Ontological Foundations of Natural Language Communication in Multiagent Systems

The paper outlines a semantic ontology as a minimal set of top-level conceptual distinctions underlying natural language communication. A semantic ontology can serve as the basis for the specification of the meaning, as the logical form, of agent messages couched in natural language. It represents a general and reusable module in the architecture of multi-agent systems involving human as well as software agents. As a practical example, we will sketch a basic multi-agent system relying on natural language communication. 1 Ontology as a Basis for Multiagent Semantics Successful communication in a multiagent system requires not only that the communicating agents share a common language, but also that they are committed to the same intended model for the semantics of this language. The semantics of a communication language is the theory that specifies the truth conditions of the messages embedded in the agents’ speech acts. Under the closed world assumption, a shared intended model may be specified as a subset of the Herbrand base, that is, the set of ground goals of the communication language. In this case an ontology can be regarded as the logic program whose declarative meaning (roughly, the set of ground goals deducible from it) is an intended model shared by a community of communicating agents. This is just a paraphrase of the classical definition of an ontology as the formal statement of a model specifying the shared understanding of cooperating agents (Gruber 1991, 1995). A semantic ontology is a conceptualisation, common to a community of agents that understand natural language, of the categories and relations that pervade the agents’ environment as a whole. It can be used to specify the logical form as the truth-functional meaning of agent messages embedded in natural language. Architecturally a semantic ontology is the most reusable component of multiagent systems involving a human-computer interface. A semantic ontology has to reflect the wired-in conceptual framework human agents are equipped with. In (Schneider 2001), a minimal semantic ontology was drafted which drew its inspiration from two sources: the semantical analysis of natural language and philosophical accounts of the commonsense view of reality. Indeed, Parsons’ (1990) account of the semantics of verbs in terms of underlying events as well as the parts played in the latter by objects can be ideally complemented by Strawson’s (1959) “descriptive metaphysics”, an attempt to specify the basic entity-types of commonsense. The paper is structured as follows. In Section 2, we present the basic conceptual distinctions that are required by a minimal semantic ontology. Events and processes have to be differentiated from the objects that participate in them; the participants of events are either physical objects or persons. In Section 3, we show how to classify different ways of participation according to the kinds of events and participants involved, giving an ontological reading of verb complementation. Section 4 sketches the role of a semantic ontology as a basis of natural language communication between agents by using a simple multi-agent architecture involving human and software agents as an example. 2 Basic distinctions in semantic ontology Dependent and Independent Entities It seems to be a fundamental feature of the human conceptual scheme that some kinds of entities, like physical objects or persons, are considered as basic, while other types of individuals, like qualities or boundaries, are regarded to be somehow dependent on the former. According to Strawson (1959:16-17), this dependence has to be understood in terms of identification in an agent’s environment. A class of particulars A (say, colours or boundaries) is identification-dependent on a class of particulars B (say, physical objects) if and only if, in order to be able to identify an instance of A, an agent has to single out an instance of B first. The commonsense distinction between dependent and independent entities is also acknowledged by recent computational upper-level ontologies, like BFO (Smith 2002) or DOLCE (Gangemi et al. 2002). In particular, the dichotomy between objects and the characteristics dependent on them is fundamental for a semantic ontology underlying natural language communication, as it motivates the grammatical difference between nouns and adjectives. The common role of nouns is to refer to objects or kinds of objects, while adjectives usually denote attributes. Of course, there are exceptions to that rule, but nominalisations of adjectives, such as “Green” or “wisdom” seem to be recognised by speakers as exceptions to a more basic semantic rule. Persons and Bodies Another distinction that is crucial for a semantic ontology is that between between mental or private characteristics (e.g. beliefs, intentions, desires) on the one hand, and physical or public characteristics (e.g. weight, colour) on the other hand. According