Date: Thu, 18 Aug 2005 09:24:38 -0700 (PDT) From: Magda Dumitru Subject: The Proper Treatment of Events
AUTHORS: van Lambalgen, Michiel; Hamm, Fritz TITLE: The Proper Treatment of Events SERIES: Explorations in Semantics PUBLISHER: Blackwell Publishing YEAR: 2005
Magda Dumitru, unaffiliated scholar
The volume comes fourth in the series "Explorations in Semantics" edited by Susan Rothstein, and dedicated to students and researchers in linguistics and philosophy of language. It has three parts, the first on linguistic issues ("Time, Events, and Cognition", comprising chapters 1, 2, and 3), the second on computational and logical formalisms ("The Formal Apparatus", including chapters 4, 5, and 6), and the third on their "marriage" ("A Marriage Made in Heaven -- Linguistics and Robotics"- chapters 7 through 12). An Appendix, "The Basic of Logic Programming", is last to follow.
The semantics of tense and aspect is studied from the vantage point of cognitive science (truth and reference are cognitive, not real entities; tense and aspect are discussed in terms of goals, actions and consequences) and is approached in a fully computational way with clear pedagogical intent -- exercises are proposed at the end of several chapters, and the reader is addressed directly: "Readers not familiar with this notion are advised to read the Appendix first" (p. 51); "The reader for whom this is all new is advised to read only the statements of the theorems and skip the proofs" (p. 58); "the reader who is having trouble here is advised to refer back to that section" (p. 108); "it would be clear (at least to a mathematically inclined reader) what to do" (p. 223). An errata, supplementary material (proofs), and slides for eventual instructors can be found at http://staff.science.uva.nl/~michiell.
Chapter 1 "Time" The main question the authors ask is: 'what must our minds be like for tensed talk to make sense?'. Their answer is that human minds manage to integrate past, present, and future via planning (constructing and prioritizing events in memory). Therefore, temporal perspective ought to be the main conceptual view on time, from which other views (time as duration and time as succession) are derived. Goal-plan knowledge can relate past, present, and future when constructing narratives (executions of plans toward given goals), for instance, via episodic memory -- defined as a "generalized capacity for imagining or constructing possible worlds" (p. 13). The authors adduce psychological and neurological evidence supporting their conclusion.
Chapter 2 "Events and Time" The chapter is meant to better define the notion of event, from a cognitive point of view: "the basic building block of the human construction of time is the event" (p. 15). Semantics, defined as cognitive representation of events, must include both temporal and non-temporal relations; the latter involve goals and actions undertaken to achieve them, as well as cause and effect. An example discussed here is the French Passé Simple (PS), which may refer either to an event taken as a whole, or to its left and right boundaries. The authors further favor the Walker construction of events, instead of the Russell-Kamp model, because it is able to better account for the semantics of the PS, and because it has other important features: it allows for instants to be directed and to mark change, it assumes the notion of spacious present, and insures the derivation of continuous time from events. Humans individuate events (and hence construct time) by using object recognition (by distinguishing figure from ground) and by forming goals and plans (the raising of a hand may be part of a handshake or of a grasping action).
Chapter 3 "Language, Time, and Planning" Language production may be linked to planning via linearization. Moreover, planning is instrumental in establishing discursive coherence via anaphoric and temporal binding. The authors maintain that tense itself should be interpreted in discourse, not in single sentences (issue taken to be a counterargument to generativism). An example supporting this argument is, again, the French PS, proven to introduce events into discourse, which are then ordered thanks to world (causal) knowledge. Planning plays an important role in semantics: English future tenses involving auxiliaries always encode goals; human cognition seems to be goal-oriented, since the perfect is future oriented (the reference time lies in the future of the event time -- following Reichenbach); "Language production can be characterized as transforming a semantic structure" (p. 27).
Chapter 4 "Events Formalized" The chapter discusses event calculus (EC), a formalism for planning -- nonmonotonic (default) reasoning -- involved both in time cognition and tensed language. EC requires a first-order logic whose vocabulary includes individual objects, real numbers (to represent time), time-dependent properties (states and activities), variable quantities, and event types (marking the limits of time-dependent properties). EC allows for reification/ nominalization (properties may fill an argument slot in a predicate); perfect nominalizations are temporally related (are modeled as event types), whereas imperfect nominalizations are modeled as fluents and are not temporally related. EC also allows for predicates to include an explicit parameter for time. As for tenses, the past applies to events, whereas progressives apply to fluents. Causality, a relation between events, is taken to belong to physics and comes in two varieties: instantaneous change and continuous change. In instantaneous change, event types (e.g. collision) may become event tokens by combining with a time parameter. Event types are initiated or terminated by a fluent (fluents are time-dependent properties -- sets of intervals including initiating and terminating instants; fluents are considered as both functions and objects, which may further take a parameter x). Relevant predicates are 'Initially', 'Happens', 'Initiates', and 'Terminates'. In continuous change (e.g. drinking a glass of wine) the relevant predicates 'Trajectory', 'Releases', 'Clipped', 'Declipped', and 'HoldsAt' insure that some properties ('f-relevant events') escape the inertia specific to the first form of causation; "continuous change occurs due to a force, not an event, and hence absence of relevant events does not always entail absence of change". The definition of events includes the 'Happens' predicate, while the definition of fluents includes the 'HoldsAt' predicate.
