File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/79/j79-1056_metho.xml
Size: 86,731 bytes
Last Modified: 2025-10-06 14:11:16
<?xml version="1.0" standalone="yes"?> <Paper uid="J79-1056"> <Title>American Journal of Computational Linguistics Microf ichc 56 PROCESSING CASE</Title> <Section position="1" start_page="0" end_page="0" type="metho"> <SectionTitle> PROCESSING CASE YORICK WILKS PRESENT ADDRESS: DEPARTMENT OF LANGUAGE AND LINGUISTICS, UNIVERSITY OF ESSEX </SectionTitle> <Paragraph position="0"/> </Section> <Section position="2" start_page="0" end_page="0" type="metho"> <SectionTitle> ABSTRACT </SectionTitle> <Paragraph position="0"> This paper has three purposes: firstly, to describe how case information IS distributed in the preference semantics system of language understahding, and tc show what practical use is made of that informatloq. Second] y, to argue that that way of doing thlngs has adv,~ntages over tao alternatives: (a) putting all case information In one place, and (b) not using any case information at all, but only t.he names of English prepositidns. Thirdly, I wish to- use the positions established earl ler counter some recent arguments by and others that the noti~n of case is not in fact functioning in any natural language understanding systems that fall within what could be called the Artificial Intelligence paradigm. A theme that recurs in the paper is that tendentlous distlb1ctions, such as surface&quot;, &quot;deep&quot; and &quot;conceptual&quot; case, must be expounded in processing terms if tbey are to make sense.</Paragraph> <Paragraph position="1"> The paper owes a great deal to discussions with Eugene Chanliak, Graham Ritcshie, Marg~ret Klng and Frederick Parker-Rhodes. The mistakes, as usual, are a11 my own6</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> TABLE OF CONTENFS </SectionTitle> <Paragraph position="0"> Introduction ...................... 4 .............. Case i.n preference semanti.cs 10 . . . . Case in formuias - 10 Case in paraplates ................ 32 ....... .... Case extraction i.nferences 47 Charniak'sbriefoncase ................ 58 References ....................... 67</Paragraph> </Section> <Section position="4" start_page="0" end_page="9" type="metho"> <SectionTitle> INTRODUCTION </SectionTitle> <Paragraph position="0"> This paper is Intended to describe the case handling procedures of the preference semantics (Wilks 1972, 1975a)system of natural language understanding by giving a more complete account than In prevlous papers, and In particular I shall distinguish the application of case to the parsing of preposition structures In Engl~sh from subseqyent inferences using case, Case information in thls system IS stored in two dlffercnt places: in what are called formulas and paraplates respectively. I shall argue a reasonable*processing account of case requires this. I shall wntrast this positlon briew with those of Schank and Riesbeck, who seem to me to advocate a single type of case information, and a no case view respectively. More importantly, Ihshall argue against a recent position of Charniak that Artificial Intelligence (AI) natural language systems do not in fact make any use d case, I shall dlscuss his arguments and urge that, although there are systems to whlch his arguments do apply, they do not apply to the one described here, at least hot if case is to have anyth~ng like its normal meaning, The use of case informat~on in A1 comes from the work of F~llmore (1968), in which the underlying structur,e of a sentence is displayed, in essence, as an array of argument values far a predicate, where the -redi,:ate is the verb of tk sentenceo The corresponding values are the case parts of the sentence, each of a different case type, and, for any given verb, the general pattern of cases it takes is called the case frame of that verb.</Paragraph> <Paragraph position="1"> Thus, if the horizontal lines denote the verb predicate, the case frame for &quot;break&quot; could be written as:</Paragraph> <Paragraph position="3"> which means that this verb must take an object, and cen but need not take an agent and instrument case. These latter two cases are optional here and that optionallty is indicated by the parentheses round the case names. So, on this view: (1) John broke would be ill-formed because it gives us only the (optional) agent of breaking, which is John, but omits the obligatory object that is broken. Thls analysis can be contrasted with &quot;hit&quot;, whose frame would be</Paragraph> <Paragraph position="5"> where the overlapping brackets mean that the two cases,agent and instrumental are semi-optional, in that at least one of the two must appear. Thus The window hit would be ill-formed on that view because, although it contains the obligatory object, it contains neither the agent nor the instrument of the hitting, However, (3) The window broke would be perfectly well-formed with respect to the frame for 'break', This sort of case analysis 1s normally zalled deep case *to distinguish it from the surface case manifestation of word inflections in Greek, Lati'n and German etc. It should be noted, too, that here, as throughout the paper, no attempt is made to stick to the actual list of case names used by any author, since these vary sd much from one to the next, and no point of principle hangs on arYy particular list of cases, ~illrnore' s linguls tic theory is, naturally enough, a generative one in that the procedures it suggests would be those for generating sentences from an underlying structure of a verb plus its case argument values. What the surface form would be, glven any partlc~lar underlying structure, is determined By what Fillmore calls the &quot;subject selection rule&quot;, which says that, if there is an agent, that will be the subject of any active sentence produded; if there is no agent but there is an instrument, then that will be the subject and so on.