next up previous contents
Next: Lexicalization Up: Linguistic aspects of lexical Previous: Lexical Semantic Relations


Semantic Roles


Semantic relations were introduced in generative grammar during the mid-1960s and early 1970s ([Fil68], [Jac72], [Gru67]) as a way of classifying the arguments of natural language predicates into a closed set of participant types which were thought to have a special status in grammar. A list of the most popular roles and the properties usually associated with them is given below.
A participant which the meaning of the verb specifies as doing or causing something, possibly intentionally. Examples: subjects of kill, eat, hit, smash, kick, watch.

a participant which the verb characterizes as having something happen to it, and as being affected by what happens to it. Examples: objects of kill, eat, smash but not those of watch, hear, love.

A participant who is characterized as aware of something. Examples: subject of love, object of annoy.

A participant which is characterized as changing its position or condition, or as being in a state or position. Examples: objects of give, hand, subjects of walk, die.

The thematic role associated with the NP expressing the location in a sentence with a verb of location. Examples: subjects of keep, own, retain, know, locative PPs.

Object from which motion proceeds. Examples: subjects of buy, promise, objects of deprive, free, cure.

Object to which motion proceeds. Examples: subject of receive, buy, dative objects of tell, give. (adapted from ([Dow89])

In linguistic theory, thematic roles have traditionally been regarded as determinant in expressing generalizations about the syntactic realization of a predicate's arguments (see [EAG96]).


The theoretical status of semantic roles in linguistic theory is still a largely unresolved issue. For example, there is considerable doubt about whether semantic roles should be regarded as syntactic, lexical or semantic/conceptual entities. Another open issue, connected with the previous one, is whether semantic roles should be considered a primitive part of linguistic knowledge (see, among others, [Fil68] [Fil77], [Dik89], [Wil81], [Cho86], [Bel88]) or as a derivative notion of some specific aspect of the form-meaning mapping ([Jac72], [Jac90], [Rap88]). However, the most common understanding is that semantic roles are semantic/conceptual elements (see, among others, [Jac72], [Jac90], [Rap88], [Dik89]).

Most characterizations of thematic roles have been carried out in terms of primitive semantic properties of predicates. For example, [Jac72] suggested that thematic relations should be defined in terms of the three semantic subfunctions CAUSE, CHANGE and BE which constitute some of the primitive building blocks of lexical conceptual representations. According to this treatment, the lexical-conceptual representation of a transitive verb like open would be as shown below where NP1 is interpreted as agent and NP2 as theme.


An analogous proposal was developed by [Dow79] within a Montague Grammar framework and later adopted and extended by [Fol84].

In addition to [Dow79], other model-theoretic formalizations are presented in [Car84] and [Dow89]. In [Dow89], Dowty defines thematic role types as abstractions above individual thematic roles of specific verbs. The individual thematic role of a verb is defined as the set of all properties which the verb entails for a given argument position. For example the individual role for the subject argument of the verb love would correspond to the set of properties which can be attributed to the first argument of the predicate love through semantic entailment, e.g. the properties which characterize the lover participant. A thematic role type can then be defined as the intersection of some set of individual thematic roles. For example, the role type ``recipient'' would be the set of all entailed properties shared by a particular individual role of verbs such as give, sell, buy, receive and tell.

As Dowty himself hastens to point out, this method is not guaranteed to yield useful results. Even assuming that each individual role will effectively intersect with at least another individual role, the number of resulting role types might just be too big to be useful at all. More generally, the identification of an appropriate set of semantic roles [Pal94] is problematic; in practice this means the number of roles varies significantly across different proposals.

The problems just pointed out have led several scholars -- e.g. [Som87], [Roz89], [Dow91] -- to put forward alternative conceptions of semantic roles. Although from different angles, these all criticize the use of necessary and sufficient conditions for the identification of roles, and advocate more flexible approaches. These approaches appear particularly suitable for the construction of large scale lexicons since they overcome many problems of role identification inherent to traditional approaches, i.e. the difficulty in enumerating precise criteria which qualify the conceptual makeup of a given semantic role.

