An OWL Ontology of Time Jerry R. Hobbs with contributions from George Ferguson, James Allen, Richard Fikes, Pat Hayes, Drew McDermott, Ian Niles, Adam Pease, Austin Tate, Mabry Tyson, and Richard Waldinger July 2004 1. Introduction A number of sites, DAML contractors and others, have developed ontologies of time (e.g., DAML-S, Cycorp, CMU, Kestrel, Teknowledge). A group of us have decided to collaborate to develop a representative ontology of time for DAML, which could then be used as is or elaborated on by others needing such an ontology. It is hoped that this collaboration will result in an ontology that will be adopted much more widely than any single site's product would be. We envision three aspects to this effort: 1. An abstract characterization of the concepts and their properties, expressed in first-order predicate calculus. 2. A translation of the abstract ontology into OWL code, to whatever extent possible given the current state of DAML expressivity. 3. Mappings between the DAML ontology and individual sites' ontologies. DAML is under development and is thus a moving target, and that is why separating 1 and 2 is desirable. Level 1 can stabilize before DAML does. A mapping in 3 may be an isomorphism, or it may be something more complicated. The reason for 3 is so DAML users can exploit the wide variety of resources for temporal reasoning that are available. Moreover, it will aid the widespread use of the ontology if it can be linked easily to, for example, the temporal portion of Teknowledge's IEEE Standard Upper Ontology effort or to Cycorp's soon-to-be widely used knowledge base. The purposes of the temporal ontology are both for expressing temporal aspects of the contents of web resources and for expressing time-related properties of web services. The following document outlines the principal features of a representative DAML ontology of time. It is informed by ontology efforts at a number of sites and reflects but elaborates on a tentative consensus during discussions at the last DAML meeting. The first three areas are spelled out in significant detail. The last three are just sketches of work to be done. There are a number of places where it is stated that the ontology is silent about some issue. This is done to avoid controversial choices in the ontolgy where more than one treatment would be reasonable and consistent. Often these issues involve identifying a one-dimensional entity and a zero-dimensional entity with one another. In general, functions are used where they are total and have a unique value; predicates are used otherwise. The order of arguments usually follows the subject-object-object of preposition order in the most natural use in an English sentence (except for "Hath", where topicalization applies). A note on notation: Conjunction (&) takes precedence over implication(-->) and equivalence (<-->). Formulas are assumed to be universally quantified on the variables appearing in the antecedent of the highest-level implication. Thus, p1(x) & p2(y) --> q1(x,y) & q2(y) is to be interpreted as (A x,y)[[p1(x) & p2(y)] --> [q1(x,y) & q2(y)]] At the end of each section there is a subsection on MAPPINGS. These are sketches of the relations between some highly developed temporal ontologies and the one outlined here. 2. Topological Temporal Relations 2.1. Instants and Intervals: There are two subclasses of temporal-entity: instant and interval. Instant(t) --> TemporalEntity(t) Interval(T) --> TemporalEntity(T) These are the only two subclasses of temporal entities. (A T)[TemporalEntity(T) --> [Instant(T) v Interval(T)]] As we will see, intervals are, intuitively, things with extent and instants are, intuitively, point-like in that they have no interior points. (In what follows, lower case t is used for instants, upper case T for intervals and for temporal-entities unspecified as to subtype. This is strictly for the reader's convenience, and has no formal significance.) _begins_ and _ends_ are relations between instants and temporal entities. begins(t,T) --> Instant(t) & TemporalThing(T) ends(t,T) --> Instant(t) & TemporalThing(T) For convenience, we can say that the beginning and end of an instant is itself. The converses of these rules are also true. Instant(t) <--> begins(t,t) Instant(t) <--> ends(t,t) The beginnings and ends of temporal entities, if they exist, are unique. TemporalEntity(T) & begins(t1,T) & begins(t2,T) --> t1=t2 TemporalEntity(T) & ends(t1,T) & ends(t2,T) --> t1=t2 As will be seen in Section 2.4, in one approach to infinite intervals, a positively infinite interval has no end, and a negatively infinite interval has no beginning. Hence, we use the relations "begins" and "ends" in the core ontology, rather than defining functions "beginning-of" and "end-of", since the functions would not be total. They can be defined in an extension of the core ontology that posits instants at positive and negative infinity. _inside_ is a relation between an instant and an interval. inside(t,T) --> Instant(t) & Interval(T) This concept of inside is not intended to include beginnings and ends of intervals, as will be seen below. It will be useful in characterizing clock and calendar terms to have a relation between instants and intervals that says that the instant is inside or the beginning of the interval. (A t,T)[beginsOrIn(t,T) <--> [begins(t,T) v inside(t,T)]] time-between is a relation among a temporal entity and two instants. timeBetween(T,t1,t2) --> TemporalEntity(T) & Instant(t1) & Instant(t2) The two instants are the beginning and end points of the temporal entity. (A t1,t2)[t1 =/= t2 --> (A T)[timeBetween(T,t1,t2) <--> begins(t1,T) & ends(t2,T)]] The ontology is silent about whether the time from t to t, if it exists, is identical to the instant t. The ontology is silent about whether intervals _consist of_ instants. The core ontology is silent about whether intervals are uniquely determined by their beginnings and ends. This issue is dealt with in Section 2.4. We can define a ProperInterval as one whose beginning and end are not identical. (A T)[ProperInterval(T) <--> Interval(T) & (A t1,t2)[begins(t1,T) & ends(t2,T) --> t1 =/= t2]] A half-infinite or infinite interval, by this definition, is proper. The ontology is silent about whether there are any intervals that are not proper intervals. 2.2. Before: There is a before relation on temporal entities, which gives directionality to time. If temporal-entity T1 is before temporal-entity T2, then the end of T1 is before the beginning of T2. Thus, before can be considered to be basic to instants and derived for intervals. (A T1,T2)[before(T1,T2) <--> (E t1,t2)[ends(t1,T1) & begins(t2,T2) & before(t1,t2)]] The before relation is anti-reflexive, anti-symmetric and transitive. before(T1,T2) --> T1 =/= T2 before(T1,T2) --> ~before(T2,T1) before(T1,T2) & before(T2,T3) --> before(T1,T3) The end of an interval is not before the beginning of the interval. Interval(T) & begins(t1,T) & ends(t2,T) --> ~before(t2,t1) The beginning of a proper interval is before the end of the interval. ProperInterval(T) & begins(t1,T) & ends(t2,T) --> before(t1,t2) The converse of this is a theorem. begins(t1,T) & ends(t2,T) & before(t1,t2) --> ProperInterval(T) If one instant is before another, there is a time between them. Instant(t1) & Instant(t2) & before(t1,t2) --> (E T) timeBetween(T,t1,t2) The ontology is silent about whether there is a time from t to t. If an instant is inside a proper interval, then the beginning of the interval is before the instant, which is before the end of the interval. This is the principal property of "inside". inside(t,T) & begins(t1,T) & ends(t2,T) --> before(t1,t) & before(t,t2) The converse of this condition is called Convexity and is discussed in Section 2.4. The relation "after" is defined in terms of "before". after(T1,T2) <--> before(T2,T1) The basic ontology is silent about whether time is linearly ordered. Thus it supports theories of time, such as the branching futures theory, which conflate time and possibility or knowledge. This issue is discussed further in Section 2.4. The basic ontology is silent about whether time is dense, that is, whether between any two instants there is a third instant. Thus it supports theories in which time consists of discrete instants. This issue is discussed further in Section 2.4. 2.3. Interval Relations: The relations between intervals defined in Allen's temporal interval calculus (Allen, 1984; Allen and Kautz, 1985; Allen and Hayes, 1989; Allen and Ferguson, 1997) can be defined in a relatively straightforward fashion in terms of "before" and identity on the beginning and end points. It is a bit more complicated than the reader might at first suspect, because allowance has to be made for the possibility of infinite intervals. Where one of the intervals could be infinite, the relation between the end points has to be conditionalized on their existence. The standard interval calculus assumes all intervals are proper, and we will do that here too. The definitions of the interval relations in terms of "before" relations among their beginning and end points, when they exist, are given by the following axioms. In these axioms, t1 and t2 are the beginning and end of interval T1; t3 and t4 are the beginning and end of T2. (A T1,T2)[intEquals(T1,T2) <--> [ProperInterval(T1) & ProperInterval(T2) & (A t1)[begins(t1,T1) <--> begins(t1,T2)] & (A t2)[ends(t2,T1) <--> ends(t2,T2)]]] intBefore(T1,T2) <--> ProperInterval(T1) & ProperInterval(T2) & before(T1,T2) (A T1,T2)[intMeets(T1,T2) <--> [ProperInterval(T1) & ProperInterval(T2) & (E t)[ends(t,T1) & begins(t,T2)]]] (A T1,T2)[intOverlaps(T1,T2) <--> [ProperInterval(T1) & ProperInterval(T2) & (E t2,t3)[ends(t2,T1) & begins(t3,T2) & before(t3,t2) & (A t1)[begins(t1,T1) --> before(t1,t3)] & (A t4)[ends(t4,T2) --> before(t2,t4)]]]] (A T1,T2)[intStarts(T1,T2) <--> [ProperInterval(T1) & ProperInterval(T2) & (E t2)[ends(t2,T1) & (A t1)[begins(t1,T1) <--> begins(t1,T2)] & (A t4)[ends(t4,T2) --> before(t2,t4)]]]] (A T1,T2)[intDuring(T1,T2) <--> [ProperInterval(T1) & ProperInterval(T2) & (E t1,t2)[begins(t1,T1) & ends(t2,T1) & (A t3)[begins(t3,T2) --> before(t3,t1)] & (A t4)[ends(t4,T2) --> before(t2,t4)]]]] (A T1,T2)[intFinishes(T1,T2) <--> [ProperInterval(T1) & ProperInterval(T2) & (E t1)[begins(t1,T1) & (A t3)[begins(t3,T2) --> before(t3,t1)] & (A t4)[ends(t4,T2) <--> ends(t4,T1))]]]] The inverse interval relations can be defined in terms of these relations. intAfter(T1,T2) <--> intBefore(T2,T1) intMetBy(T1,T2) <--> intMeets(T2,T1) intOverlappedBy(T1,T2) <--> intOverlaps(T2,T1) intStartedBy(T1,T2) <--> intStarts(T2,T1) intContains(T1,T2) <--> intDuring(T2,T1) intFinishedBy(T1,T2) <--> intFinishes(T2,T1) In addition, it will be useful below to have a single predicate for "starts or is during". This is called "starts-or-during". startsOrDuring(T1,T2) <--> [intStarts(T1,T2) v intDuring(T1,T2)] It will also be useful to have a single predicate for intervals intersecting in at most an instant. nonoverlap(T1,T2) <--> [intBefore(T1,T2) v intAfter(T1,T2) v intMeets(T1,T2) v intMetBy(T1,T2)] We could have as easily defined these in terms of "before" relations on the beginnings and ends of the intervals. So far, the concepts and axioms in the ontology of time would be appropriate for scalar phenomena in general. 