</Paragraph> <Paragraph position="6"> In this paper, I Ishall concentrate, as is normal if not desirable in A1 and computational linguistics, Qn questions of analys~s rather than generation. The general problen in analysis (that is not emphasised in generation) is that of the selection restrictions on the cases, which can be illustrated by looking at the simplest of the systems using case analysis (Simmons 1973). Simmons wishes to map all of the sentences (4) (7) onto the same semantic network because a11 four, in some sense, refer to the szme event: (4) John broke the window with a hammer (5) John broke the window (6) The hammer broke the window (7) The window broke All are perfectly well-formed with respect to the franc for &quot;break1' given earlier. Simmons parses such sentences using an augmented transition network (Woodst1970) and a notion of case paradigm due to Celce-Murcia (1972). This paradigm, for active forms of a verb llke &quot;break&quot;, has the form:</Paragraph> </Section> <Section position="5" start_page="9" end_page="11" type="metho"> <SectionTitle> (8) AGENT * OBJECT INSTRUMENT AGENI * CBJECT INSTRUMEIJT * OBJECT OBJECT A </SectionTitle> <Paragraph position="0"> The lines of (8) are patterns that must match input word strings in left right order so as to assign the cases they contain, The lines of (8) match each of (4) - (7) in turn, where marks the posltion of the verb (break) in each line of the paradigm, The lines of (8) are no more than the possible case combinations allowed by the case frame for &quot;break&quot; together with an analytic version of the subject selection rule&quot;, which always makes the Agent the first (subject) item In any line of the paradigm (8) in which it occurs, There is no need for the lines of (8) to be ordered In their application to input sentences, although there is one additional. item of information required before they can be applied at all: the selection restrictions, These tell us what it is to be an agent of &quot;break&quot;: in Simons's scheme a noun marked ANIMATE, The selection restrictions attached to the cases in (8) are essential t~ the application of the paradlgrn, for only thus could we know that &quot;~ohn&quot; in (4) was matched by ACZENT in the flrst line of (8). It should be noted that the restriction of AGENT in (8) to nouns m%rked ANIMATE is not necessarily a restriction peculiar to 'break', but rather to the class of verbs for which (8) is the (active) paradigm. Conversely, the ANIMATE restriction on AGEPU'Ts in (8) IS not necessarily on AGENTS as such although it might turn out to be so, These points will be important when we come to Chamiak's arguments later.</Paragraph> <Paragraph position="1"> Notice too, that there are not two different ways In which a sentence can be ill-formed with respect to the paradigm: one with respect to selection restrictions and one with respect to the case frame (as has been argued by Bruce 1975). The case frame expressed by the paradigm, and the corresponding selection restrictions are indivisible. So, for example, (9) John broke is ill-formed with respect to the case frame for &quot;break&quot; as explained earlier, However, if we look at that fact in procedural /terms, such as those provided by Simmons' paradigm, we cannot deem that failure as one of matching a line of (8) as distinct from (Bruce's view) not meeting the select~on restr~ctions PHYSOB , say, on the case OBJECT on the fourth line of (8). For the selection restriction associated with a case defines what ~t is to match a corresponding line of (8).</Paragraph> <Paragraph position="2"> The only way in which a sentence could independently fall to match any line in the paradigm of (8) would be the trivial one of having some number of arguments (say, four or zero) not corresponding to any line of (8),</Paragraph> </Section> <Section position="6" start_page="11" end_page="11" type="metho"> <SectionTitle> CASE IN PREFERENCE SEMANT?CS </SectionTitle> <Paragraph position="0"> Case in formulas This system builds meaning structures and inference rules from eighty primitive semantic elements, These are of eight types, one of which consists of the case elements as follows; *DIRE the general DIRECTION ca'se element. Like all the primitive elements whose names are preceded by an asterisk, it is equivalent to a class of other primitives, in this case the following four: TO direction towards FROM direction away from something UP in an upwards direction</Paragraph> </Section> <Section position="7" start_page="11" end_page="12" type="metho"> <SectionTitle> SUBJ the AGENT case,. indicating the ins ti gator of an action, </SectionTitle> <Paragraph position="0"> 'subject' here being taken to refer to a semantic, rather than a surface, subject WITH the ACCOMPANIMENT case, indicating the acco~npanier of an entity POSS the POSSESSIVE case, indicating who owns some thlng These case primitives are ultimately the names of relations in the system of semantic representstion, just as in the familiar semantlc net representations (Simmons 1973) that indicate the instrumentality of say, the action of striking by a labelled arc such as: strike 'hammer The representations described here are not of this superficial form for three reasons: a) Semantic nets do not immediately suggest their associated processes, whereas the representations here are intended to be directed towards the processes that operate on them.</Paragraph> <Paragraph position="1"> b) There is a clear distinction in the present system of represent- null ation between the knowledge stored and the patterns sought in language, on the one hand, and the language text actually represented found, on the other; where the latter may not exactly match what was being looked for, This distinction 1s not always easy to work into a semantic net structure.</Paragraph> <Paragraph position="2"> c) The present system of representation is intended to be more &quot;habitable&quot; in Watt's (1968) sense of providing a language of semantic representation that is appropriate to the way humans express themselves. That is a highly subjective notion, perhaps, but here it is taken to require at least a dynamrc, or readable structure, which nets do not haverThj srequirement rests upon another assuniption: that our representation must have the &quot;one thing after another&quot; feature that text,$ have, rather than being static and timeless like most semantic nets (thoughdNorman and Rumelhart (1975) hdve constructed nets containing ordered assertions, though these are no longer semantic nets in the classic sense).</Paragraph> <Paragraph position="3"> The case primitives function within a semantic dependency grammar (Hays ?964), intended to express the meaning of word senses and, by extension, of texts. Each of the casg primitives above will have a dependent, which is a type of entity for all the case prirfiitives except WAY and GOAL, which take an assertion as dependent, The case primitive and its dependent (entity or assertion) fom a case group which is in turn dependent on a primitive action (except for WITH and POSS which depended on an entity , and may therefore .be only semi-cases).</Paragraph> <Paragraph position="4"> This is best seen by example of the first structure in the system, the formula - which expresses word sense in the dictionary, The formula for the action sense of &quot;break&quot; is as follows: The general structure of such formulas has been explained In Wifks (1968, 1972, 1975a, 1975b). They are intended to express the fnterlingual meaning of the sense of the word, and the primitives that comprise them are intehded to be interlingual (as are Fillmore's cases) even though they happen to be mostly Anglo-Saxon monosyllables. 