1, claret:1 – claret:2, port:7 – port:2, tipple:1 – tipple:2, sup:1 – sup:1, taste:3 – taste:5. These verb senses are all hyponyms of consume:2 “serve oneself to, or consume regularly” and all nouns denote the food or drink consumed, thus are hyponyms of food:1, so the subtree rooted in these two synsets (consume:2 and food:1) are a grouping in our understanding of the term. Within the same relation (Undergoer) we also find another such grouping with the hyponyms of gather:1 “assemble or get together” linked to hyponyms of food:1: berry:1 – berry:1, blackberry:1 – blackberry:1, mushroom:1 – mushroom:5, oyster:1 – oyster:2, snail:1 – snail:2 where the noun senses denote animals or plants serving as food that are being collected (the meaning encoded by the respective verb). A similar coherent grouping of verb – noun sense pairs is formed by hyponyms of hunt:1 “pursue for food or sport” and nouns denoting animals and/or fish, e.g. whale:1 – whale:2, seal:6 – seal:9, fowl:2 – fowl:1, crab:3 – crab:1, shark:2 – shark:1, etc. An interesting venue for further research would be a more detailed comparison with observations on such groupings provided in existing literature.

To analyze the regularities and interdependencies between derivational models (zero derivation in particular) and semantic relations between verb and noun senses that emerge from the data, we performed semantic grouping of verb – noun pairs for each morphosemantic relation. For a given relation we consider PWN verb (sub)trees and we investigate the verb senses that enter a morphosemantic relation R and their semantic primes (verb groupings). Secondly, in many verb trees for a relation R there are noticeable groups of verb senses which exhibit similar preferences for certain types of noun senses – in terms of their semantic prime and in terms of their place in the nouns’ PWN tree (noun groupings).

Further, we illustrate how the study of morphosemantic relations between the members of subtrees reveals the morphological and semantic relatedness of hierarchically structured parts of the lexicon. In Figures 4 and 5 subtrees rooted in a given synset are represented as circles of different color and size, some of which are subsets of larger supersets. The numbers in the brackets in each set represent the number of pairs in the current set (the number on the left) out of the number of pairs in the closest superset that contains it (the number on the right).

Figure 4:

The interdependence between groupings for V.communication – N.communication with Event.

Figure 5:

The interdependence between groupings for V.change – N.substance with Uses.

In Figure 4 we show one semantic grouping for the relation Event with prime pairs displaying both zero and affixal derivation. We notice that different verb groupings tend to enter morphosemantic relations with particular noun groupings. For example, with regards to the Event relation the synsets in the tree of communicate:2 predominantly relate to nouns in the tree of communication:2, while verbs in the tree of state:1 relate to nouns in the subgroups of speech act:1 and statement:1.

Similar trends are observed when we focus on groupings within zero-derivation pairs. In Figure 5 we show one semantic grouping for the relation Uses with pairs displaying zero derivation. For 18 pairs from our dataset, the noun subtree is rooted in the node matter:3 and the verbal one stems from the node change:1. We can notice that 6 verbs belong to the subsubtree rooted in process:1, while 16 nouns belong to the subsubtree rooted in substance:1. Out of these 16 nouns, 11 are hyponyms of material:1, and 6 of these 11 are grouped under the node chemical:1.

A further in-depth study into such groupings will provide a more profound look into the frequency and semantic patterns involved in zero derivation (with a recourse to affixation as well) and how they pattern across semantic classes. From a practical perspective these observations also help identify: (i) morphosemantic relations that are not typical for pairs of noun and verb subtrees, that is, another relation is prevalent between the members of these trees; hence, the untypical one may be a signal for a wrong automatic assignment: e.g., in the subtree of cover:1 there is a grouping of verbs related to nouns from the subtree of food:1 which in the majority of cases enter into the relation of Uses (egg:2 – egg:2, flour:1 – flour:1, etc.), with the exception of butter:1 where the relation assigned is Undergoer; thus Undergoer needs to be further validated or ruled out; (ii) relation gaps – missing morphosemantic relations between derivationally related verb – noun sense pairs identified in pairs of subtrees with groupings of morphosemantically related pairs: e.g., in the subtrees of the pair communicate:1 – communicator:1 where we observe 15 pairs representing the Agent relation, we also identify the derivationally related pair of synsets supplicate:1 – supplicant:2, suppliant:1, for which the relation Agent has not been encoded in PWN. These observations also facilitate the definition of more derivational models, involving a wider set of suffixes and taking into consideration phonetic transformations. These questions will be pursued in our future work.

6 Conclusions

The detailed analysis of the relationship between derivationally related word senses by taking into consideration their semantic primes in the PWN framework sheds light into the frequency of zero derivation versus affixation, into the frequency of zero derivation across relations, and of zero derivation with respect to pairs of semantic primes and smaller semantic groupings (PWN subtrees).

We have shown that the zero suffix is highly frequent in our dataset, highly underspecified with respect to the semantic relations it can express, and covers the whole spectrum of (14) relations defined over the dataset. It stands out among the affixes present in the data, as none of the overt ones, although sometimes highly underspecified too, may express all the relations. The zero suffix manifests preferences with respect to the relations expressed: its main meaning is Event, but Result, By-means-of, Undergoer and Uses are also important. At least in the data we worked with, it is also the main affix for creating new words for most of the relations (By-means-of, Body-part, Event, Location, Property, Result, Undergoer, Uses and Vehicle), it comes second in this respect with other relations (Agent, Instrument, State) and occupies the third position for the relation Material.

With respect to the semantic conditions under which zero derivation occurs, we have adopted a comparative approach between zero derivation and affixal derivation, and the most frequent pairs of semantic classes were presented for each morphosemantic relation and for each type of derivation. They show that although most of these combinations occur with both types of derivation, some of them are better represented for one type or the other. There are cases when, although zero derivation prevails with a certain relation, some pairs of semantic primes are more frequent with affixal derivation: see the pair V.social – N.person for Undergoer, or V.change – N.substance or V.change – N.state for Result: both relations are dominated by zero derivation, but these prime pairs are dominated by affixal derivation.

The pairs are extracted from a language resource (PWN) whose structure favors the analysis in terms of morphosemantically related groupings of words. They highlight semantic subtrees in which certain derivational tendencies can be noticed and, starting from them, undiscovered ones can be brought to light.

The results presented here are obtained by studying a large number of pairs of verbs and nouns. They offer more concrete arguments to already formulated linguistic characterizations of the phenomena under study, reveal certain tendencies and combinatorial preferences, as well as refine previous analyses. Although we cannot argue for the representativity of our PWN-based dataset for the entire English language, it is large enough to allow for generalizations about the semantics of zero and overt suffixes, which could be tested in future studies.

Further research should reconsider these results by adding another dimension to their interpretation, namely the direction of derivation.