2.4. Optional Extensions: In the basic ontology we have tried to remain neutral with respect to controversial issues, while producing a consistent and useable axiomatization. In specific applications one may want to have stronger properties and thus take a stand on some of these issues. In this section, we describe some options, with the axioms that would implement them. These axioms and any subsequent theorems depending on them are prefaced with a 0-argument proposition that says the option is being exercised. Thus the axiom for total ordering is prefaced by the proposition Total-Order() --> Then to adopt the option of total ordering, one merely has to assert Total-Order() Total or Linear Ordering: In many applications, if not most, it will be useful to assume that time is linearly or totally ordered. The axiom that expresses this is as follows: Total-Order() --> (A t1,t2)[Instant(t1) & Instant(t2) --> [before(t1,t2) v t1 = t2 v before(t2,t1)]] This eliminates models of time with branching futures and other conflations of time and possibility or limited knowledge. Infinity: There are two common ways of allowing infinitely long intervals. Both are common enough that it is worth a little effort to construct the time ontology in a way that accommodates both. The statements of the axioms have been complicated modestly in order to localize the difference between the two approaches to the choice between two pairs of simple existence axioms, which are themselves conditioned on 0-argument propositions indicating the choice of that option. In the first approach, one posits time instants at positive and negative infinity. Half-infinite intervals are then intervals that have one of these as an endpoint. Rather than introduce constants for these in the core ontology, we will have two predicates -- "posinf" and "neginf" -- which are true of only these points. The 0-argument proposition corresponding to the choice of this approach will be Pts-at-Inf() In the second approach, there are intervals that have no beginning and/or end. "posinf-Interval(T)" says that T is a half-infinite interval with no end. "neginf-Interval(T)" says that T is a half-infinite interval with no beginning. The 0-argument proposition corresponding to this option will be No-Pts-at-Inf() In the first approach, "posinf-interval" and "neginf-interval" will not be true of anything. In the second approach "posinf" and "neginf" will not be true of anything. The axioms that specify the properties of posinf, neginf, posinf-interval, and neginf-interval will be conditioned on the existence of such temporal entities. Thus, if an approach does not include them, the condition will never be satisfied. These axioms can thus be part of the core theory. Axioms in the core theory will make the two approaches mutually exclusive. Then which approach one takes amounts on which of two pairs of existence axioms one uses. This choice is further localized to the decision between asserting "Pts-at-Inf()" and asserting "No-Pts-at-Inf()". The arguments of the predicates "posinf" and "neginf", if they exist, are instants. posinf(t) --> Instant(t) neginf(t) --> Instant(t) The principal property of the point at positive infinity is that every other instant is before it. (A t,t1)[Instant(t1) & posinf(t) --> [before(t1,t) v t1 = t]] The next axiom entails that there are infinitely many instants after any given instant other than the point at positive infinity. (A t1)[Instant(t1) & ~posinf(t1) --> (E t2)[Instant(t2) & before(t1,t2)]] Note that these two axioms are valid in an approach that does not admit a point at positive infinity; the antecedent of Axiom 2.4-4 will never be satisfied, and the second conjunct in the antecedent of Axiom 2.4-5 will always be satisfied, guaranteeing that after every instant there will be another instant. The principal property of the point at negative infinity is that it is before every other instant. (A t,t1)[Instant(t1) & neginf(t) --> [before(t,t1) v t1 = t]] The next axiom entails that there are infinitely many instants before any given instant other than the point at negative infinity. (A t1)[Instant(t1) & ~neginf(t1) --> (E t2)[Instant(t2) & before(t2,t1)]] Likewise these axioms are valid in an approach that does not admit a point at negative infinity. In the second approach instants at positive and negative infinity are not posited, but intervals can have the properties "posinf-interval" and "neginf-interval". Because of Axiom 2.1-4, if an interval has an end, it is not a positive infinite interval. Thus, a positive infinite interval cannot have an end. An instant inside a postive half-infinite interval has infinitely many instants after it. (A t1,T)[posinf-Interval(T) & inside(t1,T) --> (E t2)[before(t1,t2) & inside(t2,T)]] This axiom is valid in the first approach as well, since "posinf-interval" will never be true and the antecedent will never be satisfied. Because of Axiom 2.1-3, if an interval has a beginning, it is not a negative infinite interval. Thus, a negative infinite interval cannot have a beginning. Corresponding to Axiom 2.4-8 is the following axiom for "neginf-interval": (A t1,T)[neginf-Interval(T) & inside(t1,T) --> (E t2)[before(t2,t1) & inside(t2,T)]] It may be useful to have two more predicates. An interval is (at least) a half-infinite interval if either "posinf-interval" or "neginf-interval" is true of it. (A T)[halfinf-Interval(T) <--> [posinf-Interval(T) v neginf-Interval(T)]] An interval is an infinite interval if it is both positively and negatively infinite. (A T)[inf-Interval(T) <--> [posinf-Interval(T) & neginf-Interval(T)]] Again these axioms are valid in the first approach because the antecedents will never be true. Finally for the core ontology, we probably want to stipulate that one either uses the "posinf" approach or the "posinf-interval" approach. This is accomplished by the following axiom. [(E t) posinf(t)] <--> ~[(E T) posinf-Interval(T)] Similarly, [(E t) neginf(t)] <--> ~[(E T) neginf-Interval(T)] Note that one could use one approach for negative infinity and the other for positive infinity, although this development does not support it. This completes the treatment of infinite time in the core ontology. The following two axioms give points at infinity if we are using the first approach, indicated by the proposition "Pts-at-Inf()" in the antecedent. Pts-at-Inf() --> (E t) posinf(t) Pts-at-Inf() --> (E t) neginf(t) That is, there are instants out at positive and negative infinity, respectively, when the Points at Infinity approach is taken. Again, to adopt this approach, simply assert Pts-at-Inf() When one adopts this approach, one can also, for convenience, extend the language to include the two constants, PositiveInfinity and NegativeInfinity, where posinf(PositiveInfinity) neginf(NegativeInfinity) One can also extend the language to include the functions "beginning-of" and "end-of", defined as follows: beginning-of(T) = t <--> begins(t,T) end-of(T) = t <--> ends(t,T) We stipulated the uniqueness of "begins" and "ends" in Section 2.1, and Axioms 2.4-12 and 2.4-13 rule out intervals with no beginnings or ends, so the functions will be total. The following two axioms guarantee the existence of half infinite intervals if one takes the "No Points at Infinity" approach. No-Pts-at-Inf() --> (A t)[Instant(t) --> (E T)[posinf-Interval(T) & begins(t,T)]] No-Pts-at-Inf() --> (A t)[Instant(t) --> (E T)[neginf-Interval(T) & ends(t,T)]] To specify that we are using the second approach, we would assert No-Pts-at-Inf() Suppose we wish to map between the two ontologies. Suppose the predicates and constants in the theory using the first approach are subscripted with 1 and the predicates in the theory using the second approach are subscripted with 2. The domains of the two theories are the same. All predicates and functions of the two theories are equivalent with the exception of "begin", "ends", "beginning-of", "end-of", "posinf", "neginf", "posinf-interval", and "neginf-interval". These are related by the following two articulation axioms. posinf1(end-of1(T)) <--> posinf-interval2(T) neginf1(beginning-of1(T)) <--> neginf-interval2(T) Density: In some applications it is useful to have the property of density, that is, the property that between any two distinct instants there is a third distinct instant. The axiom for this is as follows, where the 0-argument predicate indicating the exercising of this option is "Dense()": Dense() --> (A t1,t2)[Instant(t1) & Instant(t2) & before(t1,t2) --> (E t)[Instant(t) & before(t1,t) & before(t,t2)]] This is weaker than the mathematical property of continuity, which we will not axiomatize here. Convexity: In Section 2.2 we gave the axiom 2.2-8: inside(t,T) & begins(t1,T) & ends(t2,T) --> before(t1,t) & before(t,t2) The converse of this condition is called Convexity and may be stronger than some users will want if they are modeling time as a partial ordering. (See Esoteric Note below.) To choose the option of Convexity, simply assert the 0-argument proposition "Convex()". Convex() --> [begins(t1,T) & ends(t2,T) & before(t1,t) & before(t,t2) --> inside(t,T)] In the rest of this development anny property that depends on Convexity will be conditioned on the proposition "Convex()". Convexity implies that intervals are contiguous with respect to the before relation, in that an instant between two other instants inside an interval is also inside the interval. Convex() --> [before(t1,t2) & before(t2,t3) & inside(t1,T) & inside(t3,T) --> inside(t2,T)] Extensional Collapse: In the standard development of interval calculus, it is assumed that any intervals that are intEquals are identical. That is, intervals are uniquely determined by their beginning and end points. We can call this the property of Extensional Collapse, and indicate it by the 0-argument proposition "Ext-Collapse()". Ext-Collapse() --> (A T1,T2)[intEquals(T1,T2) --> T1 = T2] If we think of different intervals between the end points as being different ways the beginning can lead to the end, then Extensional Collapse can be seen as collapsing all these into a single "before" relation. In the rest of this development we will point it out whenever any concept or property depends on Extensional Collapse. We often think of time as isomorphic to the real numbers. The set of real numbers is one model of the theory presented here, including when one assumes total ordering, no points at infinity, density, convexity, and extensional collapse. Esoteric Note: Convexity, Extensional Collapse, and Total Ordering are independent properties. This can be seen by considering the following four models based on directed graphs, where the arcs define the before relation: 1. An interval is any subset of the paths between two nodes. (For example, time is partially ordered and an interval is any path from one node to another.) 