'E'ormulas are trees of left-rignt dependencies but the dependence is interpreted differently acccrding to the type of the subformula. We have already mentioned the dependence of an entity (or assertion) on a case primrtive to form a case group, such as (*'tIb% SUBJ), which means that an agent is (preferably) human. These case groups (except WITH and POSS) a11 depend rightwards on some action, so that (*HUM SUN) in (10) depends on the main primitive actioh of the whole formula STRTK, The whole formula is to be interpreted as &quot;breaking&quot; being a ~'k~~~ing, done preferably to a *PHYSOSJect, and by a *HUMan SUBJect, using an INSTrument that is a THING and with the GOAL of CAUSing the *PINSOBJect to BE NOTWHOLE.</Paragraph> <Paragraph position="5"> This interpre~ation can be constructed from the following _generaL rules for the building and interpretation of formulas: i) Each subgroup in the formulacons~sts of a left mernber depending on a right member, and left or right may be either a single prim~tive element or another group, Thus, in (*EIUM SUBJ) we have a case group, known to be such because the rightmost member of its pair is the governor and SUBJ is the primitive element naming the Agent caqe. One level higher (*HUM SUBJ) . depends on STRIK, the main primitive of the whole formula, to form an assertion group. Similarly, each of the other maln subparts of the formula (whose heads are respectively OBJE, INST an6 GOAL) depends on STRIK to form in each case an action group which always consists of an action and any case group that is not an Agent group (nor a POSS or WITH group), ii) The dependency within a group 1s interpreted differently according to the type of the group.</Paragraph> <Paragraph position="6"> Within an acti~n graup ((*PKYSOB OBJE) STRIK) the dependepce is that of an action's object on the action and OWE does no more than name that relation. In the cases group (*PHYGOB OBE) the relation internally is no sore than the preferred type of case fLller Cphysical object) on the name of the case. With a substantive group like (LINE THIYG) the dependence is interpreted as specification, i.e. linear object. *PHYSOB is a name ~f a class of primitive elements which includes THING, but also other primitives like MAN. In case subfarniulas, e&c@pt afid WAY, the leftmost item is always the preferred entlty type, to fuyctlon in the corresponding rightwards-nwed case. This lef trnost item 19, if you will, the 'selection restriction' for that case role for whatever action is being coded: i. e. in the formula above, for 'break'. The reader should not confuse this with being a restriction for the associated primitive STRIK, This point will be discussed later, but for now the formula is to be taken as no more than a formal expression of the meaning of the action 'break' that can be used in subsequent It inference and parsing routines, However, selection restriction&quot; here is to be read as Ipreferring the agent of &quot;break&quot; to be hurnaa1,say, I have described elsewhere (Wilks 197 c) how when text representations are assembled that will be only a preference on the agent of &quot;break&quot;, and the system will not baulk at assembling a representation for &quot;The dog broke his bowl&quot; where the agent is not human, though the system would prefer a hwnan agent if it could find one. I write of &quot;assembling representationsu because the elements like *Hull in the formula above are not slots to be filled by, in this case, the agent of some breaking, The formulas are 'blueprints' for how representations are to b. assembled elsewhere from whole formulas. When a representstion for &quot;The man broke .the window&quot; is assembled the whole formula above (together with a PAST element) will stand at some apptopriate node of a higher-level representation.</Paragraph> <Paragraph position="7"> It is this feature of the system that explains why the head, or principal, element of a formula is easily accessible at one end QE it --rather than buried in the center as it would be if the formula w2reL in SVO form, rather than SOV form, as above, Formulas exist for all parts of speech so, for ex@mple, a formula for an entity will have not a primitive action head like STRIK, but an entity head like THING or MAN or STUFF (for substance). Note too, that the preference restrictions for case need not be simple as above but can be as complex as required, including further case restrictions recursively. So, for example, if we had a formula i3r 'sew' it might well have a case sub-fomrrula to be interpreted as: done with an INSTrument that is preferably a LINEar THING WITH (accompaniment case) an aperture (THRU PART).</Paragraph> <Paragraph position="8"> iii) In order to makc the formula &quot;habitablet' the agents and objects are compressed, in that they can be agknts and objects for more than one primitive action, Agents and objects of actions in a formula are normally sought to the left of the primitive action element.</Paragraph> <Paragraph position="9"> If the whole formula is for an action (as above for 'break') the two leSStmost subparts of the formula will always be the preferred agent and object of the head primitive, in that order. For any actions within the formula (such as CAUSE in the formula for 'break') its preferred agent and object are normally the next agent and object to its left -- which of course, as in the case of 'break' may turn out -to yield the same entity as the preferred agent of the whole formula, though this need not be the case. Moreover, in the case of actions within a formula (i.e. not constituting the head) the agent need not be marked though the object must be if it is an entity type. This proviso does not apply In the formula above since the agent is the same for CAUSE and STRIK, and CAUSE takes an assertion as object, but within a formula a group (MAN STRIK) would always be interpreted as an group, MAN being an unmarked agent of STRIK, and not as a man being struck which would require a marked object in the action group i , e, ( (MAN OBJE) STRIK) .</Paragraph> <Paragraph position="10"> All this implies that some of the subgroups in the formula for break' are not the apparent ones i.e, the dependent of GOAL, as mentioned earlier, must be an assertion, whereas it is brncketted to only (((NOTWHOLE K1ND)BE)CAUSE) which can only (during inference procedures,called 'extraction' to be described later) become an assertion group by the addition of an agent found to the left namely (*HUM SUBJ), CAUSE also requires a dependent object that is an assertian (hence (*PH'ISOB OBJE) will not do as its abject 'taken alone) and can take, a,s dependent of that group, an entity to its left marked either OBJE or SUBJ kfnichever is closest, Hence the dependent of ((NOTWHOLE KIND)BE) is *PHYSOB and the &quot;'real&quot; dependent of CAUSE (found by inference) is (*PHYSOB ( (NOTWHOLE KIND) BE) ) and the real dependent of GOAL is ( (*HUM SUBJ) (*PHYSOB ( (NOTWHOLE KIND) BE) ) )CAUSE) . This compresai& of expression can be argued to be &quot;habitable&quot; for a formula maker, It also avoids to a large extent the defect of some fuller conceptual representations of this general type, pointed out by Sandewall (1972), that if the entities like (*HbM SUBJ) are put into the representation many times but are inteqded to refer to THE SAME IlUMAN, then this must be indicated as it frequently was not, Where such identity must be specific in formulas,but cannot be achieved by the above compressed expressions, it is obtained by means of the primitives SAME and NOTSAME: the same (or not), that is to say, as the first encountered token of the associated primitive when working in from the top level of the formula, One point that should emerse from this is that those.who want to use case names as the names of relations, as in semantic nets, and also deal with surface language, must be prepared td extract a number of such relations from a single occurrence of certain formula subparts, Thus, the formula for 'break' ab~ve would contain not only the net links: which asserts that the preferred agent of breaking will be human.