The Perceptual Origin of Thematic Relations

On the Acquisition of Lexical Entries: The Perceptual Origin of Thematic Relations

James Pustejovsky Department of Computer Science Brandeis University Waltham, MA 02254 617-736-2709 jamesp~br andeis.csnet-relay

Abstract that the conceptual abstraction of thematic information does not develop arbitrarily but along a given, predictable path; namely, a developmental path that starts with tan- This paper describes a computational model of concept gible perceptual predicates (e.g. spatial, causative) to acquisition for natural language. We develop a theory later form the more abstract mental and cognitive predi- of lexical semantics, the Eztended Aspect Calculus, which cates. In this view thematic relations are actually sets of together with a ~maxkedness theory" for thematic rela- thematic properties, related by a partial ordering. This tions, constrains what a possible word meaning can be. effectively establishes a maxkedness theory for thematic This is based on the supposition that predicates from the roles that a learning system must adhere to in the acqui- perceptual domain axe the primitives for more abstract sition of lexical entries for a larlguage. relations. We then describe an implementation of this We will discuss two computational methods for model, TULLY, which mirrors the stages of lexical acqui- concept development in natural language: sition for children. (1) F~ature Relaxation of particular features of the arguments to a verb. This is performed by a constraint propagation method. I. Introduction (2) Thematic Decoupling of semantically incorporated In this paper we describe a computational model of con- information from the verb. cept acquisition for natural language making use of po- When these two learning techniques are combined with sitive-only data, modelled on a theory of lexical seman- the model of lexical semantics adopted here, the stages tics. This theory, the Eztende~t Aspect Calculus acts to- of development for verb acquisition are similar to those gether with a maxkedness theory for thematic roles to acknowledged for child language acquisition. constrain what a possible word type is, just as a grammar defines what a well-formed tree structure is in syntax. We argue that linguistic specific knowledge and learning principles are needed for concept acquisition from positive 2. Learnabillty Theory and Concept De- evidence alone: Furthermore, this model posits a close in- velopment teraction between the predicates of visual perception and the early semantic interpretation of thematic roles as used Work in machine learning has shown the useful- in linguistic expressions. In fact, we claim that these re- ness to an inductive concept-learning system of inducing lations act as constraints to the development of predicate "bias" in the learning process (cf. [Mitchell 1977, 1978], hierachies in language acquisition. Finally, we describe [Michalski 1983]). An even more promising development TULLY, an implementation of this model in ZETALXSP is the move to base the bias on domain-intensive models, and discuss its design in the context of machine learning as seen in [Mitchell et al. 1985], [Utgoff 1985], and [Win- research. ston et al. 1983 I. This is an important direction for those concerned with natural language acquisition, as it con- There has been little work on the acquisition of verges with a long-held belief of many psychologists and thematic relation and case roles, due to the absence of linguists that domain-specific information is necessary for any consensus on their formal properties. In this research learning (cf. [Slobin 1982], [Pinker 1984], {Bowerman we begin to address what a theory of thematic relations 1974], [Chomsky 1980]). Indeed, Berwick (1984) moves in might look like, using learnabUity theory as a metric for exactly this direction. Berwick describes a model for the evaluating the model. We claim that there is an impor- acquisition of syntactic knowledge based on a restricted tant relationship between visual or imagistic perception X-syntactic parser, a modification of the Marcus parser and the development of thematic relations in linguistic us- ([Marcus 1980]). The domain knowledge specified to the age for a child. This has been argued recently by Jackend- system in this case is a parametric parser and learning off (1983, 1985) and was an assumption in the pioneering system that adapts to a particular linguistic environment, work of Miller and Johnson-Laird (1976). Here we argue given only positive data. This is just the sort of biasing necessary to account for data on syntactic acquisition.