2. An interval is the complete set of paths between two nodes. 3. An interval consists of the beginning and end nodes and all the arcs between the beginning and end nodes but no intermediate nodes. So inside(t,T) is never true. (This is a hard model to motivate.) 4. The instants are a set of discrete, linearly ordered nodes. There are multiple arcs between the nodes. The intervals are paths from one node to another, including the nodes. (For example, the instants may be the successive states in the situation calculus and the intervals sequences of actions mapping one state into the next. Different actions can have the same start and end states.) Model 1 has none of the three properties. Model 2 has Convexity and Extensional Collapse, but is not Totally Ordered. Model 3 is Totally Ordered and has Extensional Collapse but not Convexity. Model 4 is Totally Ordered and Convex, but lacks Extensional Collapse. 2.5. Linking Time and Events: The time ontology links to other things in the world through four predicates -- atTime, during, holds, and timeSpan. We assume that another ontology provides for the description of events -- either a general ontology of event structure abstractly conceived, or specific, domain-dependent ontologies for specific domains. The term "eventuality" will be used to cover events, states, processes, propositions, states of affairs, and anything else that can be located with respect to time. The possible natures of eventualities would be spelled out in the event ontologies. The term "eventuality" in this document is only an expositional convenience and has no formal role in the time ontology. The predicate atTime relates an eventuality to an instant, and is intended to say that the eventuality holds, obtains, or is taking place at that time. atTime(e,t) --> Instant(t) The predicate "during" relates an eventuality to an interval, and is intended to say that the eventuality holds, obtains, or is taking place throughout that interval. during(e,T) --> Interval(T) If an eventuality obtains during an interval, it obtains at every instant inside the interval and during every subinterval. during(e,T) & inside(t,T) --> atTime(e,t) during(e,T) & intDuring(T1,T) --> during(e,T1) Note that this means that an intermittant activity, like writing a book, does not hold "during" the interval from the beginning to the end of the activity. Rather the "convex hull" of the activity, as defined in Section 6, holds "during" the interval. Whether a particular process is viewed as instantaneous or as occuring over an interval is a granularity decision that may vary according to the context of use, and is assumed to be provided by the event ontology. Often the eventualities in the event ontology are best thought of as propositions, and the relation between these and times is most naturally called "holds". "holds(e,T)" would say that e holds at instant T or during interval T. The predicate "holds" would be part of the event ontology, not part of the time ontology, although its second argument would be be provided by the time ontology. The designers of the event ontology may or may not want to relate "holds" to "atTime" and "during" by axioms such as the following: holds(e,t) & Instant(t) <--> atTime(e,t) holds(e,T) & Interval(T) <--> during(e,T) Similarly, the event ontology may provide other ways of linking events with times, for example, by including a time parameter in predications. p(x,t) The time ontology provides ways of reasoning about the t's; their use as arguments of predicates from another domain would be a feature of the ontology of the other domain. The predicate timeSpan relates eventualities to instants or intervals. For contiguous states and processes, it tells the entire instant or interval for which the state or process obtains or takes place. In Section 6 we will develop a treatment of discontinuous temporal sequences, and it will be useful to remain open to having these as time spans of eventualities as well. timeSpan(T,e) --> TemporalEntity(T) & tseq(T) timeSpan(T,e) & Interval(T) --> during(e,T) timeSpan(t,e) & Instant(t) --> atTime(e,t) timeSpan(T,e) & Interval(T) & ~inside(t,T) & ~begins(t,T) & ~ends(t,T) --> ~atTime(e,t) timeSpan(t,e) & Instant(t) & t1 =/= t --> ~atTime(e,t1) Whether the eventuality obtains at the beginning and end points of its time span is a matter for the event ontology to specify. The silence here on this issue is the reason "timeSpan" is not defined in terms of necessary and sufficient conditions. In an extension of the time ontology, we also allow temporal predicates to apply directly to events, should the user wish. Thus, begins(t,e) says that the instant t begins the interval that is the time span of eventuality e; see the documentation for details: http://www.isi.edu/~pan/damltime/time-entry-documentation.txt Different communities have different ways of representing the times and durations of states and events (processes). In one approach, states and events can both have durations, and at least events can be instantaneous. In another approach, events can only be instantaneous and only states can have durations. In the latter approach, events that one might consider as having duration (e.g., heating water) are modeled as a state of the system that is initiated and terminated by instantaneous events. That is, there is the instantaneous event of the beginning of the heating at the beginning of an interval, that transitions the system into a state in which the water is heating. The state continues until another instantaneous event occurs---the stopping of the heating at the end of the interval. These two perspectives on events are straightforwardly interdefinable in terms of the ontology we have provided. This is a matter for the event ontology to specify. This time ontology is neutral with respect to the choice. MAPPINGS: Teknowledge's SUMO has pretty much the same ontology as presented here, though the names are slightly different. An instant is a TimePoint, an interval is a TimeInterval, beginning-of is BeginFn, and so on. SUMO implements the Allen calculus. Cyc has functions #startingPoint and #endingPoint that apply to intervals, but also to eventualities. Cyc implements the Allen calculus. Cyc uses a holdIn predicate to relate events to times, but to other events as well. Cyc defines a very rich set of derived concepts that are not defined here, but could be. For instant Kestral uses Time-Point, for interval they use Time-Interval, for beginning-of they use start-time-point, and so on. PSL axiomatizes before as a total ordering. 3. Measuring Durations 3.1. Temporal Units: This development assumes ordinary arithmetic is available. There are at least two approaches that can be taken toward measuring intervals. The first is to consider units of time as functions from Intervals to Reals. Because of infinite intervals, the range must also include Infinity. minutes: Intervals --> Reals U {Infinity} minutes([5:14,5:17)) = 3 The other approach is to consider temporal units to constitute a set of entities -- call it TemporalUnits -- and have a single function _duration_ mapping Intervals x TemporalUnits into the Reals. duration: Intervals x TemporalUnits --> Reals U {Infinity} duration([5:14,5:17), *Minute*) = 3 The two approaches are interdefinable: seconds(T) = duration(T,*Second*) minutes(T) = duration(T,*Minute*) hours(T) = duration(T,*Hour*) days(T) = duration(T,*Day*) weeks(T) = duration(T,*Week*) months(T) = duration(T,*Month*) years(T) = duration(T,*Year*) Ordinarily, the first is more convenient for stating specific facts about particular units. The second is more convenient for stating general facts about all units. The constraints on the arguments of duration are as follows: duration(T,u) --> ProperInterval(T) & TemporalUnit(u) The temporal units are as follows: TemporalUnit(*Second*) & TemporalUnit(*Minute*) & TemporalUnit(*Hour*) & TemporalUnit(*Day*) & TemporalUnit(*Week*) & TemporalUnit(*Month*) & TemporalUnit(*Year*) The aritmetic relations among the various units are as follows: seconds(T) = 60 * minutes(T) minutes(T) = 60 * hours(T) hours(T) = 24 * days(T) days(T) = 7 * weeks(T) months(T) = 12 * years(T) The relation between days and months (and, to a lesser extent, years) will be specified as part of the ontology of clock and calendar below. On their own, however, month and year are legitimate temporal units. In this development durations are treated as functions on intervals and units, and not as first class entities on their own, as in some approaches. In the latter approach, durations are essentially equivalence classes of intervals of the same length, and the length of the duration is the length of the members of the class. The relation between an approach of this sort (indicated by prefix D-) and the one presented here is straightforward. (A T,u,n)[duration(T,u) = n <--> (E d)[D-duration-of(T) = d & D-duration(d,u) = n]] At the present level of development of the temporal ontology, this extra layer of representation seems superfluous. It may be more compelling, however, when the ontology is extended to deal with the combined durations of noncontiguous aggregates of intervals. 