</Paragraph> <Paragraph position="11"> However the top link must not be interpreted as saying thzt the preferred agent of the primitive STRIK is human, because that is not an assertion in the system at all. All the top link can say is that the primitive action STRIK sometines takes human agents. In soae other formula, for another surface action whose underlying primitive was also STRIK, the preferred agent might be &quot;ANI, a wider class. Hence, in this system there are not specific semantic restrictions on. the dependents of the primitives, as in, say Schank (1973), In thp present system, such a restriction could emerge only inductively from a survey of a considerable body of formulas, It is worth clarifying this issue here: what do underlying representations constitute case frames for? The issue is related to two others: first null , the different roles of case frames in analysis and gen*ratian, and secot~dfy, the procedural opposition between case frame blueprints, like formulas3 and the 'fuller*' representations of generative sunantic> null ists trees and Schankian ~conceptua~izations'.</Paragraph> <Paragraph position="12"> First , let us note that it has neyer been as clear as mrght be wished what case frames awe for in Fillmore's work, The normal introductory account given earlier states that they are for surface verbs like 'break', but in Fillmore (1975) he argues that it is not so simple because he would want to admit sentences like: (12) Noon found Harry sleeping as perfectly well-formed, while not taking account of the usgge in the %entive part of the case frame for 'find'. On the other hand, he wishes to avoid the position of saying that case frames are for underlying verbs like semantic primitives in formulas, or the underlying verbs of Generative Semanti~srepresentations which avoid cases altogether, as in the following for &quot;I broke the glass with a rock&quot;. I have followed Fillmore's (1975j device here of making (13) easier to read by putting ~t in SVO rather than the usual VSO (predicate first) form, It will be seen that it is pretty similar to the above fnmula for &quot;break1' except that, in order to avoid case notation, they have had to resort to such philosophically suspect devices as separating the act of using from the basic 'act' inside the tree, even though there was really only one action in the whole business An extreme version of the view that case frames belong only to the underlying structure is Schank's (1973) view that case frames awe for underlying primitive acts and that all cases that a primitive act takes, it takes obligatorily.</Paragraph> <Paragraph position="13"> Thus, for example, Schank's primitive act TRANS expresses the underlying content of such actions as &quot;buy&quot;, &quot;sell&quot; and &quot;take&quot;, and .ke would begin the representation of (14) The man took a book as (Schank 1973, p. 196) -to. man</Paragraph> <Paragraph position="15"> where the arrows labelled R and 0 indicate Recipient and Objective ..- case respectively, and the - Agentive case is in fact indicated by the double arrow linking man&quot;, the agent, to the act TRANS. The details here need not concern us, the point being that Schank is setting up case frames, not for surface verbs of English, like Fillmore (1968), but for these primitive acts, of which he has about twelve. From the point of view on case expressed in the system described here, both these strong positions have drawbacks, indeed they have complementary ones, In .the first place, illm more's (1968) system, with the aid of which he wants to contrast verbs by means of their frames, only becomes significant if Interpreted with the aid of some non-surface'representation of actions. So, for example, Charniak has pointed out (personal communication) that the earlier contrast of the frames for &quot;hit&quot; and &quot;break&quot; is significant only if there is Pome common, underlying, action that the two verbs share, and which can be thought of as being substituted for the horizontal line in the frames. Eor, if that is no& so, theq the contrast of the frames for &quot;kT11&quot; and murder&quot; is of no more inkerest than the contrast between &quot;kill&quot; and I I remember&quot;, or any other random verb, In other words, it is only because the two vtrbs plready' have something in common, over and above their case frames, that the comparison has point. Thus, the contrast of the case frames of only surface verbs is, if unsupplemented, unsatisfactory, null Conversely, there may be certain problems inherent In Schank's attempt to both (a) relate surface verbs to underlying primitive actions, and then discuss only the latter, and (b) at the same time make all participants in the case frames for primitives obligatory. So, for example, Fillmore would express the case frame for see&quot; as (OBJECT DATIVE) and for &quot;learn&quot; (OBJECT AGENT). While Schank (1973 pp. 220-1) expresses both verbs by an underlying primitive PITRANS together with a case frame, for the primitive, containing at least A, 0 and R (Fillmore would call R by D), The individual letters for cases assigned by different authors need not detain us, nor need their contrasting interpretation of the case names, for the present point is the perPSectly general one that, whether or not Fillmore is right with this particular verb pair, it is highly likely that there are pairs L_.</Paragraph> <Paragraph position="16"> of surface verbs like this one whose surface case frames are different and whose Schankian primitive act is the same, It follows-from (b) above that, for Schank, their &quot;deep&quot; case frame must therefore be the same too.</Paragraph> <Paragraph position="17"> Since, for him, every surface vetb has a main act expressing it, there is clearly going to be a problem with this consequence of (b) unless he is prepared to say that there is no necessary relation at all between a verb's case frame and the case frame of its corresponding primitive act.