172 One area of language acquisition that has not been How does this principle relate to lexical seman- sufficientlyaddressed within computational models is the tics and the way thematic relations are mapped to syn- acquisition of conceptual structure. For language acquisi- tactic positions? We claim that the connection is very tion, the problem can be stated as follows: How does the direct. Concept learning begins with spatial, temporal, child identify a particular thematic role with a specific and causal predicates being the most salient. This follows grammatical function in the sentence? This is the prob- from our supposition that these are innate structures, or lem of mapping the semantic functions of a proposition are learned very early. Following Miller and Johnson- into specified syntactic positions in a sentence. Laird (1976), [Miller 1985], and most psychologists, we Pinker (1984) makes an interestingsuggestion (due assume the prelinguistic child is already able to discern originally to D. Lebeaux) in answer to this question. He spatial orientations, causation, and temporal dependen- proposes that one of the strategies available to the lan- cies. We take this as a point of departure for our theory guage learner involves a sort of ~template matching" of of markedness, which is developed in the next section. argument to syntactic position. There are canonicalcon- j~gurat{ortsthat are the default mappings and non-cano- 3.0 Theoretical Assumptions nicoJ mappings for the exceptions. For example, the template consists of two rows, one of thematic roles, and the 3.1 The Extended Aspect Calculus other of syntactic positions. A canonical mapping exists In this section we outline the semantic framework if no lines joining the two rows cross. Figure 1 shows a which defines our domain for lexical acquisition. In the canonical mapping representing the sentence in (1), while current linguistic literature on case roles or thematic re- Figure 2 illustratesa noncanonical mapping representing lations, there is littlediscussion on what logical connec- sentence (2). tion exists between one e-role and another. Besides being the workhorse for motivating several principles of syntax (cf. [Chomsky 1981], [Willi~ms 1980]) the most that 0-roles: ~~L is claimed is that Universal Grammar specifies a reper- toire of thematic relations (or case roles), Agent, Theme, Syntactic roles: SUBJ OBJ OBL Patient, Goal, Source, Instrument, and that every NP must carry one and only one role. It should be remem- Figure 1 bered, however, that thematic relations were originally conceived in terms of the argument positions of semantic predicates such as CAUSE and DO. * That is a verb e-roles: A Th G/S/L didn't simply have a list of labelled arguments 2 such as Agent and Patient,but had an interpretation in terms of Syntactic ro~O BL more primitive predicates where the notions Agent and Patient were defined. The causer of an event (following Jackendoff (1976)) is defined as an Agent, for example, Figure 2 c ,4u s E(=, ,) -. Ag,.~(=). Similarly, the first argument position of the pred- (1) Mary hit Bill. icate GO is interpreted as Theme, as in GO(=,y,z). The (2) Bill was hit by Mary. second argument here is the SOURCE and the third is called the GOAL. The model we have in mind acts to constrain the With this principle we can represent the productivity of space of possible word meanings. In this sense it is similar verb forms that are used but not heard by the child. We to Dowty's aspect calculus but goes beyond it in embed- will adopt a modified version of the canonical mapping ding his model within a markedness theory for thematic strategy for our system, and embed it within a theory of how perceptual primitives help derive linguisticconcepts. types. Our model is a first-orderlogic that employs sym- bols acting as special operators over the standard logical As mentioned, one of the motivations for adopt- vocabulary. These are taken from three distinct semantic ing the canonical mapping principle is the power it gives fields. They are: causal, spatial, and aspectual. a learning system in the face of positive-only data. In terms of learnability theory, Berwick (1985) (following The predicates associated with the causal field are [Angluin 1978]) notes that to ensure successful acquisi- Cau~e, (C,), C~se~ (C2), and l.stru,ne.t(I). The spatial field has only one predicate, which is predicated tion of the language after a finite number of positive ex- Locatiue, of an object we term the Th~me. Finally, the aspectual amples, something llke the Subset Principle is necessary. We can compare this principle to a Version Space model of inductive learning( [Mitchell 1977, 1978]), with no neg- i CfiJackendoff(1972, 1976) for a detailed elaboration of ative instances. Generalization proceeds in a conservative this theory. fashion, taking only the narrowest concept that covers the data. 2 This is now roughly the common assumption in GB, GPSG, and LFG.