[Dow91] proposes to abandon the use of discrete role types to provide a total indexing of verbal arguments in favour of a weaker method where the relation between role types and clusters of entailments of verb meanings need not be unique. Dowty assumes that there are only two ``thematic-role-like concepts'' for verbal predicates: the proto-agent and proto-patient role. Proto-roles are conceived of as ``cluster-concepts'' which are determined for each choice of predicate with respect to a given set of semantic properties. The properties which contribute to the definition of the proto-agent and proto-patient roles are listed below.

Contributing Properties for the Proto-Agent Role
sentience (and/or perception)
causes event

Contributing Properties for the Proto-Patient Role
change of state (including coming-to-being, going-out-of-being)
incremental theme (i.e. determinant of aspect, see section on lexical aspect below)
vcausally affected by event
stationary (relative to movement of Proto-Agent)

According to Dowty, proto-roles are essentially meant for argument selection, e.g. lexical assignment of grammatical functions to subcategorized arguments.

The work of [Dow91] has been taken as the starting point of the EAGLES recommendations on the encoding of thematic roles [EAG96]).

[San92a,San92b,San93a,San93b] propose to extend the functionality of Dowty's prototype roles by including in the defining clusters properties which are instrumental for the identification of semantic verb (sub)classes. For example, it is well known that at least six subtypes of psychological verbs can be distinguished according to semantic properties of the stimulus and experiencer arguments (see [Jac90] and references therein), as shown in Table 2.1.

Table 2.1: Psychological verb subclasses
non-causative source neu. reactive emotive experience
non-causative source pos. reactive emotive admire
non-causative source neg. reactive emotive fear
neutral caus. source neu. affected emotive interest
positive caus. source pos. affected emotive delight
negative caus. source neg. affected emotive scare

This characterization of psychological verbs can be formally rendered by defining a lattice of thematic sorts relative to the stimulus and experiencer arguments which extend prototype agent and patient roles, as shown in Fig 2.1.
Figure 2.1: Protoroles sorts for psychological predicates.

[Ash95] assume that both causation and change can be specified along the following dimensions so as to yield a thematic hierarchy such as the one described in the lattice structure in Fig 2.2.

Figure 2.2: Thematic hierarchy.

[SanFCa] proposes to enrich this characterization by

Mapping between Semantic Roles and Grammatical Relations

Semantic roles are assumed to be the source of grammatical relations in many linguistic theories. Grammar frameworks such as Government and Binding (GB), Lexical Functional Grammar (LFG) and Functional Grammar (FG) all posit a level of semantic, or thematic, relations to which grammatical relations systematically relate. In particular, semantic roles are the standard devices used for organising predicate argument structures within the lexicon, where arguments are identified on the basis of semantic roles. GB, LFG and FG follow a lexicalist approach to grammar which makes the lexicon the source of syntactic representations; this implies that grammatical relations are, one way or another, projected from predicate argument structures specified within the lexicon.

The principles guiding the mapping of lexical representations onto syntactic structures vary across the different theories. A first distinction can be made between multi-stratal frameworks such as Government and Binding and mono-stratal ones such as Lexical Functional Grammar and Functional Grammar: whereas in the former the mapping is onto D- structure representations, in the latter the projection is directly onto surface representations. Hence, in GB the attention is focused on the way in which thematic roles are mapped onto structural positions at D-structure; the actual syntactic realization of these roles in the surface of the sentence is then accounted for at the level of the mapping between D- and S-structure. By contrast, LFG and FG link semantic relations directly to their surface syntactic realization. From this it follows that the mapping conditions in the two cases are different. In multi-stratal frameworks, D-structure appears to be a pure structural representation of thematic relations, regardless of their syntactic expressions; this could explain why GB lexical representations do not systematically have to specify the syntactic expression of arguments. In mono-stratal frameworks, such mapping conditions have to account for the variation in the syntactic realization of the same semantic relation in the surface of the sentence.