3.2. Concatenation and Hath: The multiplicative relations above don't tell the whole story of the relations among temporal units. Temporal units are _composed of_ smaller temporal units. A larger temporal unit is a concatenation of smaller temporal units. We will first define a general relation of concatenation between an interval and a set of smaller intervals. Then we will introduce a predicate "Hath" that specifies the number of smaller unit intervals that concatenate to a larger interval. Concatenation: A proper interval x is a concatenation of a set S of proper intervals if and only if S covers all of x, and all members of S are subintervals of x and are mutually disjoint. (The third conjunct on the right side of <--> is because beginsOrIn covers only beginning-of and inside.) concatenation(x,S) <--> ProperInterval(x) & (A z)[beginsOrIn(z,x) --> (E y)[member(y,S) & beginsOrIn(z,y)]] & (A z)[ends(z,x) --> (E y)[member(y,S) & ends(z,y)]] & (A y)[member(y,S) --> [intStarts(y,x) v intDuring(y,x) v intFinishes(y,x)]] & (A y1,y2)[member(y1,S) & member(y2,S) --> [y1=y2 v nonoverlap(y1,y2)]] The following properties of "concatenation" can be proved as theorems: There are elements in S that start and finish x: concatenation(x,S) --> (E! y1)[member(y1,S) & intStarts(y1,x)] concatenation(x,S) --> (E! y2)[member(y2,S) & intFinishes(y2,x)] Except for the first and last elements of S, every element of S has elements that precede and follow it. These theorems depend on the property of Convexity. Convex() --> [concatenation(x,S) --> (A y1)[member(y1,S) --> [intFinishes(y1,x) v (E! y2)[member(y2,S) & intMeets(y1,y2)]]]] Convex() --> [concatenation(x,S) --> (A y2)[member(y2,S) --> [intStarts(y2,x) v (E! y1)[member(y1,S) & intMeets(y1,y2)]]]] The uniqueness (E!) follows from nonoverlap. Hath: The basic predicate used here for expressing the composition of larger intervals out of smaller temporal intervals of unit length is "Hath", from statements like "30 days hath September" and "60 minutes hath an hour." Its structure is Hath(N,u,x) meaning "N proper intervals of duration one unit u hath the proper interval x." That is, if Hath(N,u,x) holds, then x is the concatenation of N unit intervals where the unit is u. For example, if x is some month of September then "Hath(30,*Day*,x)" would be true. "Hath" is defined as follows: Hath(N,u,x) <--> (E S)[card(S) = N & (A z)[member(z,S) --> duration(z,u) = 1] & concatenation(x,S)] That is, x is the concatenation of a set S of N proper intervals of duration one unit u. The type constraints on its arguments can be proved as a theorem: N is an integer (assuming that is the constraint on the value of card), u is a temporal unit, and x is a proper interval: Hath(N,u,x) --> integer(N) & TemporalUnit(u) & ProperInterval(x) This treatment of concatenation will work for scalar phenomena in general. This treatment of Hath will work for measurable quantities in general. 3.3. The Structure of Temporal Units: We now define predicates true of intervals that are one temporal unit long. For example, "week" is a predicate true of intervals whose duration is one week. second(T) <--> seconds(T) = 1 minute(T) <--> minutes(T) = 1 hour(T) <--> hours(T) = 1 day(T) <--> days(T) = 1 week(T) <--> weeks(T) = 1 month(T) <--> months(T) = 1 year(T) <--> years(T) = 1 We are now in a position to state the relations between successive temporal units. minute(T) --> Hath(60,*Second*,T) hour(T) --> Hath(60,*Minute*,T) day(T) --> Hath(24,*Hour*,T) week(T) --> Hath(7,*Day*,T) year(T) --> Hath(12,*Month*,T) The relations between months and days are dealt with in Section 4.4. MAPPINGS: Teknowledge's SUMO has some facts about the lengths of temporal units in terms of smaller units. Cyc reifies durations. Cyc's notion of time covering subsets aims at the same concept dealt with here with Hath. Kestrel uses temporal units to specify the granularity of the time representation. PSL reifies and axiomatizes durations. PSL includes a treatment of delays between events. A delay is the interval between the instants at which two events occur. 4. Clock and Calendar 4.1. Time Zones: What hour of the day an instant is in is relative to the time zone. This is also true of minutes, since there are regions in the world, e.g., central Australia, where the hours are not aligned with GMT hours, but are, e.g., offset half an hour. Probably seconds are not relative to the time zone. Days, weeks, months and years are also relative to the time zone, since, e.g., 2002 began in the Eastern Standard time zone three hours before it began in the Pacific Standard time zone. Thus, predications about all clock and calendar intervals except seconds are relative to a time zone. This can be carried to what seems like a ridiculous extreme, but turns out to yield a very concise treatment. The Common Era (C.E. or A.D.) is also relative to a time zone, since 2002 years ago, it began three hours earlier in what is now the Eastern Standard time zone than in what is now the Pacific Standard time zone. What we think of as the Common Era is in fact 24 (or more) slightly displaced half-infinite intervals. (We leave B.C.E. to specialized ontologies.) The principal functions and predicates will specify a clock or calendar unit interval to be the nth such unit in a larger interval. The time zone need not be specified in this predication if it is already built into the nature of the larger interval. That means that the time zone only needs to be specified in the largest interval, that is, the Common Era; that time zone will be inherited by all smaller intervals. Thus, the Common Era can be considered as a function from time zones (or "time standards", see below) to intervals. CE(z) = T Fortunately, this counterintuitive conceptualization will usually be invisible and, for example, will not be evident in the most useful expressions for time, in Section 4.5 below. In fact, the CE predication functions as a good place to hide considerations of time zone when they are not relevant. (The BCE era is similarly time zone dependent, although this will almost never be relevant.) Esoteric Aside: Strictly speaking, the use of CE as a function depends on Extensional Collapse. If we don't want to assume that, then we can use a corresponding predicate -- CEPred(e,z) -- to mean era e is the Common Era in time zone z. We have been refering to time _zones_, but in fact it is more convenient to work in terms of what we might call the "time standard" that is used in a time zone. That is, it is better to work with *PST* as a legal entity than with the *PST* zone as a geographical region. A time standard is a way of computing the time, relative to a world-wide system of computing time. For each time standard, there is a zone, or geographical region, and a time of the year in which it is used for describing local times. Where and when a time standard is used have to be axiomatized, and this involves interrelating a time ontology and a geographical ontology. These relations can be quite complex. Only the entities like *PST* and *EDT*, the time standards, are part of the _time_ ontology. If we were to conflate time zones (i.e., geographical regions) and time standards, it would likely result in problems in several situations. For example, the Eastern Standard zone and the Eastern Daylight zone are not identical, since most of Indiana is on Eastern Standard time all year. The state of Arizona and the Navajo Indian Reservation, two overlapping geopolitical regions, have different time standards -- one is Pacific and one is Mountain. Time standards that seem equivalent, like Eastern Standard and Central Daylight, should be thought of as separate entities. Whereas they function the same in the time ontology, they do not function the same in the ontology that articulates time and geography. For example, it would be false to say those parts of Indiana shift in April from Eastern Standard to Central Daylight time. In this treatment it will be assumed there is a set of entities called time standards. Some relations among time standards are discussed in Section 4.5. 4.2. Clock and Calendar Units: The aim of this section is to explicate the various standard clock and calendar intervals. A day as a calender interval begins at and includes midnight and goes until but does not include the next midnight. By contrast, a day as a duration is any interval that is 24 hours in length. The day as a duration was dealt with in Section 3. This section deals with the day as a calendar interval. Including the beginning but not the end of a calendar interval in the interval may strike some as arbitrary. But we get a cleaner treatment if, for example, all times of the form 12:xx a.m., including 12:00 a.m. are part of the same hour and day, and all times of the form 10:15:xx, including 10:15:00, are part of the same minute. It is useful to have three ways of saying the same thing: the clock or calendar interval y is the nth clock or calendar interval of type u in a larger interval x. This can be expressed as follows for minutes: minit(y,n,x) If the property of Extensional Collapse holds, then y is uniquely determined by n and x, it can also be expressed as follows: minitFn(n,x) = y For stating general properties about clock intervals, it is useful also to have the following way to express the same thing: clockInt(y,n,u,x) This expression says that y is the nth clock interval of type u in x. For example, the proposition "clockInt(10:03,3,*Minute*,[10:00,11:00))" holds. Here u can be a member of the set of clock units, that is, one of *Second*, *Minute*, or *Hour*. In addition, there is a calendar unit function with similar structure: calInt(y,n,u,x) This says that y is the nth calendar interval of type u in x. For example, the proposition "calInt(12Mar2002,12,*Day*,Mar2002)" holds. Here u can be one of the calendar units *Day*, *Week*, *Month*, and *Year*. The unit *DayOfWeek* will be introduced below in Section 4.3. The relations among these modes of expression are as follows: sec(y,n,x) <--> secFn(n,x) = y sec(y,n,x) <--> clockInt(y,n,*Second*,x) minit(y,n,x) <--> minitFn(n,x) = y minit(y,n,x) <--> clockInt(y,n,*Minute*,x) hr(y,n,x) <--> hrFn(n,x) = y hr(y,n,x) <--> clockInt(y,n,*Hour*,x) da(y,n,x) <--> daFn(n,x) = y da(y,n,x) <--> calInt(y,n,*Day*,x) mon(y,n,x) <--> monFn(n,x) = y mon(y,n,x) <--> calInt(y,n,*Month*,x) yr(y,n,x) <--> yrFn(n,x) = y yr(y,n,x) <--> calInt(y,n,*Year*,x) Weeks and months are dealt with separately below. The