173 field has three predicates, representing the three temporal culmination to the process of giving. This is captured by intervals t~, beginning, t2, middle, and t3, end. From the reference to the final subinterval, is. The linking between interaction of these predicates all thematic types can be = and the L associated with tt is interpreted as Source, derived. We call the lexicalspecification for this aspectual while the other linked arguments, y and z are Theme (the and thematic information the Thematic Mapping Indez. book) and Goa/, respectively. Furthermore, = is specified As an example of how these components work to- as a Causer and the object which is marked Theme is also gether to define a thematic type, consider first the dis- an affected object (i.e. Patient). This will be one of the tinction between a state, an activity (or process), and an TMIs for an accomplishment. accomplishment. A state can be thought of as reference In these examples the three subsystems are shown to an unbounded interval, which we will simply call t2; as rows, and the configuration given is lexicallyspecified. 4 that is, the state spans this interval.3 An activity or pro- tess can be thought of as referring to a designated initial 3.2 A Markedness Theory for Thematic Roles point and the ensuing process; in other words, the situation spans the two intervals tt and t2. Finally, an event As mentioned above, the theory we are outlining can be viewed as referring to both an activity and a des- here is grounded on the supposition that all relations in ignated terminating interval; that is, the event spans all the language are suffiently described in terms of causal, three intervals,it, t2, and is, spatial and aspectual predicates. A thematic role in this Now consider how these bindings interact with the view is seen as a set of primitive properties relating to the other semantic fieldsfor the verb run in sentence (8) and predicates mentioned above. The relationship between give in sentence (9). these thematic roles is a partial ordering over the sets of (8) John ran yesterday. properties defining them. It is this partial ordering that (9) John gave the book to Mary. allows us to define a markedness theory for thematic roles. We associate with the verb run an argument structure of Why is this important? simply rim(=}. For give we associate the argument struc- If thematic roles are assigned randomly to a verb, ture ~v,(=,v, =). The Thematic Mapping Index for each is then one would expect that there exist verbs that have given below in (10) and (11). only Patient or Instrument, or two Agents or Themes, for 00) example. Yet this is not what we find. What appears to be the case is that thematic roles are not assigned to a verb independently of one another, but rather that some thematic roles are fixed only after other roles have been established. For example, a verb will not be assigned a GOAL if there is not a THEME assigned first. Similarly, L/!, a LOCATIVE is dependent on there being a THEME (11) present. This dependency can be viewed as an acquisition strategy for learning the thematic relations of a verb. Now let us outline the theory. We begin by estab- lishing the most unmarked relation that an argument can Th bear to its predicate. Let us call this role Them,~. The t ,!) only semantic information this carries is that of an exis- tt t2 tential quantifier. It is the only named role outside of the three interpretive systems defined above. Normally, we The sentence in (8) represents a process with no logical culmination, and the one argument is linked to the named think of Them, as an object in motion. This is only half case role, Theme. The entire process is associated with correct, however, since statives carry a Theme readings as both the initialinterval t~ and the middle interval t2. The well. It is in fact the feature [±motion] that distinguishes argument = is linked to C~ as well, indicating that it is the role of Mary in (1) and (2) below. an Actor as well as a moving object (i.e. Theme). This represents one TMI for an activity verb. (1) Stative: l-motion I Mary sleeps. The structure in (9) specifies that the meaning of give carries with it the supposition that there is a logical (2) Active: [+motion] Mary fell. This is a simplication of our model, but for our This gives us our first markedness convention: purposes the difference is moot. A state is actually interpreted as a primitive homogeneous event-sequence, with downward closure. Cf. [Pustejovsky, 1987], (3) Therr=ee--Theme.~/[+motion]

4 [Jacl~endofftOSS] develops a similaridea, but vide in/ra (3) Themery-..Themes/[-motior=] for discussion.

174 where ThemeA is an "activity" Theme, and Themes is a (8) Agent{+CompleU] <~ Agent[-CompleteJ stative. Within the spatial subsystem, there is one variable These changes define a new concept, "effector', which is type, Location, and a finite set of them L1, L~... L~. The a superset of the previous concepts given in the system. most unmarked location is that carrying no specific aspec- The same can be done with C'~ to arrive at the concept of tual binding. That is, the named variables are Ls and Lz an "effected object." We see the difference in interpreta- and are commonly referred to as Source and Goal. Thus, tion in the sentences below. Lu is the unmarked role. The limitations on named locative variables is perhaps constrained only by the aspectual a. John intentionally broke the chair. (Agent-direct) system of the language (rich aspectual distinction, then more named locative variables). The markedness conven- b. John accidentally broke that chair when he sat tions here are: down. (Agent-indirect)

(4) Lu -* S/B c. John broke the chair when he fell. (Effector)