The long-standing semantic issue of whether fundamental temporal entities should be points or intervals is thus sidestepped; the authors maintain that events and fluents, both extended in time yet playing different roles in EC, are fundamental temporal entities. Robotics needs to derive predictions, and therefore insures situation description via axioms (holding for every situation) and scenarios (holding for a particular situation). Inertia ought to be defined by a suitable logic, working alongside EC and restricting the class of models. Computational limitations lead the authors to operate with minimal models exclusively (defined as 'closed worlds' including only events and causal influenced between events, as specified for a given scenario), not with substructures of the 'real', as Discourse Representation Theory does, for instance. Developing discourse is modeled as a nonmonotonic progression (as a form of planning), which accounts for the meaning of the progressive in English and for the phenomenon of coercion.
Chapter 5 "Computing with Time and Events" "if semantics wants to make contact with the huge psycholinguistic literature on language comprehension and production, it had better become computational" (p. 49). Accordingly, the authors adopt Moschovakis' idea that the sense of an expression identifies with the algorithm that computes the denotation of that expression. The authors also cite evidence by Mani and Johnson-Laird, according to whom uniquely determined cognitive models facilitate the recall of a piece of discourse, and hence could be stored in memory. Models (which include denotations) are computed by the sense of expressions. The semantic theory proposed includes two languages: a language of predicates in EC, and a language of constraints (formulas). Both languages define a programming language: constraint logic programming, which differs from standard logic programming in that unification proceeds via equation without substitution. The logic program consists thus of a scenario and the axioms of EC. A minimal model of the logic program is defined as the completion of the program.
Chapter 6 "Finishing Touches" Nominalization is taken to be instrumental in the construction of tense and aspect, since "nominalized VPs are the basic units of semantic computation, in their guise as fluents and events" (p. 71). Montague semantics, dynamic semantics, and Davidson's analysis of events are all rejected as possible bases for analysis for different reasons: the first treats intensions via possible worlds, the second treats computation at an abstract level, and the third is not representing "an extensional two-sorted first order language in which predicates may have a time parameter" (p. 72). Events and fluents, not language formulas, are the basic computational entities, since they are involved in causal relationships.
Chapter 7 "Aktionsart" "Aktionsarten form a continuum, organized around five prototypes, rather than a discrete set" (p. 93). An Aktionsart is not "a permanent feature attached to a V or a VP" (p. 83), since verbs can be coerced into any Aktionsart. Aktionsarten are seen as eventualities(in the sense of Emmon Bach) attached to a VP, and defined as a quadruple containing an event and up to three fluents. An eventuality determines a scenario by specifying the meaning of the elements present in the quadruple. Verbs are of the form run(x, t), where x is the subject position, and t the time parameter, over which it can be abstracted in two ways (as an event type -- which is a function of time, and as a fluent -- which is an object), yielding perfect and imperfect nominals. Users don't have direct access to the time parameter, yet do have direct access to derived event types and fluents, represented as "definable finite unions of intervals and points" (p. 88). Tense, expressed by the predicates 'Happens' and 'HoldsAt', applies once the abstractions obtain. The imperfect paradox is solved by dissociating the goal (incremental theme) from the activity leading to that goal, which are then fused together in a single event.