am/pm designation of hours is represented by the function hr12. hr12(y,n,*am*,x) <--> hr(y,n,x) hr12(y,n,*pm*,x) <--> hr(y,n+12,x) A distinction is made above between clocks and calendars because they differ in how they number their unit intervals. The first minute of an hour is labelled with 0; for example, the first minute of the hour [10:00,11:00) is 10:00. The first day of a month is labelled with 1; the first day of March is March 1. We number minutes for the number just completed; we number days for the day we are working on. Thus, if the larger unit has N smaller units, the argument n in clockInt runs from 0 to N-1, whereas in calInt n runs from 1 to N. To state properties true of both clock and calendar intervals, we can use the predicate calInt and relate the two notions with the axiom calInt(y,n,u,x) <--> clockInt(y,n-1,u,x) Note that the Common Era is a calendar interval in this sense, since it begins with 1 C.E. and not 0 C.E. The type constraints on the arguments of calInt are as follows: calInt(y,n,u,x) --> Interval(y) & integer(n) & TemporalUnit(u) & Interval(x) We allow x to be any interval, not just a calendar interval. When x does not begin at the beginning of a calendar unit of type u, we take y to be the nth _full_ interval of type u in x. Thus, the first year of World War II, in this sense, is 1940, the first full year, and not 1939, the year it began. The first week of the year will be the first full week. We can express this constraint as follows: calInt(y,n,u,x) --> starts-or-during(y,x) Each of the calendar intervals is that unit long; for example, a calendar year is a year long. calInt(y,n,u,x) --> duration(y,u) = 1 There are properties relating to the labelling of clock and calendar intervals. If N u's hath x and y is the nth u in x, then n is between 1 and N. calInt(y,n,u,x) & Hath(N,u,x) --> 0 < n <= N There is a 1st small interval, and it starts the large interval. Hath(N,u,x) --> (E! y) calInt(y,1,u,x) Hath(S,N,u,x) & calInt(y,1,u,x) --> intStarts(y,x) There is an Nth small interval, and it finishes the large interval. Hath(N,u,x) --> (E! y) calInt(y,N,u,x) Hath(N,u,x) & calInt(y,N,u,x) --> intFinishes(y,x) All but the last small interval have a small interval that succeeds and is met by it. calInt(y1,n,u,x) & Hath(N,u,x) & n < N --> (E! y2)[calInt(y2,n+1,u,x) & intMeets(y1,y2)] All but the first small interval have a small interval that precedes and meets it. calInt(y2,n,u,x) & Hath(N,u,x) & 1 < n --> (E! y1)[calInt(y1,n - 1,u,x) & intMeets(y1,y2)] 4.3. Weeks A week is any seven consecutive days. A calendar week, by contrast, according to a commonly adopted convention, starts at midnight, Saturday night, and goes to the next midnight, Saturday night. There are 52 weeks in a year, but there are not usually 52 calendar weeks in a year. Weeks are independent of months and years. However, we can still talk about the nth week in some larger period of time, e.g., the third week of the month or the fifth week of the semester. So the same three modes of representation are appropriate for weeks as well. wk(y,n,x) <--> wkFn(n,x) = y wk(y,n,x) <--> calInt(y,n,*Week*,x) As it happens, the n and x arguments will often be irrelevant, when we only want to say that some period is a calendar week. The day of the week is a calendar interval of type *Day*. The nth day-of-the-week in a week is the nth day in that interval. dayofweek(y,n,x) <--> day(y,n,x) & (E n1,x1) wk(x,n1,x1) The days of the week have special names in English. dayofweek(y,1,x) <--> Sunday(y,x) dayofweek(y,2,x) <--> Monday(y,x) dayofweek(y,3,x) <--> Tuesday(y,x) dayofweek(y,4,x) <--> Wednesday(y,x) dayofweek(y,5,x) <--> Thursday(y,x) dayofweek(y,6,x) <--> Friday(y,x) dayofweek(y,7,x) <--> Saturday(y,x) For example, Sunday(y,x) says that y is the Sunday of week x. Since a day of the week is also a calendar day, it is a theorem that it is a day long. dayofweek(y,n,x) --> day(y) One correspondance will anchor the cycle of weeks to the rest of the calendar, for example, saying that January 1, 2002 was the Tuesday of some week x. (A z)(E x) Tuesday(dayFn(1,monFn(1,yrFn(2002,CE(z)))),x) We can define weekdays and weekend days as follows: weekday(y,x) <--> [Monday(y,x) v Tuesday(y,x) v Wednesday(y,x) v Thursday(y,x) v Friday(y,x)] weekendday(y,x) <--> [Saturday(y,x) v Sunday(y,x)] As before, the use of the functions wkFn and dayofweekFn depend on Extensional Collapse. 4.4. Months and Years The months have special names in English. mon(y,1,x) <--> January(y,x) mon(y,2,x) <--> February(y,x) mon(y,3,x) <--> March(y,x) mon(y,4,x) <--> April(y,x) mon(y,5,x) <--> May(y,x) mon(y,6,x) <--> June(y,x) mon(y,7,x) <--> July(y,x) mon(y,8,x) <--> August(y,x) mon(y,9,x) <--> September(y,x) mon(y,10,x) <--> October(y,x) mon(y,11,x) <--> November(y,x) mon(y,12,x) <--> December(y,x) The number of days in a month have to be spelled out for individual months. January(m,y) --> Hath(31,*Day*,m) March(m,y) --> Hath(31,*Day*,m) April(m,y) --> Hath(30,*Day*,m) May(m,y) --> Hath(31,*Day*,m) June(m,y) --> Hath(30,*Day*,m) July(m,y) --> Hath(31,*Day*,m) August(m,y) --> Hath(31,*Day*,m) September(m,y) --> Hath(30,*Day*,m) October(m,y) --> Hath(31,*Day*,m) November(m,y) --> Hath(30,*Day*,m) December(m,y) --> Hath(31,*Day*,m) The definition of a leap year is as follows: (A z)[leap-year(y) <--> (E n,x)[year(y,n,CE(z)) & [divides(400,n) v [divides(4,n) & ~divides(100,n)]]]] We leave leap seconds to specialized ontologies. Now the number of days in February can be specified. February(m,y) & leap-year(y) --> Hath(29,*Day*,m) February(m,y) & ~leap-year(y) --> Hath(28,*Day*,m) A reasonable approach to defining month as a unit of temporal measure would be to specify that the beginning and end points have to be on the same days of successive months. The following rather ugly axiom captures this. month(T) <--> (E t1,t2,d1,d2,n,m1,m2,n1,y1,y2,n2,e) [begins(t1,T) & ends(t2,T) [beginsOrIn(t1,d1) & beginsOrIn(t2,d2) & da(d1,n,m1) & mon(m1,n1,y1) & yr(y1,n2,e) & da(d2,n,m2) & [mon(m2,n1+1,y1) v (E y2)[n1=12 & mon(m2,1,y2) & yr(y2,n2+1,e)]]] The last disjunct takes care of months spaning December and January. So the month as a measure of duration would be related to days as a measure of duration only indirectly, mediated by the calendar. It is possible to prove that months are between 28 and 31 days. To say that July 4 is a holiday in the United States one could write (A d,m,y)[da(d,4,m) & July(m,y) --> holiday(d,USA)] Holidays like Easter can be defined in terms of this ontology coupled with an ontology of the phases of the moon. Other calendar systems could be axiomatized similarly. and the BCE era could also be axiomatized in this framework. These are left as exercises for interested developers. 5. Describing Times and Durations 5.1. Time Stamps: Standard notation for times list the year, month, day, hour, minute, and second. It is useful to define a predication for this. timeOf(t,y,m,d,h,n,s,z) <--> beginsOrIn(t,secFn(s,minFn(n,hrFn(h,daFn(d, monFn(m,yrFn(y,CE(z)))))))) Alternatively (and not assuming Extensional Collapse), timeOf(t,y,m,d,h,n,s,z) <--> (E s1,n1,h1,d1,m1,y1,e) [beginsOrIn(t,s1) & sec(s1,s,n1) & min(n1,n,h1) & hr(h1,h,d1) & da(d1,d,m1) & mon(m1,m,y1) & yr(y1,y,e) & CEPred(e,z)] For example, an instant t has the time 5:14:35pm PST, Wednesday, February 6, 2002 if the following properties hold for t: timeOf(t,2002,2,6,17,14,35,*PST*) (E w,x)[beginsOrIn(t,w) & Wednesday(w,x)] The second line says that t is in the Wednesday w of some week x. The relations among time zones can be expressed in terms of the "timeOf" predicate. Two examples are as follows: h < 8 --> [timeOf(t,y,m,d,h,n,s,*GMT*) <--> timeOf(t,y,m,d-1,h+16,n,s,*PST*)] h >= 8 --> [timeOf(t,y,m,d,h,n,s,*GMT*) <--> timeOf(t,y,m,d,h-8,n,s,*PST*)] timeOf(t,y,m,d,h,n,s,*EST*) <--> timeOf(t,y,m,d,h,n,s,*CDT*) 5.2 Calendar-Clock Descriptions To express calInt(y,n,u,x) and clockInt(y,n,u,x) directly in a description logic-based markup language, such as OWL, is inconvenient since x is itself a clock or calendar interval that requires description. So we have defined a calendar-clock or time description in OWL for specifying both calendar and clock information for a calendar-clock interval. A calendar-clock description has the following properties or fields: unitType, yearOf, monthOf, weekOf, dayOf, hourOf, minuteOf, secondOf, and timeZoneOf. The property unitType specifies the temporal unit type of the calendar-clock description, and its domain is the set of temporal units. For example, the unit type of 10:30 is minute, and the unit type of March 20, 2003 is day. The unit type is required. For a given temporal unit type, all the fields or properties for smaller units will be ignored. For instance, if the temporal unit type is day, the values of the fields or properties hourOf, minuteOf, and secondOf, if present, will be ignored. Since calendar-clock descriptions are for describing calendar-clock intervals, we have defined a property or relation, called calendarClockDescriptionOf with CalendarClockDescription as the range, for calendar-clock intervals. To express calInt(12Mar2002,12,*Day*,Mar2002), for example, using a calendar- clock description, we need an instance of CalendarClockDescription that has values only for unitType (day), yearOf (2002), monthOf (3), and dayOf (12). clockInt(10:03,3,*Minute*,[10:00, 11:00]) can be expressed similarly. More details about calendar-clock descriptions, as well as duration descriptions, together with examples used in OWL-S can be found in Pan and Hobbs [2004]. 