(s) L~ -- C/E Given the manner in which the features of primitive thematic roles are able to change their values, we are Within the causal subsystem there are three predicates, Cl, C2, and I. We call C2, (the traditional Patient defining a predictable generalization path that relations role) is less marked than c~, but is more marked than I. incorporating these roles will take. In other words, two These conventions give us the core of the primitive concepts may be related thematically, but may have very semantic relations. To be able to perform predicate gen- different extensional properties. For example, give and eralization over each relation, however, we define a set of take are clearly definable perceptual transfer relations. features that applies to each argument within the seman- But given the abstractions available from our marked- tic subsystems. These are the abstraction operators that ness theory, they are thematically related to something allow a perceptual-based semantics to generalize to non- as distant as "experiencer verbs", e.g. please, as in "The perceptual relations. These features also have marked book pleased John." This relation is a transfer verb with an incorporated Theme; and unmarked values, as we will show below. There are namely, the "pleasure." s four features that contribute to the generalization process If we apply these features in the spatial subsystem, in concept acquisition: we can arrive at generalized notions of location, as well as abstracted interpretations for Theme, Goal and Source.

(a) l±~b,tra,t] (b) [+d~r,~t] For example, given the thematic role Th - A with the feature [-Abstract] in the default setting, we can generalize (c) [±,o,.pl,t,] (d) [±.~i~t,] to allow for abstract relations such as like, where the object is not affected, but is an abstract Theme. Similarly, The first feature, abttract, distinguishes tangible the Theme in a sentence such as (a) can be concrete and objects from intangible ones. Direct will allow a gradi- direct, or abstract, as in (b). ence in the notion of causation and motion. The third feature, cornplete,picks out the extension of an argument (a) have(L, rh) Mary has a book. as either an entire object or only part of it. Ani~v~ac~lhas the standard semantics of labeling an object as alive or (b) have(L, Yh) Mary has a problem with Bill. not. Let us illustrate how these operators abstract over In conclusion, we can give the following dependencies be- primitive thematic roles. By changing the value of a fea- tween thematic roles: ture, we can alter the description, and hence, the set of objects in its extension. Assume, for example, that the predicate C1 has as its unmarked value, [+Direct]. {r~eme}

(6) C,[UDir,,tl --[+Vir,ctl {~} {s, c} By changing the value of this feature we allow CI, the direct agent of an event, to refer to an indirect causer. {c,} (7) Ae,.t[+D~rect I <@ Aee,~tl-Dir,ct ]

Similarly, we can change the value of the default setting for the feature I+Complet~] to refer to a subcausation (or s Cf. Pustejovsky (1987) for an explanation of this term causation by part). and a full discussion of the extended aspect calculus.

175 4.2 The Acquisition Procedure The generaliztion features apply to this structure to build We now turn to the design of the learning program hierarchicalstructures (Cf. {Keil 1979], [Kodratoff 1986]). itself. TULLY can be characterized as a domain-intensive This partial ordering allows us to define a notion of cov- inductive learning system, where the generalizations pos- crs'ng, as with a semi-lattice,from which a strong princi- sible in the system are restricted by the architecture im- ple of functional uniqueness is derivable (of. [Jackendoff posed by the semantic model. We can separate clearly 1985]). The mapping of a thematic role to an argument what is given from what is learned in the system, as shown follows the following principle: in Figure 1.

(9) Maximal Assignment Principle An argument GIVEN will receive the maximal interpretation consistent Extended Aspect Calculus with the data. 0-Markedness Theory Canonical Mapping This says two things. First, it says that an Agent, for Rule Execution Loop example, will always have a location and theme role associated with it. Furthermore, an Agent may be affected by its action, and hence be a Patient as well. Secondly, ACQUIRED this principle says that although an argument may bear Verbal Lexical semantics many thematic roles, the grammar picks out that function Argument-function mapping which is mazimall!; specific in its interpretation, accord- Predication Hierarchy ing to the markedness theory. Thus, the two arguments might be Themes in "John chased Mary", but the the- Figure 1 matic roles which maximally characterize their functions In order to better understand the learning mecha- in the sentence are A and P, respectively. nism, we will step through an example run of the system. First, however, we will give the rule execution loop which 4. The Learning Component the system follows. 4.1 The Form of the Input Rule Execution Loop The input is a data structure pair; an event sequence expression and a sentence describing the event. 1. Instantiate Existing Thematic Indexes The event-sequence is a simulated output from a middle- INSTANTIATE: Attempt to do a semantic analy- level vision system where motion detection from the low- sis of word given using existing Thematic Mapping level input has already been associated with particular Indexes. If the analysis fails then go to 2. object types. 6 The event-sequence consists of three instantaneous 2. Concept.acquisition phase. descriptions (IDa) of a situation represented as intervals. Note failure: Credit assignment. These correspond to the intervals t~, t2, and ts in the Link arguments to roles according to Canonical aspect calculus. The predicates are perceptual primi- Mapping. tives, such as those described in Miller and Johnson- 3. Build new Thematic Mapping Index Laird (1976) and Maddox and Pustejovsky (1987), such LINK and SHIFT: Constructs new index accord- as [Ar(t~, ~) ~ ~ = [O,V(,,, d t, ,4,,,,,.~t,(,,) ~, Mo,,,~(~,) ~, ...]]. The second object is a linguistic expression (i.e. a sen- ing to the Extended Aspect Calculus using infor- tence), parsed by a simple finite state transducer. ~ mation from credit assignment in (2). If this fails then go to (4).