</Paragraph> <Paragraph position="18"> It may well be possible to defend such a position within his theory, but he will still be left with the difficulty that: verbs with quite different semantic behaviour (for Fillmorian case differences ate not superPS ici81) have identical behaviour in his system, There is bound to be a lack of discrimination consequent upon Schank's all-cases-are-obligatory view unless some careful avoiding action is taken, that he has not yet, to my knowledge embarked upon, However, Schank would probably not wish to take advantage of this last possibility because he does not himself hold the view that conceptual case is entirely independent of surface structure considerations&quot;, as was wrongly attributed to him by Bruce (ibid.p,338), and for the simple reason that he intends that case structure in conceptualizations shall resolve the case ambiguities present in English preposition constructions.</Paragraph> <Paragraph position="19"> Schank (1973) makes this quite clear, and I shall return tc~ ilt when discussing preposition constructions in the next section. Thus, since Schankian case frames for primitives are - not independent of all surface structure considerations, he does have the problem above presented by the contrast of &quot;see&quot; and learn&quot;, The burden of the last arguments have been to show that three different positioas on the question of &quot;what are case frames - for?&quot; are unsatisfactory : (i) that they are simply for surface verbs : Fillmore's 1968 position, argued against along the lines sketched above as early as Schank (1969), ( ii) that underlying structures containing primitive actions do not require case frames at all: the Generative Semaqtics position, for example, Postal (1971), (iii) that underlying structures containing primitive actions are case frames for those primitives and, moteover, are all obligatory case frames: Schank's position just discussed.</Paragraph> <Paragraph position="20"> Let me now restate the position of the present system, as it concerns formulas, Formulas are meaning structures for surface word senses, Formulas for surface verbs can be interpreted as case frame$ for the verbs, in that they contain case subparts at the top level (i,eo depending directly on the head action primitive) thdt the formula maker has considered necessary to express as part of the meaning of the verb. Thus, the formula for &quot;break&quot; given earlier (10) contains case subparts at the top level (i.eo depending directly on the head primitive STRIK) INSTrument, GOAL, OBJect and Agent(=underlylng SUBJect), The formula maker is concerned only with specifying, as best as he can, the semantic preferences of the particular surface word in question, If it were a verb he would semantically specify the preferred agent, say, of that action, witbno implicit reference to agenthood as such. The same goes for the other cases: he would attempt to put those cases into a fopula that he th~ught necessary to specify the meaning of the action, Consider (16) He lives in Lugano I and (17) He drank some Barbera in Lugano, It is logically true that one must, if one drinks, do it somewhere, but no one wohld hold that the notion of location yas involved in explaining the meaning of drinking, However, one might well hpld that it was impossible to explain the notion of living, in the sense of inhabiting, without making clear that it was done in some location. Thus only in the formula for live&quot; would we expect a location case subpart (o.,..LOCA), This description of the insertion of case specification into the formula for some given surface verb does not correspond in any clear way to Fillmore's obligatory or optional distinction, though it seems clear that any formula should contain at least - Fillmore's obligatory cases for that surface verb and, as we saw , (10) above, for &quot;break&quot;, does this, since the qnly obligatory case for &quot;break&quot; is object, The main reason for this difference remains the essentially generative quality of a Fillmorean case frame, Much sophisticated linguistic argument in the past decade has gone into aptempting to prove that the term &quot;gen'erate&quot; is neutral Getween &quot;analysetl and 11 produce&quot;.</Paragraph> <Paragraph position="21"> One of the best efforts is to be found in~yons' (1968, p.155), But, although it is easy t~ see the productive role of, say, illm more's subject selection rule (SSR), it is very hard to see what analytic signrficance it could have; the surface subject is, after all, usually revealed by simple methods not requiring the notion of case. The SSR tells one how to cho~se the subject 8iven - the case structure and in that sense is inherently generative in its non-neutral sense J meaning 'productive'. The system presented here however is inherently analytic (its productive aspects have been described in(He~skovits (1973)), and this accounts for a great deal of the difference of approach to the notipn of meaning structure5 Those with practical acquaintance with sentence analysis and production will need li~tle persuasion that the two processes are not in any sense simple inverses of each other (nor does Fillmore himself believe they are, see illmore 1972 p, 23).</Paragraph> <Paragraph position="22"> Charniak has argued (1975) that the method af formula coding results in top level case subformulas that cannot appear in the surface form of the verb, and he cites (ibid, p,16) the formula .for &quot;drink&quot; : where the containment group (SELF IN) indicates that the liquid object (FLOW STUFF) is moved into the animate agent (SELF), and the directlon group ((MAN(Tmu FART))TO) indicates that this movement is in the direction of a human aperture, These two groups appear at the top level of the formula, and thus dependent on the head primitive action C~~~~. However, later inferential procedures of extraction (see below) would show, 8s with &quot;break&quot; (10) earlier, that the real dependency of the two groups was on MOVE, However, the requirement 1s observed that the Fillmorean obligatory cases (only Agent presumably) for &quot;dr~nk&quot; appear at the top level, and the other case groups - containment and directlon most certainly could appear at the surface with 'drink' as in: (19) John drank the beer up thrwgh his nose with a straw and into his brain, where both cases appear at the surface. Charniak is confusing what can appear at the surface with what it would normallxbe redundant to say.</Paragraph> <Paragraph position="23"> The diifelrence is made clear in (19) if we choose to say something remarkable, and false.