5.3 Duration Descriptions There are two systems of time, based on different astronomical facts. The year-month system is based on the revolution of the Earth around the Sun. The week-day-hour-minute-second system is based on the rotation of the Earth around its axis. As long as we donít mix these two systems, temporal arithmetic is simple. But they donít align well, and when we try to relate days and months, complications arise, as we have already seen. We cannot simply rule out months as units, as some have suggested. Monthly rates play a very important role in commerce. If you pay $1000 a month in rent, you are paying more per day for your apartment in February than in March, and often when rents are prorated, the number of days in that specific month is used in the calculation, although in some industries months have been normalized to 28 or 30 days. It is therefore important to build a consistent system of duration measurement that involves both months and days. Here we introduce duration descriptions, in which the duration of an arbitrary finite interval can be described as a concatenation of years, months, weeks, days, hours, minutes, seconds, and fractions of seconds. The primary convention we will follow is that followed by car rental and other companies that have different rates for different periods of time. From the beginning of the interval, we fit in as many as possible of the largest unit type. Then into the remainder we fit in as many as possible of the next largest unit type, and so on. For example, when we rent a car, we pay the weekly rate for as many full weeks as we keep the car, then we pay the daily rate for any leftover full days, then the hourly rate for any leftover hours. The predication durationOf(T,y,m,w,d,h,n,s) says that duration of the interval T is y years, m months, w weeks, d days, h hours, n minutes, and s seconds. The values of the numeric arguments can be any real number, although indeterminacies will arise if we try to determine the identity of a duration described as a fractional number of months and a duration described in terms of days. The reason we allow real numbers, rather than restricting the values to integers, is that we frequently talk about such durations as one and a half months. However, for the rest of this development we will assume all of the numeric arguments are integers. The predicate durationOf can be defined in the following rather cumbersome manner: durationOf(T,y,m,w,d,h,n,s) <--> (E S,T1)[concatenation(T,S U {T}) & card(S) = y & (A v)[member(v,S) --> year(v)] & intFinishes(T1,T) & durationOf(T1,0,m,w,d,h,n,s)] durationOf(T,0,m,w,d,h,n,s) <--> (E S,T1)[concatenation(T,S U {T}) & card(S) = m & (A v)[member(v,S) --> month(v)] & intFinishes(T1,T) & durationOf(T1,0,0,w,d,h,n,s)] durationOf(T,0,0,w,d,h,n,s) <--> (E S,T1)[concatenation(T,S U {T}) & card(S) = w & (A v)[member(v,S) --> week(v)] & intFinishes(T1,T) & durationOf(T1,0,0,0,d,h,n,s)] durationOf(T,0,0,0,d,h,n,s) <--> (E S,T1)[concatenation(T,S U {T}) & card(S) = d & (A v)[member(v,S) --> day(v)] & intFinishes(T1,T) & durationOf(T1,0,0,0,0,h,n,s)] durationOf(T,0,0,0,0,h,n,s) <--> (E S,T1)[concatenation(T,S U {T}) & card(S) = h & (A v)[member(v,S) --> hour(v)] & intFinishes(T1,T) & durationOf(T1,0,0,0,0,0,n,s)] durationOf(T,0,0,0,0,0,n,s) <--> (E S,T1)[concatenation(T,S U {T}) & card(S) = n & (A v)[member(v,S) --> minute(v)] & intFinishes(T1,T) & durationOf(T1,0,0,0,0,0,0,s)] durationOf(T,0,0,0,0,0,0,s) <--> (E S,T1)[concatenation(T,S) & card(S) = s & (A v)[member(v,S) --> second(v)]] The axiom saying that an instant has 0 duration is Instant(t) --> durationOf(t,0,0,0,0,0,0,0) The predicates timeOf and durationOf can be related. Corresponding to every time is the duration of the interval from the beginning of the Common Era to that time. timeOf(t0,1,1,1,0,0,0) & timeBetween(T,t0,t) --> (A y,m,d,h,n,s)[timeOf(t,y,m,d,h,n,s) <--> durationOf(T,y-1,m-1,d-1,h,n,s)] The duration of an interval can have many different descriptions. An interval can be 1 day 2 hours, or 26 hours, or 1560 minutes, and so on. It is useful to be able to talk about these descriptions in a convenient way as independent objects and to talk about their equivalences. Thus, we define a specific kind of individual called a ěduration description", together with a number of functions relating the duration description to the values of each of the eight arguments of durationOf. Thereby we convert the 8-ary predicate durationOf into eight binary relations that are more convenient for description logic-based markup languages, such as OWL. Here is the definition of the duration description: (A T,y,m,w,d,h,n,s)[durationOf(T,y,m,w,d,h,n,s) <--> (E D)[durationDescriptionOf(D,T) & DurationDescription(D) & yearsOf(D) = y & monthsOf(D)=m & weeksOf(D)=w & daysOf(D)=d & hoursOf(D)=h & minutesOf(D)=n & secondsOf(D)=s]] We will say that a duration description is canonical if the number of weeks is zero and the number of all other units is less than the number of those units in the next higher unit. That is, there is an arbitrarily large number of years, less than 12 months, less than 24 hours, less than 60 minutes, and less than 60 seconds. The number of days is less than the number that could be consumed by one more month, given where the interval is anchored in time. The definition of canonicalDurDescr(D) is as follows: canonicalDurDescr(D) <--> [0 <= monthsOf(D) < 12 & weeksOf(D) = 0 & 0 <= hoursOf(D) < 24 & 0 <= minutesOf(D) < 60 & 0 Ł secondsOf(D) < 60 & (E T,T1,T2,t,t1,t2) [durationOf(T1,yearsOf(D),monthsOf(D),0,0,0,0,0) & durationDescriptionOf(D,T) & begins(t1,T) & begins(t1,T1) & month(T2) & intMeets(T1,T2) & ends(t2,T2) & ends(t,T) & before(t,t2)]] The existentially quantified expression at the end requires explanation. T is the interval that D describes. T is the interval starting at the same point and including only Dís year and month segments. T is a month- long interval that is appended to the end of T . The daysOf slot of D is canonical if and only if T ends before T does. The complexities of day-month arithmetic are hidden in the predicate month. MAPPINGS: Teknowledge's SUMO distinguishes between durations (e.g., HourFn) and clock and calendar intervals (e.g., Hour). Time zones are treated as geographical regions. The treatment of dates and times via functions follows Cyc's treatment. Kestrel's roundabout attempts to state rather straightforward facts about the clock and calendar are an excellent illustration of the lack of expressivity in DAML+OIL. The ISO standard for dates and times can be represented straightforwardly with the timeOf predicate or the unitFn functions. 6. Temporal Granularity Useful background reading for this note includes Bettini et al. (2002), Fikes and Zhou, and Hobbs (1985). Very often in reasoning about the world, we would like to treat an event that has extent as instantaneous, and we would like to express its time only down to a certain level of granularity. For example, we might want to say that the election occurs on November 5, 2002, without specifying the hours, minutes, or seconds. We might want to say that the Thirty Years' War ended in 1648, without specifying the month and day. For the most part, this can be done simply by being silent about the more detailed temporal properties. In Section 2.5 we introduced the predication "timeSpan(T,e)" relating events to temporal entities, the relation "temporal-description(d,t)" relating a temporal entity to a description of the clock and calendar intervals it is included in, and the functions "second-of(d)", "minute-of(d)", "hour-of(d)", "day-of(d)", "month-of(d)", and "year-of(d)". Suppose we know that an event occurs on a specific day, but we don't know the hour, or it is inappropriate to specify the hour. Then we can specify the day-of, month-of, and year-of properties, but not the hour-of, minute-of, or second-of properties. For example, for the election e, we can say timeSpan(t,e), temporal-description(d,t), day-of(d) = 5, month-of(d) = 11, year-of(d) = 2002 and no more. We can even remain silent about whether t is an instant or an interval. Sometimes it may be necessary to talk explicitly about the granularity at which we are viewing the world. For that we need to become clear about what a granularity is, and how it functions in a reasoning system. A granularity G on a set of entities S is defined by an indistinguishability relation, or equivalently, a cover of S, i.e. a set of sets of elements of S such that every element of S is an element of at least one element of the cover. We will identify the granularity G with the cover. (A G,S)[cover(G,S) <--> (A x)[member(x,S) <--> (E s)[member(s,G) & member(x,s)]]] Two elements of S are indistinguishable with respect to G if they are in the same element of G. (A x1,x2,G)[indisting(x1,x2,G) <--> (E s)[member(s,G) & member(x1,s) & member(x2,s)]] A granularity can be a partition of S, in which case every element of G is an equivalence class. The indistinguishability relation is transitive in this case. A common case of this is where the classes are defined by the values of some given function f. (A G,S)[G = f-gran(S,f) <--> [cover(G,S) & (A x1,x2)[indisting(x1,x2,G) <--> f(x1) = f(x2)]]] For example, if S is the set of descriptions of instants and f is the function "year-of", then G will be a granularity on the time line that does not distinguish between two instants within the same calendar year. The granularities defined by Bettini et al. (2002) are essentially of this nature. They will be discussed further after we have introduced temporal aggregates in Section 6 below. A granularity can also consist of overlapping sets, in which case the indistinguishability relation is not transitive. A common example of this is in domains where there is some distance function d, and any two elements that are closer than a given distance a to each other are indistinguishable. We will suppose d takes two entities and a unit u as its arguments and returns a real number. (A G,S)[G = d-gran(S,a) <--> [cover(G,S) & (A x1,x2)[indisting(x1,x2,G) <--> d(x1,x2,u) < a]]] For example, suppose S is the set of instants, d is duration of the interval between the two instants, the unit u is *Minute*, and a is 1. Then G will be the granularity on the time line that does not distinguish between instants that are less than a minute apart. Note that this is not transitive, because 9:34:10 is indistinguishable from 9:34:50, which is indistinguishable from 9:35:30, but the first and last are more than a minute apart and are thus distinguishable. Both of these granularities are uniform over the set, but we can imagine wanting variable granularities. Suppose we are planning a robbery. Before the week preceeding the robbery, we may not care what time any events occur. All times are indistinguishable. The week preceeding the robbery we may care only what day events take place on. On the day of the robbery we may care about the hour in which an event occurs, and during the robbery itself we may want to time the events down to ten-second intervals. Such a granularity could be defined as above; the formula would only be more complex. The utility of viewing the world under some granularity is that the task at hand becomes easier to reason about, because distinctions that are possible in the world at large can be ignored in the task. One way of