In spite of these different conceptions of the mapping between semantic and syntactic relations, all frameworks considered here share the general assumption that the relationship between semantic and syntactic relations is constrained by some sort of hierarchy of semantic roles. This idea dates back to [Fil68] who first formulated the view that subject selection is in some way sensitive to a hierarchy of ``cases'', i.e. semantic relations. Following Fillmore, most theories invoke (though to a different extent) a mapping between an ordered list of semantic (i.e. a hierarchy) and an ordered list of grammatical relations (either expressed as different positions within phrase markers, or explicitly or implicitly organized in hierarchical terms). Given a thematic/semantic hierarchy (agent > theme ...) and a syntactic hierarchy (subject > object ...), the general form of the mapping is as follows: map the semantic roles of a given argument structure, which have been ordered according to the hierarchy, into the syntactic hierarchy from left to right. Under this view, the mapping is controlled by hierarchical, i.e. relative, strategies (that is, ``higher'' semantic roles are mapped onto ``higher'' syntactic relations), rather than invariable correspondence relations (such as a given semantic role always maps onto a given grammatical relation).

However, this mapping between hierarchies is not always sufficient to predict the syntactic realization of an argument. Whenever this is the case, the mapping is constrained through additional information on the syntactic realization of arguments; this practice is adopted, though to a different extent, within the GB and LFG frameworks. In GB, when the syntactic expression of an arguments cannot be predicted on the basis of general rules, as in the case of psychological verbs, this information is specified at the level of lexical representations in the form of Case-grid [Bel88]. By contrast, LFG lexical representations systematically include the ``syntactic function assignment'', i.e. the explicit stipulation of the syntactic realization of verb's arguments; according to latest developments [Bre89], this specification is made in underspecified form. Due to its peculiar conception of grammatical relations, FG never contains specifications of this kind: since subject and object selection is made on the basis of pragmatic considerations, the mapping between semantic and syntactic functions only defines the range of possible syntactic realizations, thus stating preferences rather than constraints within the range of possible mappings.

A kind of regular mapping between grammatical relations and semantic roles is also assumed in Dowty's conception of proto-roles and further developments. In fact, proto-roles are related to argument selection through the so-called ``Argument Selection Principle'', according to which the argument for which the predicate entails the greatest number of Proto-Agent properties will be lexicalized as the subject of the predicate and the argument having the greatest number of Proto-Patient properties will, all else being equal, be lexicalized as the direct object. The basic idea underlying this approach to argument selection is that the ranking according to which the arguments of a verb compete with one another with respect to subjecthood and objecthood is provided by the clustering of semantic properties, rather than by the mapping between specific positions (say between Agent and Subject). This is to say that argument selection of subject and object is determined by the total number of Proto-Agent entailments and Proto-Patient entailments shown by each argument of a verb.

To sum up, three different aspects have been taken as defining features of the kind of mapping between lexical and syntactic representations, i.e. whether:

All grammar frameworks considered in this brief survey have been characterised with respect to these features.

Comparing Approaches

Semantic roles are dealt with under a number of different names in the linguistic literature, including thematic relations, participant roles, deep cases, semantic case/roles and theta roles. Though many of these terminological distinctions carry specific implications regarding theoretical status, there is a shared underlying intuition that certain characteristic ``modes of participation'' in an event can be identified and generalised across a variety of verbs. Such modes of event participation are spelled out in terms of basic concepts such as cause, change, be. All the approaches reviewed do in some sense follow this practice, although they differ as to the number and type of basic concepts used.

Concerning formalization, three treatments can be distinguished. First, approaches which rely on an informal specification such as Jackendoff's lexical conceptual structures [Jac90]. Second, approaches such as those proposed by [Dow79], [Dow89] which are developed within a model-theoretic framework. Third approaches which provide an algebraic specification within a typed feature structure formalism; these tend to be more oriented towards NLP applications (e.g. [San92b], [San93a], [San93b], [SanFCa]).