</Paragraph> <Paragraph position="24"> In this section on formulas I have tried to justify the interrol e of case description formulas : by claiming BOTH that they give a meaning structure that, under inferences processes, can provide a real underlying structure for text AND that they must include enough of the surface case behaviour (of the verb meaning being expressed) to give procedural power in analyzing the input surface sentences. It seems to me essential to preservehoth these roles, and to avoid the path taken, in their different manners, by Schank, Fillmore and the Generative Semanticists, qf displa$ing a full underlying structure directly without the processes that reach it. I argued earlier that each of those three gave only a filled-in or- fina1,structurg which in itself gives no hints as to how you get there.</Paragraph> <Paragraph position="25"> The generative linguist argues, of course, that he is not trying to reach such a stfucture at all, but to generate surface structbres C_ from it, and the &quot;artificial intelligence critique&quot; of this aspect of generative linguistics is familiar by now (my own version is in ilks &quot; 1975~). The wqy in which Schank emphasises a filled-in structure is a quite different and more interesting matter, His conceptualizations are filled-in structures, with no procedural capacity.</Paragraph> <Paragraph position="26"> However, at the back of his early papers (i.e. the Appendix to Bchank et a:. 1970) are dictionary entries or skeletons for conceptualizations, which contain selection restrictions on the slots in the conccp tualizations (and in (Schank 1973, p. 229)). These objects are much more fundamental to his approach than is generally realized, for they do give it more procedural power than one could envisage from looking only at the final conceptualizations, Their role is somewhat like that of the formulas for actions: they are blue-prints showing what the system would like to be the usage and context of a given action. But the difference from formulas is this: the Schank dictionary entry has slots, marked, say, HUMAN which, when they are all filled by surface vords yield the filled-in conceptualization, The formula too has case specifications, like (*AN1 SUBJ), but these are not slots,to be filled in, but directions as to how to fill in an agent slot in a higher older entity called a template whlch consists of whole formulas: essentially, a network based on an agent formula, an action formula and an object fdrmula (although any of these may be dumi es) . Schank has no equivalent to formulas for nouns or adjectives, or any part of speech other than verbs.</Paragraph> <Paragraph position="27"> Thus, if the formula (10) for &quot;break&quot; is placed at the action node of a Lemplate, its agent preference txies to ensure that a formula for an animate entity will go at the corresponding agent node. There is a metric for this initial parsing and matching described in (Wilks 1972, 1975a,, 1975bJ. The role of the formulas as data for parsing I here is as follows: templates, matched onto sentences and clauses expressed as strings of formulas, try to pick up surface subjects and use the formulas as blue prints for electing the best template, by trying to ensure that the surface subject is also the agent. In the initial matching of the he harmner broke the window'' this will not be possible because the formula for hammer&quot;, the surface subject, will be placed at the first (or agent) node of the template, However, since the formula for &quot;break1' will be at the corresponding action notle the system will know at my point in later processing that the surface subject is not the underlying (animate) agent as preferred by the &quot;break&quot; formula at the action node. of khe same template, This clash will, as we shall see, lead to later ext'raction inferencing that deals 'with the ei!&quot;gatiw paradigm in a unikrm manner. But this clash - is only observable in a system which builds structures that retain thelr preferences built in: i.e. not in one like Schanks' that just fills ia slots in dictionary entries. Fillmore's anomalous Naon found Harry tl sleeping&quot; causes no trouble here: the formula for find&quot; expresses a preference for an animate agent, In the template for this sentence, It that is not satisfied by the surface subject (a formula for noon&quot;, with a head (WN POINT)) but no case frame is violated, Formulas are not used as parsing mechanisms to deal with preposition structures in English, This is done by other structures called paraplates, to which I will now turn, so as to givc a fuller account of them than in the past, in (Wilks 1975a), In what follows templates will be written in short form: not as triples of complex formula trees, but as square brackets round the surface phrase or clause that the template is for, The words are clustered within the brackets so as to represent the three nodes of the template, Case in paraplates Paraplates are structures employed to assert a connectivity between two templates, typically between a template representing a main clause and one representing a prepositional phrase, Thus, if we were analysing (20) John left his clothes at the cleaners which would be represented initially in the system by two templates, the correct paraplate, when applied, would assert a spatial location Each paraplate corresponds to one of the cases in the inventory, which is the same as the inventory used to construct formulas. Many paraplates, however, may correspond to a single case, A paraplate has the form of two template-skeletons connected by a label indicating the case expressed by the paraplate, where by template-skeleton I mean an entity like a tepplate except that, instead of a formula at each of its threenodes, it has a function rangkg over formulas. Any template that ma?ches either part of the paraplate mst have formulas that satisfy the functiops in the corresponding part of the paraplate, If the functions in both parts of a paraplate are satisfied by a pair of templates (and the template for the prepositional phrase is nlormally considered to be the right-hand part, though this need not correspond to actual order of occurrence in text') then the case label of the paraplate is asserted in the representation as holding between the two templates.</Paragraph> <Paragraph position="28"> In earlier descriptions I have distinguished paraplates from inference rules, but in fact they can perfectly well be seen as a form of inference rules as Schanis has argued. However, the essential role of paraplates is as parsing structures for prepositional phrases. So, if we were representing &quot;~ohn picked up the statue made oul of wood on the table after lunch&quot; we would expect paraplates for the various case dependencies to create ties as follows: where SOUR indicates source case, and TLOCA, time location.</Paragraph> <Paragraph position="29"> The preference semantics system applies paraplates immediately after matching templates and chaosing the ''most preferred&quot; ones (Wilks 1975@). In operation, the system inputs small English paragrdphs online, produces a semantic structure for them, and from that generates a French translation (Wilks 197313, 1975a; ~eljSkovits 1973), Thus, as regards Engliqh, the system is an analytic one, and is faced with the standard problem that a single English preposition can introduce many cases (and can, of course, often be translated by a number of French prepositioris depending on the case).