4. Invoke Noncanonical Mapping Principle. If (3) failsto build a mapping for the lexical item in the input, then the rule INTERSECT is invoked. This allows the lines to cross from any of the interpretive levels to the argument tier.

s For a detailed discussion of how the visual processing 5. Generalization Step. and linguisticsystems interact,cf. Maddox and Pustejovsky This is where the markedness theory is invoked. (1987). Induction follows the restrictions in the theory, where generalization is limited to one of the stated We are not addressing any complex interaction between types. syntactic and semantic acquisition in this system. Ideally, we would like to integrate the concept acquisition mechanisms here with a parser such as Berwick's, Cf. Berwick 1985.

176 ture relaxation on the Theme, together with propagated Assume that the first input to the system is the constraints to the Source and Goal (the subject and ob- sentence ~Mary hit the cat," with its accompanying event ject, respectively); the difference is that the Theme is sequence expression, represented as a situation calculus incorporated said is not grammatically expressed. expression. INSTANTIATE attempts to map an exist- John pleased his mother. ing Thematic Mapping [ndez onto the input, but fails. please(z ~ ~, y ffi G, Th : incorporated) Stage (2) is entered by the failure of (1), and credit assignment indicates where it failed. Heuristics will indicate which thematic properties are associated with each argu- Conclusions ment, and stage (3) links the arguments with the proper roles, according to Canonical Mapping. This links Mary In this paper we have outlined a theory of acquisi- to Agent and the cat to Patient. tion for the semantic roles associated with verbs. Specifi- One important point to make here is that any cally, we argue that perceptual predicates form the foun- information from the perceptual expression that is not dation for later conceptual development in language, and grammatically expressed will automatically be assumed propose a specific algorithm for learning employing a the- to be part of the verb meaning itself. In this case, the ory of markedness for thematic types and the two strate- instrument of the hitting (e.g. Mary's arm) is covered by gies of thematic decoupling and constraint relazation and the lexical semantics of hit. propagation. The approach sketched above will doubtless There are two forms of generalization performed need revision and refinement on particular points, but is by the system in step (5): constraint propagation and claimed to offer a new perspective which can contribute to thematic decoupling. In a propagation procedure (Cf. the solution of some long-standing puzzles in acquisition. [Waltz, 1975]), the computation is described as operat- ing locall!/, since the change has local consistency. To Acknowledgements illustrate, consider the verb entry for have, as in (1), (I) John has a book. have(z =/;, y = Th) I would like to thank Sabine Bergler who did the where the object carries the feature [-abstract].Now, con- first implementation of the algorithm, as well as Anthony sider how the sense of the verb changes with a feature Maddox, John Brolio, Ken Wexler, Mellissa Bowermxn, change to [~abetract], as in (2). and Edwin Williams for useful discussion. All faults and (2) John has an idea. errors are of course my own. In other words, there is a propagation of this feature to the subject, where the sense of locative becomes more abstract, e.g. menta/. These types of extensions give rise to other verbs with the same thematic mapping, but with References ~relaxed" interpretations. * [I] Angluin, D. "Inductive Inference of formal Lan- The other strategy employed here is that of the- guages from positive data." In[ormation and Con- matic decoupling, where thematic information becomes trol 45:117-135. disassociated from the lexical semantics for a verb. ' [2] Berwick, Robert C. The Acquisition of Syntactic The narrower interpretation of a verb's meaning will be Information, MIT Press, Cambridge, MA. 1985. arrived at after enough training instances are given; for example, from cut as meaning a particular action with a [3] Berwick, Robert C., "Learning from Positive-Only knife, to cut as an action that results in a certain state. Examples: The Subset Principle and Three Case It is interesting to speculate on how these strate- Studies," in Michalski et al, 1986. gies facilitatethe development from perceptual relations [4] Bowerman, Mellissa "Learning the Structure of Cau to more abstract ones. The verb tell,for example, can be satire Verbs," in Clark (ed) Papers and reports on viewed as a transfer verb with a [+abstract]Theme, and the child language development, No. 8, Stanford Uni- accompanying contraint propagation (Cf. [Pinker, 1984] versity Committee on Linguistics. 1974 and [Jackendoff, 1983]). Similarly,experiencer verbs such [5] Chomsky, Noam Rules and Representation, Colum- as please, upset, and anger can be seen as combining both bia University Press, 1980 strategies: they are similar to transferverbs, but with lea- [6] Chomsky, Noam Lectures on Government and Bind- ing, Foris, Holland, 1981. s For further discussion of constraint propagation as [7] Dowry, David R., Word Meaning and Montague a learning strategy, cf. Pustejovsky (1987b). Grammar, D. Reidel, Dordrecht, Holland, 1979. [8] Jackendoff, Ray, Language and Cognition, MIT 9 Results given in Nygren (1977) indicate that chil- Press, Cambridge, MA. 1983. dren have fully incorporated instruments for verbs such as hammer, cut, and saw, and only at a later.age do they [9] Jackendoff, Ray, ~The Role of Thematic Relations abstract to a verb sense without a particular and constant in Linguistic Theory,", ms. Brandeis University, instrument interpretation. 1985