Chapter 8 "Tense" The formalization proposed for tense makes reference to the concept of 'integrity constraints', borrowed from database theory, and expressing "obligations and prohibitions that the states of the database must satisfy if they fulfill a certain condition" (p. 99); when updating a database, care must be taken for these integrity constraints to hold under the new circumstances. Mostly event types (hence eventualities viewed perfectively) are located in time (either in a given event structure or in their representation as instants or, ideally, on the axis of the reals). Reals are needed to represent aspect. Cognition works with both event- and instant structure. The starting point for representing time is Reichenbach's tripartite structure (reference time R, event time E, and utterance time S). R is defined as an "integrity constraint formulated in terms of fluents, which typically puts constraints upon possible temporal locations of event types, including the events [fluents and event types] constructed by hierarchical planning" (p. 103). Integrity constraints help define truth for tensed sentences, in the sense that sentences are true if the integrity constraint is satisfied, and false otherwise. R is taken to be fundamental, since it is fixed by an integrity constraint, and represents an anchor for E. The (semantic) present tense takes R, E, and S to coincide. The past tense situates both R and E before now; it operates on an event type (e.g. running, writing, etc.). Unlike the perfect, the past tense requires an explicitly established past reference point. The past progressive applies to an event fluent and does not entail the completion of the event type, unlike the past tense. The author's aim is to define tense for each Aktionsart separately; however, when applying the progressive to Aktionsarten other than activities and accomplishments, the scenario must be extended in order to coerce the VP to either an activity or an accomplishment. Future events can be conceptualized as events per se, or as goals (through the use of the auxiliary 'will', the 'be going to' VP, and the futurate progressive).
Chapter 9 "Tense in French: Passé Simple and Imparfait" The chapter sketches the semantics of the French Passé Simple and Imparfait, as an application of the machinery developed in the previous chapters. Of particular interest is the ordering of events in time described by the PS; the authors maintain that the ordering can be established by computing the meaning of the PS sentences, together with world knowledge, and anaphora relations. PS obeys an integrity constraint such that an eventuality is represented as perfective and located in the past; however, PS itself cannot insure ordering relations by itself. As for the Imparfait, it obeys an integrity constraint such that it becomes represented as an "anaphoric, imperfective past tense" (p. 136). This is supplementary evidence against analyzing sentences in isolation.
Chapter 10 "Grammatical Aspect" The perfect is taken to belong to the category of grammatical aspect, and therefore it "sits uncomfortably between the two stools of tense and aspect" (p. 151). The present perfect includes, in its meaning, the idea of current relevance of an event, while the reference time is set at 'now'. The past perfect either provides a past reference time while its consequences still hold, or it is only a past in the past. The integrity constraint at work for the present perfect introduces a time at which the start event happens. The progressive is coercing an expression into either an activity or an accomplishment; moreover, the authors argue for a third truth value (indeterminate) for cases such as the 'Multiple-Choice Paradox', yet offer no convincing arguments.
Chapter 11 "Coercion" Since 'Aktionsart is not fully determined prior to the sentence level' (p. 168), aspectual coercion insures that constructions such as the progressive force a VP to change aspectual category. The authors describe three types of coercion: additive, subtractive, and cross- coercion, which mean, respectively, that a given scenario is built (by adding a direct object to a verb, for instance), that parts of a scenario are deleted, or that parts of a scenario are unified (a constant and a parameter, for instance). An interesting issue is the representation of nominals: "In our setup, the denotation of a house is as it were distributed over the changing partial object, the canonical terminating event, and the consequent state whose relations are governed by the scenario" (p. 177). In passive sentences, where NPs move to the subject position, the nominal may be reinterpreted as a real object (and hence adjectives can be applied). The authors assume that aspect imposes temporal structure on events and therefore may override the Aktionsart of a verb. The process is made possible if one interprets Frege's notion of sense as a basis for determining reference; otherwise the authors consider events not to have any canonical referents in the world. The sense of an expression is its associated scenario; one may navigate between the senses an expression may have by appealing to coercion.
Chapter 12 "Nominalization" Although the authors concede that 'Nominalization is a very complicated phenomenon" and that "we can barely scratch the surface here" (p. 207), they begin the analysis with a distinction Vendler once made, between two kinds of verbs: 'loose containers' (e.g. surprised us, is unlikely, is improving), and 'narrow containers' (e.g. took place yesterday). Loose containers take both nominal and verbal gerunds (corresponding to perfect nominals -- event types -- and imperfect nominals -- fluents, respectively), while narrow containers take only nominal gerunds. Coercion may explain this asymmetry, if it is assumed that there are two kinds of nominals (eventive and factive), and that an eventive noun is reinterpreted as a factive noun when combined with a loose container. The chapter then presents a historical view on the English gerundive system, in order to highlight the origins (phonological and morphological processes) and function of the verbal gerund (a newer type of gerund that came to function alongside the old type -- the nominal gerund). The authors agree with theories such as Houston's, according to which verbal gerunds (and participles) used to have mostly an adverbial function and introduce background information; therefore verbal gerunds are taken to introduce fluents.