to Strawson, our conceptual equipment is such as to posit the distinction between two types of spatio-temporal objects, namely bodies, to which only physical attributes can be ascribed, and persons which both mental and physical characteristics can be attributed to (Strawson 1959:102-103). Many natural languages reflect this distinction explictly, by gender or other systems of noun classifications. We will see that the Person/Body dichotomy even underlies the semantical subcategorisation or complementation of verbs. Thus cognitively oriented ontologies like DOLCE (Gangemi et al. 2002) have to include the difference between agentive and non-agentive objects in their taxonomies. Objects and Events Following Davidson (Davidson 1980), Parsons defends the view that the semantics of verbs and verb phrases implies the existence of events and processes (1990:4, 186-187): verbs may be considered to represent kinds of processes or events. However, the idea that the grammatical distinction between nouns and verbs is grounded on the ontological dichotomy of objects versus events or processes has been always intuited by natural language syntacticians (Tesnière 1959). Objects persist through time in virtue of core characteristics that are fully present throughout their life. Processes exist in time by having different phases at different instants, except events as instantaneous boundaries of processes (Simons 1987). Strawson argues that events or processes are dependent on objects with regard to their identification (1959:39, 45-46). Objects enjoy an ontological priority over events or states. The dependence of an occurrence on an object is called participation in DOLCE (Gangemi et al. 2002). 3 Object Participation in Language The different ways objects participate in occurrences (processes or events) have been studied by linguists interested in the phenomenon of verb complementation or thematic roles. These are partly syntactic, partly semantic relations between noun phrases and the main verb of a sentence. Thematic roles correspond to the different parts that referents play in the occurrence expressed by the verb (Parsons 1990:72-73). Table 1 shows Parsons’ (1990:73-78) list of thematic roles together with their definitions and examples. Obviously, Parsons’ empirically assembled list lacks an ontological systematisation. We count three subject-related roles: Agent, Experiencer and Performer, where the Person/Thing and Private/Public distinctions are muddled together. In (Schneider 2001), a coherent ontological account of thematic roles and ways of participation is given according to the following lines. Firstly, we will consider as basic only those thematic roles which express mere specifications of the participation relation that are neutral as to the types of occurrences or objects involved. The result is shown in Table 2. Secondly, Agent, Experiencer and Performer are defined using our four elementary thematic roles and the basic particular-types. We are here in the presence of two orthogonal oppositions: 1. Agent or Experiencer vs. Performer : the difference between personal and non-personal origins of occurrences; Thematic Roles Definition Example sentences Agent Person initiating John writes a book. the event The book is signed by John. Theme Entity affected Mary reads a book. by the event Mary blushed at his sight. Goal Addressee John gives Mary a rose. Anna writes a letter to Mary. Benefactive Entity to whose Mary gave Anne a party. benefit the John signs a book for Mary. event occurs Experiencer Person the event is Mary sees a rose. an experience of John thinks about Mary. Instrument Thing the event is John opens the letter with a knife. accomplished with Performer Thing initiating The knife opened the letter. the event Table 1. Parson’s Classification of Thematic Roles 2. Agent or Performer vs. Experiencer : the difference between a physical and a mental occurrence of which the object is an origin. To clean this orthogonal classifications up, we first define a new thematic role, namely Initiator, and redefine Performer, as restrictions of the Origin role to the object-types Person and Body respectively. An object x is an initiator of an occurrence y if and only if x is an origin of y and x is a person. An object x is a performer of an occurrence y if and only if x is an origin of y and x is a body. The thematic roles of Agent and Experiencer are then characterised as specifications of the Initiator -roles. Indeed, if x is an initiator of an occurrence y, then x is an agent of y if and only if y is a public or physical occurrence; x is an experiencer of y if and only if y is a mental or private occurrence of x. Thematic Roles Definition Example sentences Origin Entity initating John writes a book. the event A stone hits the window. The book is signed by John. The window was hit by a stone. Theme Entity affected Mary reads a book. by the event Mary blushed at his sight. Addressee Entity the John gives a rose to Mary. event is Mary gives water to her flowers.