177 [I0] Kodratoff, Yves, and J-G. Ganascia, "Improving [26] Winston, Patrick H., "Learning by Augmenting the Generalization Step in Learning", in Michal- Rules and Accumulating Censors," in Michalski et skiet el (eds.), Machine Learning II, Morgan Kauf- al, 1986. mann, [27] Winston, Patrick, Binford, Katz, and Lowry, "Learn [11] Marcus, Mltch, A Theory of Syntactic Recogni- ing Physical Descriptions from Functional Defini- tion for Natural Language, MIT Press, Cambridge, tions, Examples, and Precedents, Proceedings of 1980 AAAI, Washington, 1983 [12] Michalski, R.S., "A Theory and Methodology of Inductive Learning,", in Michalski et al (eds.), Ma- chins Learning L [13] Miller, George, "Dictionaries of the Mind" in Pro- ceedings of the 23rd Annual Meeting of the As- sociation for Computational Linguistics, Chicago, 1985. [14] Miller, George and Philip Johnson-Laird, Language and Perception, Belknap, Harvard University Press, Cambridge, MA. 1976. [15] Mitchell, Tom, "Version Spaces: A Candidate Elim- ination Approach to Rule Learning," in IJCAI-77, 1977 [16] Mitchell, Tom, Version Spaces: An Approach to Concept Learning, Ph.D. thesis Stanford, 1978. [17] Nygren, Carolyn, "Results of Experiments with In- strumentals," ms. UMASS, Amherst, MA. [18] Pilato, Samuel F. and Robert C. Berwick, "Re- versible Automata and Induction of the English Auxiliary System", in Proceedings of the 23rd An- num Meeting of the Association for Computational Linguistics, Chicago, 1985. [19] Pinker, Steven, Lan#uage Learnability and Lan- guage D~velopmcnt, Harvard University Press, Cam bridge, 1984 [20] Pustejovsky, James, "A Theory of Lexical Seman- tics for Concept Acqusition in Natural Language", to appear in /n~ernatioaa/Journal of Intelligent Systems [21] Pustejovsky, James and Sabine Bergler, "On the Acquisition of the Conceptual Lexicon", paper sub- mitted to AAAI-1987, Seattle, WA. [22] Slobin , D. "Universals and Particulars in Lan- guage Acqusition", in Gleitmann, Language Ac- quisition, Cambridge, 1982 [23] Waltz, David "Understanding line drawings of sce- nces with shadows," in The Psychology of Com- puter Vision, P. Winston ed. New York, McGraw- Hill, pp. 19-92. [24] Waltz, David "Event Space Descriptions," Pro- ceedings of the AAAI-82, 1982 [25] Williams, Edwin, "Predication", Linguistic Inquiry, 1980

178