In this paper I present an extension of the controlled natural language PENG Light that is motivated by the formal properties of the Event Calculus, a narrative-based formal language for events, uents, and the periods for which these uents hold. I show how this extension can be used to express domain-specic axioms in a natural way and sketch how PENG Light and the Event Calculus can be brought together for reasoning about events and time.
2. The weather is bad. The wind is strong and the sea is rough. and to make conditional statements such as: 3. If the weather is good then John boards the hydroplane at 14:10 and arrives on Marettimo Island at 15:35. If the weather is bad then John stays in Trapani and goes to the Albergo Maccotta.
PENG Light sentences can be translated incrementally during the writing
process via discourse representation structures [
Note that in this case no additional background knowledge is required to answer these questions since the relevant facts can be looked up in the resulting knowledge base. The situation, however, is dierent for questions such as: 5. Where is John now? Where is John at 13:30? Why does John stay in Trapani?
In contrast to a machine, a human observer can easily infer from the text and his or her commonsense knowledge that John is on the bus at 12:00 but not anymore at the airport of Palermo at that time and that John is in Trapani at 13:30 but not anymore on the bus. One possible way of thinking about this problem is that if an event happens at a specic time point then this event initiates a state that holds after that time point and continues to hold until another event terminates this state. Thus, the event of John getting on the bus at a given point in time initiates that John is on the bus, and the event of John getting o the bus terminates that John is on the bus. Moreover, using the information expressed by the conditional statements above, a human observer can construct an explanation via abduction that tells us why John stays in Trapani. How can we model this behaviour using a machine? 3
The Event Calculus [
happens(E,T1), initiates(E,F), before(T1,T2), \+ clipped(T1,F,T2). 7. clipped(T1,F,T3) :
happens(E,T2), terminates(E,F), before(T1,T2), before(T2,T3).
The rst rule (6) species that a uent F holds at a time point T2 if an event E happens at a time point T1 that occurs before T2 and initiates the uent F, and it cannot be shown that F is clipped (ceases to hold) between the time point T1 and T2. Note that the negation as failure operator \ + is used to specify a form of default persistence here. The second rule (7) species that a uent F is clipped between the time point T1 and T3 if an event E happens at a time point T2 and terminates F, and T1 occurs before T2 and T2 before T3.
The relationship between events and uents is usually modeled in the Event Calculus by a set of domain-specic initiates and terminates clauses whereas the events are represented as terms and the uents as functions, for example:
A particular course of events is then represented as a set of ground happens and event clauses while the before clause keeps track of the temporal ordering: 10. happens(e1,‘10:10’). event(e1,arriving(sk1,sk2)). happens(e2,‘11:30’). event(e2,leaving(sk3,sk2)). before(‘10:10’,‘11:30’).
This gives us the required machinery that allows us to infer together with the formalised information of (1) that John is located at the airport of Palermo at 11:00 and, let’s say in Trapani at 14:00. In Section 5, we will slightly modify the above notation in order to integrate the Event Calculus smoothly with the formal output of the controlled language processor. 4
Speaking about Events and Eects in PENG Light
The discourse in (1) that communicates the relevant events can be expressed
directly in PENG Light but this information needs to be augmented with
additional domain-specic axioms that dene the eects of these events for the Event
Calculus. As explained above, events initiate or terminate uents (or release
them from the commonsense law of inertia [
Note that in contrast to (11) and (12), the eect in (13) is not only dependent on an event but also on a uent that must hold in order to terminate another uent. In PENG Light we can further restrict the domain and the range of an event by replacing the variables in the axioms by suitable nominal expressions, for example: 14. If a person gets on a vehicle then this event initiates that the person is located in the vehicle.
Note that this leads to additional constraints on the right hand side of the
rules in the Event Calculus, and we need to specify additionally in controlled
language that John is a person and that a bus is a vehicle.
PENG Light sentences are translated via discourse representation structures [
Two things are important to note here: rst, these terms are additionally wrapped by fact/1 (not shown for reasons of space); and second, since the sentence in (16) does not use an explicit temporal marker, a timestamp ( timex(sk7, ‘11:15’)) is generated in our scenario during the processing of the sentence.
The following rule then sets up the interface between the facts in (17) and the domain-independent axioms (6) and (7) of the Event Calculus: 18. happens(E,T) :- event(E,Type), theta(E,time,X), timex(X,T).
The translation of the domain-specic axioms that relate the events to their eects results also in a number of rules. For example, the three domain-specic axioms (11-13) are translated into the following rules: 19. initiates(E,fluent(X,located_at,Y)) :
event(E,arriving), theta(E,theme,X), theta(E,location,Y). 20. initiates(E,fluent(X,located_in,Y)) :
event(E,getting_on), theta(E,agent,X), theta(E,theme,Y). 21. terminates(E,fluent(X,located_at,Z)) :event(E,leaving), theta(E,agent,Y), theta(E,theme,Z), holds_at(fluent(X,located_in,Y),now).
Given this knowledge, we can now infer with the help of the Event Calculus that John is in the bus at 11:20 and still in Palermo at that time. This is because the uents in (19) and (20) that have been initiated by the occurrence of the events in (17) continue to hold until the occurrence of further events will terminate them. For example, as soon as the information becomes available that the bus leaves the airport at 11:30, the conditions in rule (21) will trigger and terminate the uent that John is located at the airport.
The Event Calculus can be extended in many interesting ways [
In this work I showed how the controlled natural language PENG Light can be extended in a systematic way to speak about events and their eects. The controlled natural language is automatically translated into the input language of a model builder. The generated formal representation can be integrated directly with the Event Calculus and be used for reasoning about events and time. This makes PENG Light an interesting specication language for creating more userfriendly interfaces to knowledge systems in dynamic domains.