cashing this out in a theorem-proving framework is to treat the relevant indistinguishability relation as equality. This in effect reduces the number of entities in the universe of discourse and makes available rapid theorem-proving techniques for equality such as paramodulation. We can express this assumption with the axiom (5.1) (A x1,x2)[indisting(x1,x2,G) --> x1 = x2] for the relevant G. For a temporal ontology, if 0-length intervals are instants, this axiom has the effect of collapsing some intervals into instants. There are several nearly equivalent ways of viewing the addition of such an axiom -- as a context shift, as a theory mapping, or an an extra antecedent condition. Context shift: In some formalisms, contexts are explicitly represented. A context can be viewed as a set of sentences that are true in that context. Adding axiom (5.1) to that set of sentences shifts us to a new context. Theory mapping: We can view each granularity as coinciding with a theory. Within each theory, entities that are indistinguishable with respect to that granularity are viewed as equal, so that, for example, paramodulation can replace equals with equals. To reason about different granularities, there would be a "mediator theory" in which all the constant, function and predicate symbols of the granular theories are subscripted with their granularities. So equality in a granular theory G would appear as the predicate "=_G" in the mediator theory. In the mediator theory paramodulation is allowed with "true" equality, but not with the granular equality relations =_G. However, invariances such as if x =_G y, then [p_G(x) implies p_G(y)] hold in the mediator theory. Extra antecedent condition: Suppose we have a predicate "under-granularity" that takes a granularity as its one argument and is defined as follows: (A g)[under-granularity(g) <--> (A x1,x2)[indisting(x1,x2,g) --> x1 = x2]] Then we can remain in the theory of the world at large, rather than moving to a subtheory. If we are using a granularity G, rather than proving a theorem P, we prove the theorem under-granularity(G) --> P If the granularity G is transitive, and thus partitions S, adding axiom (5.1) should not get us into any trouble. However, if G is not transitive and consists of overlapping sets, such as the episilon neighborhood granularity, then contradictions can result. When we use (5.1) with such a granularity, we are risking contradiction in the hopes of efficiency gains. Such a tradeoff must be judged on a case by case basis, depending on the task and on the reasoning engine used. 7. Aggregates of Temporal Entities 7.1. Describing Aggregates of Temporal Entities In annotating temporal expressions in newspapers, Laurie Gerber encountered a number of problematic examples of temporal aggregates, including expressions like "every 3rd Monday in 2001", "every morning for the last 4 years", "4 consecutive Sundays", "the 1st 9 months of 1997", "3 weekdays after today", "the 1st full day of competition", and "the 4th of 6 days of voting". We have taken these as challenge problems for the representation of temporal aggregates, and attempted to develop convenient means for expressing the possible referents of these expressions. In this section, we will assume the notation of set theory. Sets and elements of sets will be ordinary individuals, and relations such as "member" will be relations between such individuals. In particular, we will use the relation "member" between an element of a set and the set, and the relation "subset" between two sets. We will use "Phi" to refer to the empty set. We will use the notation "{x}" for the singleton set containing the element x. We will use the symbol "U" to refer to the union operation between two sets. The function "card" will map a set into its cardinality. In addition, for convenience, we will make moderate use of second-order formulations, and quantify over predicate symbols. This could be eliminated with the use of an "apply" predicate and axiom schemas systematically relating predicate symbols to corresponding individuals, e.g., the axiom schema for unary predicates p, (A x)[apply(*p*,x) <--> p(x)] It will be convenient to have a relation "ibefore" that generalizes over several interval and instant relations, covering both "intBefore" and "intMeets" for proper intervals. (A T1,T2)[ibefore(T1,T2) <--> [before(T1,T2) v [ProperInterval(T1) & ProperInterval(T2) & intMeets(T1,T2)]]] It will also be useful to have a relation "iinside" that generalizes over all temporal entities and aggregates. We first define a predicate "iinside-1" that generalizes over instants and intervals and covers "intStarts", "intFinishes" and "intEquals" as well as "intDuring" for intervals. We break the definition into several cases. (A T1,T2)[iinside-1(T1,T2) <--> [T1=T2 v [Instant(T1) & ProperInterval(T2) & inside(T1,T2)] v [(E t) begins(t,T1) & ends(t,T1) & ProperInterval(T2) & inside(t,T2)] v [ProperInterval(T1) & ProperInterval(T2) & [intStarts(T1,T2) v intDuring(T1,T2) v intFinishes(T1,T2) v intEquals(T1,T2)]]]] The third disjunct in the definition is for the case of 0-length intervals, should they be allowed and distinct from the corresponding instants. A temporal aggregate is first of all a set of temporal entities, but it has further structure. The relation "ibefore" imposes a natural order on some sets of temporal entities, and we will use the predicate "tseq" to describe those sets. (A s)[tseq(s) <--> (A t)[member(t,s) --> TemporalEntity(t)] & (A t1,t2)[member(t1,s) & member(t2,s) --> [t1 = t2 v ibefore(t1,t2) v ibefore(t2,t1)]]] That is, a temporal sequence is a set of temporal entities totally ordered by the "ibefore" relation. A temporal sequence has no overlapping temporal entities. It will be useful to have the notion of a temporal sequence whose elements all have a property p. (A s,p)[tseqp(s,p) <--> tseq(s) & (A t)[member(t,s) --> p(t)]] A uniform temporal sequence is one all of whose members are of equal duration. (A s)[uniform-tseq(s) <--> (A t1,t2,u)[member(t1,s) & member(t2,s) & TemporalUnit(u) --> duration(t1,u) = duration(t2,u) The same temporal aggregate can be broken up into a set of intervals in many different ways. Thus it is useful to be able to talk about temporal sequences that are equivalent in the sense that they cover the same regions of time. (A s1,s2)[equiv-tseq(s1,s2) <--> tseq(s1) & tseq(s2) & (A t)[TemporalEntity(t) --> [(E t1)[member(t1,s1) & iinside-1(t,t1)] <--> (E t2)[member(t2,s2) & iinside-1(t,t2)]]]] That is, s1 and s2 are equivalent temporal sequences when any temporal entity inside one is also inside the other. A minimal temporal sequence is one that is minimal in that its intervals are maximal, so that the number of intervals in minimal. We can view a week as a week or as 7 individual successive days; the first would be minimal. We can go from a nonminimal to a minimal temporal sequence by concatenating intervals that meet. (A s)[min-tseq(s) <--> (A t1,t2)[member(t1,s) & member(t2,s) --> [t1 = t2 v (E t)[ibefore(t1,t) & ibefore(t,t2) & ~member(t,s)]]]] That is, s is a minimal temporal sequence when any two distinct intervals in s have a temporal entity not in s between them. A temporal sequence s1 is a minimal equivalent temporal sequence to temporal sequence s if s1 is minimal and equivalent to s. (A s1,s)[min-equiv-tseq(s1,s) <--> min-tseq(s1) & equiv-tseq(s1,s)] We can now generalize "iinside-1" to the predicate "iinside", which covers both temporal entities and temporal sequences. A temporal entity is "iinside" a temporal sequence if it is "iinside-1" one of the elements of its minimal equivalent temporal sequence. (A t,s)[iinside(t,s) <--> [TemporalEntity(t) & TemporalEntity(s) & iinside-1(t,s)] v [TemporalEntity(t) & tseq(s) & (E s1,t1)[min-equiv-tseq(s1,s) & member(t1,s1) & iinside-1(t,t1)]]] We can define a notion of "isubset" on the basis of "iinside". (A s,s0)[isubset(s,s0) <--> [tseq(s) & tseq(s0) & (A t)[member(t,s) --> iinside(t,s0)]]] That is, every element of temporal sequence s is inside some element of the minimal equivalent temporal sequence of s0. We can also define a relation of "idisjoint" between two temporal sequences. (A s1,s2)[idisjoint(s1,s2) <--> [tseq(s1) & tseq(s2) & ~(E t,t1,t2)[member(t1,s1) & member(t2,s2) & iinside(t,t1) & iinside(t,t2)]]] That is, temporal sequences s1 and s2 are disjoint if there is no overlap between the elements of one and the elements of the other. The first temporal entity in a temporal sequence is the one that is "ibefore" any of the others. (A t,s)[first(t,s) <--> [tseq(s) & member(t,s) & (A t1)[member(t1,s) --> [t1 = t v ibefore(t,t1)]]]] The predicate "last" is defined similarly. (A t,s)[last(t,s) <--> [tseq(s) & member(t,s) & (A t1)[member(t1,s) --> [t1 = t v ibefore(t1,t)]]]] More generally, we can talk about the nth element of temporal sequence. (A t,s)[nth(t,n,s) <--> [tseq(s) & member(t,s) & natnum(n) & (E s1)[(A t1)[member(t1,s1) <--> [member(t1,s) & ibefore(t1,t)]] & card(s1) = n-1]]] That is, the nth element of a temporal sequence has n-1 elements before it. It is a theorem that the first is the nth when n is 1, and that the last is the nth when n is the cardinality of s. (A t,s)[first(t,s) <--> nth(t,1,s)] (A t,s)[last(t,s) <--> nth(t,card(s),s)] Later in this development it will be convenient to have a predicate "nbetw" that says there are n elements in a sequence between two given elements. (A t1,t2,s,n)[nbetw(t1,t2,s,n) <--> (E s1)[card(s1) = n & (A t)[member(t,s1) <--> ibefore(t1,t) & ibefore(t,t2) & member(t,s)]]] It may sometimes be of use to talk about the convex hull of a temporal sequence. (A t,s)[convex-hull(t,s) <--> [tseq(s) & Interval(t) & (A t1)[first(t1,s) --> intStarts(t1,t)] & (A t2)[last(t2,s) --> intFinishes(t2,t)]]] Note,however, that we cannot simply dispense with temporal sequences and talk only about their convex hulls. "Every Monday in 2003" has as its convex hull the interval from January 6 to December 29, 2003, but if we use that interval to represent the phrase, we lose all the important information in the notice "The group will meet every Monday in 2003." The predicate "ngap" will enable us to define "everynthp" below. Essentially, we are after the idea of a temporal sequence s containing every nth element of s0 for which p is true. The predicate "ngap" holds between two elements of s and says that there are n-1 elements between them that are in s0 and not in s for which p is true. (A