The authors then propose a formalized theory of nominalization based on Feferman calculus, as follows: a perfect nominal (atemporal) can be created from a verb by suppressing its temporal parameter through existential quantification. On the contrary, imperfect nominals can be internally modified by tense and aspect, since the temporal parameter of the underlying verb is abstracted. Interestingly, the actual lexical content of perfect and imperfect nominals is of no consequence here. Determiners are taken to encode tense; event types in the restrictor are related to event tokens in the nuclear scope. Coercion regulates both the nominal and the verbal systems, considered to be interdependent. The abstract structural representation of perfect and imperfect nominals can be enriched with lexical content (determined by a scenario) and Aktionsart of underlying verbs; together, they form a temporal profile, further interacting with verbal context.
The volume has important qualities: it is very interesting and thought provoking, especially for linguists (although language engineers and philosophers may profit as well). It is an attempt at bridging the gap between artificial intelligence and linguistics in a consistent way, by discussing various core issues, such as time, tense, aspect, verbal and nominal categories, and category shifting (coercion). They all involve events, in some way or another, hence the title 'The proper treatment of events' would seem justified, though less suggestive of the wealth of topics discussed throughout the book.
As far as editing is concerned, it looks like the book was prepared for publication under considerable constraints of time and space. The style is very uneven -- part I anticipates, sometimes in great detail, conclusions and implications of the theory to be proposed and applications to be discussed later in the book; part II is rather terse, since much information is presented in very little space; part III is considerably more elaborated, sometimes too much so -- the detailed historical account of the English gerund, although very interesting, is not justified by the economy of the book, especially since the authors seem to bear it in mind -- 'Giving the full completion of the program would take up too much space' -- p. 62; 'Unfortunately, for lack of space, we will not discuss the very important topic of argument structure' -- p. 72; 'It is however a nightmare from the expository point of view to have to work with both an event structure and its representation as a continuum' -- p. 98, etc. Economy may also be the reason for not providing translations for the French example- sentences.
Other 'slips' seem to be the result of haste -- sometimes, symbols are not defined as soon as they are introduced ('r' at p. 40 is explained at p. 72; the 'imperfective paradox', first mentioned at p. 44 is discussed at p. 156,etc.), etc.; this little problem could be solved by introducing addenda. Typos are rare: 'morphologial' -- p. 151; 'as the long as' -- p. 219 etc.
Psychological and neurological evidence cited mostly in the first chapters may be interpreted as supporting the view that humans are goal-oriented agents. The authors infer from it that humans ought to apply the same cognitive mechanisms when approaching language (grammar): "It is then but one step to hypothesize that the linguistic coding of time is also driven by the future-oriented nature of our cognitive makeup". The 'step' may be very big indeed, since humans do not seem to root their goals in the present exclusively (Reichenbach's model may not be complex enough for a cognitive approach), despite the way future is encoded; also, it is not sure whether being future-oriented always means being goal-oriented. However, such considerations might link to the assumed linearity and first-order approach of the book, and hence may be seen as 'necessary evils'.
As for the celebrated "marriage" between linguistics and robotics, it does not resemble very much a union, but rather a unification, where computation seems to be what all this is about, were it not for integrity constraints: sense is defined as an algorithm leading to meaning (denotation) as given by a minimal model. Perhaps a detailed theoretical discussion here would have been too much to ask from such a dense book.
Last but not least, the authors cite a rich bibliography that, together with the pedagogical e-tools and the wealth of ideas suggested make 'The proper treatment of events' an essential book for linguistics students and researchers interested in the latest trends.
Davidson, D. 1967. The logical form of action sentences. The logic of decision and action, ed. by N. Rescher. Pittsburgh, PA, University of Pittsburgh Press.
Feferman, S. 1984. Toward useful type-free theories. Journal of Symbolic Logic, 49, pp. 75-111.
Frege, G. 1962. Sinn und Bedeutung. G. Frege: Funktion, Begriff, Bedeutung. Funf logische Studien, ed. by G. Patzig. Goettingen, Vandenhoeck.
Houston, A. 1989. The English gerund: Syntactic change and discourse function. Language change and variation, ed. by R. Fasold and D. Schiffrin. Amsterdam, Benjamins.
Kamp, H. 1979. Events, instants, and temporal reference. Semantics from different points of view, ed. by R. Baeuerle, U. Egli, and A. von Stechow, pp. 27-54, Berlin, Springer Verlag.
Mani, K. & P. N. Johnson-Laird 1982. The mental representation of spatial descriptions. Memory and Cognition, 10, pp. 181-187.
Moschovakis, Y. 1993. Sense and denotation as algorithm and value. Lecture Notes in Logic, ed. by J. Oikkonen and J. Vaananen, Natick, MA, A.K. Peters.
Reichenbach, H. 1947. Elements of symbolic logic, London, Macmillan.
ABOUT THE REVIEWER:
ABOUT THE REVIEWER
Magda Dumitru studies issues in Micro-Semantics, relating to definiteness, genericity, tense and aspect.