Relation to other Areas of Lexical Semantics

Lexical aspect

Since Dowty's and Verkuyl's pioneering work on aspect compositionality during the early seventies [Dow72], [Ver72], it has been a well known fact that many non-stative verbs can give rise to either a telic or atelic interpretation according to whether their theme NP has quantized or cumulative reference. Informally, a nominal predicate has quantized reference if it responds positively to the additivity test, e.g. sugar and sugar makes sugar. Conversely, with a quantized nominal there is no proper subpart of the nominal which has the same denotational properties of the NP, e.g. no proper subpart of a chair is a chair. The basic generalization concerning the interaction of nominal and temporal reference with respect to aspect compositionality can be briefly stated as follows:
a theme NP which has cumulative reference induces a durative reading at the sentential level, while with a theme NP which has quantized reference a terminative reading obtains.
This pattern is shown in the examples below where a sequence of two question marks indicates incompatibility -- under a single event reading -- between a quantified NP and the durative adverbial all day which forces an atelic (i.e. durative) interpretation on the sentences.
John drank beer all day
?? John drank a glass of beer all day
[Kri89], [Kri90] has argued that the contribution of nominal reference to sentence aspect should be characterized by establishing a homomorphism between algebraically structured NP and event denotata in such a way that subsequent stages of change in the NP are reflected in developmental stages of the event. Consider, for example, a sentence such as John drank a glass of wine. An intuitive way of characterizing Krifka's approach would be to say that by monitoring the level of wine in the glass we would be able to establish whether the event is at its outset, halfway done, completed etc.: for each subpart of liquid in the glass there is a subevent of drinking. Another important feature of Krifka's approach consists in regarding thematic roles as links between nominal reference and temporal constitution. More specifically, some thematic roles are regarded as having transfer properties which allow the object-to-event homomorphism referred above to take place.

Encoding in Lexical Databases

The classification of suggested in [San92b] and [San93a] has been used in the ACQUILEX lexicons (§3.10.3) to provide representations for psychological and motion verbs in English (see [Cop93]).

Thematic roles are also used in the EUROTRA MT lexica (§ 3.9.1), DELIS (§3.10.5), the EDR Concept Description Dictionary (§3.6).

Relevance to LE Applications

Thematic or semantic roles represent another type of grammatical relation - a semantic one - holding between a predicate and its arguments, which can be usefully exploited in the framework of NLP applications. Information on the thematic role borne by arguments with respect to their predicates is useful to abstract away from the issue of their concrete syntactic realization (e.g. in terms of grammatical functions such as subject, object indirect object and the like). In Machine Translation, for instance, this information can be used to map the predicate-argument structures of two translation equivalents such as, for example, English like and Italian piacere, although the syntactic realization of the two verbs is radically different: namely, what is the subject of English like becomes an indirect object in Italian piacere, while what is the object of like is turned into the subject in the Italian translation. A characterization of the argument structure of the two verbs in terms of semantic roles attains the purpose of neutralizing their different syntactic behaviour in the specific language. Thematic roles have been used in Machine Translation (§4.1) to express generalizations about complex translation equivalence (see [Dor90], [Dor93], [San92b], [Cop93]). Another possible use of semantic roles in the framework of NLP applications is based on their semantic content which, in principle, can be used to infer the semantic role borne by a given constituent in a sentence. In a syntactically - either structurally or functionally - ambiguous context, recognition of the semantic role borne by a given constituent can help to resolve the syntactic ambiguity due to the pervasive regularities observed in the mapping between semantic roles and grammatical relations (see §2.4.3). This kind of hypothesis is entertained and corroborated by psycholinguistic studies on language comprehension showing that thematic information is highly instrumental in the resolution of local ambiguities and garden-paths [Car88], [Pri88], [Sto89]. In particular, the results of the experiments carried out in these studies show that there appears to be a persistent default pattern of thematic assignment throughout natural language - animate subjects are Agents and objects are Themes, inanimate subjects are Themes. These results validate the hypothesis that thematic assignments have a bearing on syntactic parsing which is thus performed on the basis of semantic information. Yet, when the exploitation of this hypothesis is considered for disambiguation purposes in the framework of wide coverage NLP systems, the picture emerging from psycholinguistic studies changes radically due to the impossibility of having a coherent characterisation of semantic roles while keeping their being the source of coherent syntactic representations [Mon95].

next up previous contents
Next: Lexicalization Up: Linguistic aspects of lexical Previous: Lexical Semantic Relations
EAGLES Central Secretariat