t1,t2,s,s0,p,n) [ngap(t1,t2,s,s0,p,n) <--> [member(t1,s) & member(t2,s) & tseqp(s,p) & tseq(s0) & isubset(s,s0) & natnum(n) & (E s1)[card(s1) = n-1 & idisjoint(s,s1) & (A t)[member(t,s1) <--> [iinside(t,s0) & p(t) & ibefore(t1,t) & ibefore(t,t2)]]]]] The predicate "everynthp" says that a temporal sequence s consists of every nth element of the temporal sequence s0 for which property p is true. It will be useful in describing temporal aggregates like "every third Monday in 2001". (A s,s0,p,n)[everynthp(s,s0,p,n) <--> [tseqp(s,p) & tseq(s0) & natnum(n) & (E t1)[first(t1,s) & ~(E t)[iinside(t,s0) & ngap(t,t1,s,s0,p,n)]] & (E t2)[last(t2,s) & ~(E t)[iinside(t,s0) & ngap(t2,t,s,s0,p,n)]] & (A t1)[last(t1) v (E t2) ngap(t1,t2,s,s0,p,n0)]]] That is, the first element in s has no p element n elements before it in s0, the last element in s has no p element n elements after it, and every element but the last has a p element n elements after it. The variable for the temporal sequence s0 is, in a sense, a context parameter. When we say "every other Monday", we are unlikely to mean every other Monday in the history of the Universe. The parameter s0 constrains us to some particular segment of time. (Of course, that segment could in principle be the entire time line.) The definition of "everyp" is simpler. (A s,s0,p)[everyp(s,s0,p) <--> (A t)[member(t,s) <--> [iinside(t,s0) & p(t)]]] It is a theorem that every p is equivalant to every first p. (A s,s0,p)[everyp(s,s0,p) <--> everynthp(s,s0,p,1)] We could similarly define "every-other-p", but the resulting simplification from "everynthp(s,s0,p,2)" would not be sufficient to justify it. Now we will consider a number of English expressions for temporal aggregates and show how they would be represented with the machinery we have built up. "Every third Monday in 2001": In Section 4.3, "Monday" is a predicate with two arguments, the second being for the week it is in. Let us define "Monday1" as describing a Monday in any week. (A d)[Monday1(d) <--> (E w) Monday(d,w)] Then the phrase "every third Monday in 2001" describes a temporal sequence S for which the following is true. (E y,z)[yr(y,2001,CE(z)) & everynthp(S,{y},Monday,3)] Note that this could describe any of three temporal sequences, depending on the offset determining the first element of the set. "Every morning for the last four years": Suppose "nowfn" maps a document d into the instant or interval viewed as "now" from the point of view of that document, and suppose D is the document this phrase occurs in. Suppose also the predicate "morning" describes that part of each day that is designated a "morning". Then the phrase describes a temporal sequence S for which the the following is true. (E T,t)[duration(T,*Year*) = 4 & ends(t1,T) & iinside(t1,nowfn(D)) & everyp(S,{T},morning)] "Four consecutive Mondays": This describes a temporal sequence S for which the following is true. (E s0)[everyp(S,s0,Monday1) & card(S) = 4] "The first nine months of 1997": This describes the temporal sequence S for which the following is true. (E z)(A m)[member(m,S) <--> month(m,n,yrFn(1997,CD(z))) & 1 =< n =< 9] Note that this expression is ambiguous between the set of nine individual months, and the interval spanning the nine months. This is a harmless ambiguity because the minimal equivalent temporal sequence of the first is the singleton set consisting of the second. "The first full day of competition": For the convenience of this example, let us assume an ontology where "competition" is a substance or activity type and "full" relates intervals to such types. Then the phrase describes an interval D for which the following is true. (E s)[(A d)[member(d,s) <--> (E n,T)[day(d,n,T) & competition(c) & full(d,c)]] & first(D,s)] "Three weekdays after January 10": Suppose the predicate "weekday1" describes any weekday of any week, similar to "Monday1". Then this phrase describes the temporal aggregate S for which the following is true. (E d,y,T)[da(d,11,moFn(1,y)) & everyp(S,{T},weekday1) & intStarts(d,T) & card(S) = 3] That is, January 11, the day after January 10, starts the interval from which the three successive weekdays are to be taken. The last of these weekdays is the day D for which "last(D,S)" is true. If we know that January 10 is a Friday, we can deduce that the end of S is Wednesday, January 15. "The fourth of six days of voting": Let us, for the sake of the example, say that voting is a substance/activity type and can appear as the first argument of the predicate "during". Then a voting day can be defined as follows: (A d)[voting-day(d) <--> (E v,n,T)[da(d,n,T) & voting(v) & during(v,d)]] Then the phrase describes an interval D for which the following is true. (E s,s0)[everyp(s,s0,voting-day) & card(s) = 6 & nth(D,4,s)] Betti et al.'s (19??) concept of granularity is simply a temporal sequence in our terminology. All of the examples they give are uniform temporal sequences. For example, their "hour" granularity within an interval T is the set S such that "everyp(S,T,hr1)", where "hr1" is to "hr" as "Monday1" is to "Monday". Their notion of one granularity "grouping into" another can be defined for temporal sequences. (A s1,s2)[groups-into(s1,s2) <--> tseq(s1) & tseq(s2) & iinside(s1,s2) & (A t)[member(t,s2) --> (E s)[subset(s,s1) & min-equiv-tseq({t},s)]]] That is, temporal sequence s1 groups into temporal sequence s2 if every element of s2 is made up of a concatenation of elements of s1 and nothing else is in s1. Betti et al. also define a notion of "groups-periodically-into", relative to a period characterized by integers r. Essentially, every r instances of a granule in the coarser granularity groups a subset of the same number of granules in the finer granularity. (A s1,s2,r)[groups-periodically-into(s1,s2,r) <--> groups-into(s1,s2) & natnum(r) & (A t1,t2,s3)[member(t1,s2) & member(t2,s2) & nbetw(t1,t2,s2,r-1) & subset(s3,s1) & groups-into(s3,{t1}) --> (E s4)[subset(s4,s1) & groups-into(s4,{t2}) & card(s3) = card(s4)]]]]] To know the time of an event down to a granularity of one clock hour (in the context of S0) is to know which element it occurred during in the set S such that "everyp(S,S0,hr1)". A transitive granularity, as defined in Section 5, is a temporal sequence. 6.2. Durations of Temporal Aggregates The concept of "duration", defined in Section 3, can be extended to temporal sequences in a straightforward manner. If a temporal sequence is the empty set, Phi, its duration is zero. (A u)[TemporalUnit(u) --> duration(Phi,u) = 0] The duration of a singleton set is the duration of the temporal entity in it. (A t,u)[TemporalEntity(t) & TemporalUnit(u) --> duration({t},u) = duration(t,u)] The duration of the union of two disjoint temporal sequences is the sum of their durations. (A s1,s2,u)[tseq(s1) & tseq(s2) & idisjoint(s1,s2) & TemporalUnit(u) --> duration(s1 U s2,u) = duration(s1,u) + duration(s2,u)] We need to use the predicate "idisjoint" to ensure that there is no overlap between intervals in s1 and intervals in s2. The duration of the convex hull of a temporal sequence is of course not the same as the duration of the temporal sequence. Sometimes one notion is appropriate, sometimes the other. For determining what workers hired on an hourly basis should be paid, we want to know the duration of the temporal sequence of the hours that they worked, whereas for someone on an annual salary, the appropriate measure is the duration of its convex hull. It is a theorem that the duration of the convex hull of a temporal sequence is at least as great as that of the temporal sequence. (A t,s,u)[convex-hull(t,s) --> duration(t,u) >= duration(s,u)] 7.3. Duration Arithmetic five business days after January 8, 2003. 7.4. Rates 8. Deictic Time Deictic temporal concepts, such as ``now'', ``today'', ``tomorrow night'', and ``last year'', are more common in natural language texts than they will be in descriptions of Web resources, and for that reason we are postponing a development of this domain until the first three are in place. But since most of the content on the Web is in natural language, ultimately it will be necessary for this ontology to be developed. It should, as well, mesh well with the annotation standards used in automatic tagging of text. We expect that the key concept in this area will be a relation _now_ between an instant or interval and an utterance or document. now(t,d) The concept of "today" would also be relative to a document, and would be defined as follows: today(T,d) <--> (E t,n,x)[now(t,d) & beginsOrIn(t,T) & da(T,n,x)] That is, T is today with respect to document d if and only if there is an instant t in T that is now with respect to the document and T is a calendar day (and thus the nth calendar day in some interval x). Present, past and future can be defined in the obvious way in terms of now and before. Another feature of a treatment of deictic time would be an axiomatization of the concepts of last, this, and next on anchored sequences of temporal entities. 9. Vague Temporal Concepts In natural language a very important class of temporal expressions are inherently vague. Included in this category are such terms as "soon", "recently", and "a little while". These require an underlying theory of vagueness, and in any case are probably not immediately critical for the Semantic Web. This area will be postponed for a little while. References Allen, J.F. (1984). Towards a general theory of action and time. Artificial Intelligence 23, pp. 123-154. Allen, James F., and Henry A. Kautz. 1985. ``A Model of Naive Temporal Reasoning'', {\it Formal Theories of the Commonsense World}, ed. by Jerry R. Hobbs and Robert C. Moore, Ablex Publishing Corp., pp. 251-268. Allen, J.F. and P.J. Hayes (1989). Moments and points in an interval-based temporal logic. Computational Intelligence 5, pp. 225-238. Allen, J.F. and G. Ferguson (1997). Actions and events in interval temporal logic. In Oliveiro Stock (ed.), Spatial and Temporal Reasoning, Kluwer Academic Publishers, pp. 205-245. Claudio Bettini, X. Sean Wang, and Sushil Jajodia, "Solving multi-granularity temporal constraint networks", Artificial Intelligence, vol. 140 (2002), pp. 107-152. Richard Fikes and Qing Zhou, "A Reusable Time Ontology" Jerry Hobbs, "Granularity", IJCAI-85, or http://www.ai.sri.com/~hobbs/granularity-web.pdf Pan, Feng and Hobbs, Jerry R. 2004. Time in OWL-S. In Proceedings of AAAI Spring Symposium on Semantic Web Services, Stanford University, CA, March 2004.