</Paragraph> <Paragraph position="30"> Let us consider &quot;by&quot;, fudctioning in the following sentences, all of which may be considered to start, like (23), with &quot;He left Lugano by ..,&quot;, where I have indicated the apparent (though disputable) case of the last clause at the right of each line: filled, corresponding to Agerit-of-first-template, Action-of-firsttemplate, Object-of-first-template, and so on for the second template. Here are four paraplates that should match onto the templates for the sentences above with corresponding numbers, Like the sentences, the paraplates will all have the same left-hand side, which is written only The parentheses containing the formula parts are all to be interpreted as matching onto correspondibg part of a template if and only if the latter has the mentioned subparts as its head of formula. Thus paraplate (24)' matches (24) because the formula for &quot;he1' has head MAN included by *AN1 &quot;left&quot; has as head MOVE and so on for the other correspondences of sentences and paraplates, which will then assert the case label tie written at the right hand end of the sentence in each case (and on the corresponding arrow in the paraplate) as hcildipg between the correspon'ding template pair, $ is a dummy place holder, *DO covers a wide class of actioni, as does *REAL of entities, The paraplates above (24) ' to (29) ' are a sublist of those stacked under the name of the preposition &quot;by&quot; and under the primitive action MOVE that occurs in action position of the left side of each paraplate, There will also be other sublists of paraplates stored for &quot;by&quot; under other primitive actions, For example, (30) This painting is by Titian would, during its analysis, also access the paraplates under &quot;by&quot; but in this case the sublist under BE, The su6lists of paraplates, for a given preposition, that are names by different primitive actions are not ordered with respect to -each other, However, within any sublist, such as the stack (24) ' to (29)' above, the paraplates - are ordered, in that the paraplate at the top of the stack is applied first, the next only if that fails and so on. Let us return to the example to see why, First, INST case is largely a default case for MOVE as it is cued in by &quot;by&quot;, in that almost any entity can be an instrument here if we have no reason to beli@ve it is anything else. Thus the more specific (24)' must be applied before (25)' in order to match direction case for (24) ' since, if the order were reversed, (25) ' might match with what &quot;ought&quot; to match with (24)'. We could imagine something very specific in (25)' to match the formula for car (such as a formula expressing &quot;thing for moving humans&quot;), but that would risk missing rt by cattle truck&quot; which is not normally used for transporting people. However, it should be noticed here that the paraplates as described do not have the power to detect an anomaly such as &quot;John walked home by car&quot;.</Paragraph> <Paragraph position="31"> The templat&matching that sentence would be quite properly tied (as INST) by (25)'.</Paragraph> <Paragraph position="32"> Any anomaly in it would have to be revealed by later inference, This is an example of what Lyons (op.clt.) means by the criteria of ahalytic systems bein'g (initially) weaker than those of productive systems.</Paragraph> <Paragraph position="33"> A stronger but simllar argument applies to putting (28) ' a5ove (29)' in a preference stack, given the very weak criterion embodied in (29)', that almost any action upon any physical object would satisfy in default; and one could extend this to putting a (23)' above a less specific (26) ' , The main point here is as follows: it is clear that snch paraplates must be applied in some order, but it is not clear that they should be slmply ordered. For example there is no reason why (28) ' should be ordered with respect to (24)'. There is no formal trouble expressing a partial ordering of this sort procedurally. I argued in (Wilks 1975a) that in many cases we might expect more than one template attached to a fragment at this stage and that we would &quot;resolve the ambiguity&quot; by preferring whichever template matched higher up such a preference stack of paraplates. Order, then, is important in such paraplates stacks, even if it turns out to be only partial order. If such a stack is only partially ordered then we may, in the case of a prepo~itional phrase with two or more current templates competing to represent it, expect a - draw at some stage: that is to say, two possible templates for the same phrase may match with paraplates that are not ordered with respect to each other. Such a situation might well correspond to an example such as: (31) John jumped on the box where the last phrase is genuinely ambiguous between TO and LOCA case (in the sense in which the substitution of &quot;onto&quot; for &quot;on&quot; could only be interpreted as TO). Thus under &quot;on&quot; we would properly expect MOVE (for &quot;jump&quot;) paraplates for TO and LOCA to be mutually unordered. But let us turn to: (32) He beat the girl with a withered arm where the case dependence of the prepositional phrase could be either instrumental or a specification of which girl it was. In such cases one might imagine sorie expectational force to be drawn from the case content of the formulas. Thus, if &quot;beatp' has an instrument group in its formula, we can conceive of using this fact to decide the draw on the grounds that &quot;beat&quot; really expects&quot; an instrument, so why not give it one here, and settle the question. And there might indeed be somc psychological grounds for doing that, But in general, it is clear that the paraplates do not function expectationally, except in the sense that by looking first at the parablates most preferred in the stack (for a given preposition and basic action) one could be said to be expecting it, And that sense of &quot;expectation&quot; is perfectly consistent with trying more than one candidate template, for a given phrase or clause, at the same time against the paraplate stack, as we might have to in. a sentence like (33) He beat the dog with a branch where we would not only be resolving the cake of the preposit~onal phrase, but also, and at the sme time, the word sense of &quot;branch1', where the senses of &quot;branch&quot; might give rise to a corresponding number of templates and we would always prefer the one that matched further up the paraplate stack, as in (Wilks 1975a), The reader may have been struck by an overlap of case content between the paraplates and formulas, For example, in (10) the formula for &quot;break&quot;, the head action CAUSE has a dependent (THING INST). Yet, if we were analyzing the standard sentence &quot;He broke the window with a hammer&quot;, we would expect to use a stack of paraplates under &quot;with&quot;, containing a substack whose left-hand action was CAUSE, and in it find the appropriate paraplate for tying together the two templates for that sentence with an INST tie. This latter information might seem to overlap heavlly with that contained in the formula (10) and the question arises whether it need be stated twice, However, although the tw~ foms of information overlap there is no reason to believe they are the same: the paraplates reflect the case ambiguity of English prepositions while the formulas reflect only the sorts of case inferences one might subsequently want to make (in extraction, see below), There is no reason to suppose that one form for data can suffice for both these activities, However, Schank does make this assumption, and the point here may become clearer after some brief recall of his position on the point, In (Schank 1973) an account is given of case inferences from a primitive action. In (34) John shot the girl ~ith a rifle the underlying primitive is PROPEL which 5oms the center of any conceptualization for &quot;shoot&quot;. An obligat'ory INST inference is always made for PROPEL or, as ~t, was put earlier, EGT is part of the (obligatory) case frame for PROPEL, and also (aS it hawens) an optional part of the Fillmorean surface frame for &quot;shoot&quot;. In ~chank's scheme Instrumental case iz~volves the insertion of another primitive action into the conceptualizationo gut that is not essential to rhe present point which is that, for Schank, PROPEL &quot;expects&quot; an instrument, and criteria can be postulated such that a rifle will fit those criteria, in a way that hair will not, as in (35) john- shot the 6irl with long hair.</Paragraph> <Paragraph position="34"> Thus, for (34) an INST interprcktion is achieved for &quot;with&quot;, while in (35), by default, the hair is taken as a specification of the girl. These solutions are, of course, correct but the different aspects of the phenomena do not seem to connect in procedural terns. So, for example, Schank is perfectly well aware of the case ambiguity of the preposition &quot;with&quot;, and even lists four forms of it (ibid, p. 231) corresponding to diff~rent cases, aloqg with four 11 c~ncepttual realizations for the syntactic item 'with (noun)' &quot;, and specifies that they should be checked in order &quot;for conceptual validityu, just as paraplates are, The correspondence to the notions described In this paper (and in Wilks 1973, in, the same volume) is reasohably clear: the inference of an instrument from PROPEL corresponds to an instrument specification in a formula for &quot;shoot&quot; (and for Schank it also comes from shoot': rather than more generally from PROPEL, since otherwise it could not be so specific about the instrument being a gun, as PROPEL does not deal generally in guns); whereas the ordered list of case possibilities for &quot;with&quot; is not developed, but corresponds roughly to a paraplate stack for &quot;with&quot;, But here is the problem: in Schank (1973) the two forms of information do not actually meet in any general way, Schank mites as if the list of possible case functions of &quot;with&quot; is general (i.e. action independent), but we have shown that it may well be specific to different primitive actions, in that there may well be a separate paraplate substack for e~rh primitive action, and may be no short gel.yral list of functions of &quot;with&quot;,</Paragraph> <Paragraph position="36"> example (34) is resolved because PROPEL alld/or &quot;shoot&quot; &quot;expecrs an instrument&quot;$ but that piece of information does not, and cannot, take accoupt of the pre-existent ambiguity of &quot;with&quot;* . it just happens To fit this example, because PROPEL &quot;expects&quot; an instrument and instrumental case is put at the top of the ambiguity list for &quot;with&quot; (ibid, But will this coincidence hold in general? If we had been dealing not with PROPEL but with some ether primitive, it is not clear that the same order for &quot;with&quot; would help, I degcribed earlier a MOm primitive underlying leave&quot; (the primitive and its name are not important, on-ly that it is not PROPEL), and considered sentences like (36) 1 left Lugano with Ywo pounds in my pocket which shows that instrumentality is unlikely to be the expectation1' for &quot;with&quot; from whatever primitive underlies &quot;leave&quot;. Hence Schank' s list for. &quot;with&quot; is perhaps specific to its use with PROPEL, and he must develop sqme structure analogous to paraplates and not, assume $ i.e. a &quot;dictionary entry&quot; for a preposition (like a paraplate subtrack) will be a function of a primitive actioxl, that the instrumentality of PROPEL suffices: My point is simply that any attempt to introduce generality here will show that another sort of information structure is required to deal with one specific to particular prepositions, and that general case inferences (like instrumentali't~f from PROPEL) could only help in special cases.</Paragraph> <Paragraph position="37"> Any develppment of &quot;wlth&quot; llsts into paraplate substack structures would, I believe, show that the expectational inferences from the primitives play a far smaller role In analysis,because the crite~ia they contaxn will have been pracedurally expressed somewhere else, It is true, however, that there is a strong expectation of a guninstrument from &quot;shoot&quot;, and this phenomenon does not fit easlly into the paraplate scheme. It might be necessary to ensure that if a specific instrument were mentioned in a verb formula then that could override subsequent paraplate matching when it; arose, However, such specific instruments are not the norm for PROPEL, for this primitive presumably also underlies &quot;break&quot; PS-or ~chanl;, and there is no such presumption of specific instrument with that verb, Indeed, such an expectation would be bound to lead to error if any object found after &quot;with&quot; were taken as an instrument, rather than using a breadth-first approach like paraplates that considers the ranked possibilities for that preposition and that action. This is especially true for a system like Schanks that has no back-up and no possibility of recovery if it goes wrong, Formula and palaplate information do sometimes overlap, and I suspect SchankTs ~y$tem, that assumes that the same mechanism can serve both functions, draws much of its force by choosing an example from an area like (34) where they do overlap. But how, in the examples of (23)-(29), could we work in genexal with the ambiguity of &quot;by&quot; by assuming that MOVE had sdme strong single case expectation, over and above the fact that - some case paraplate (like (24)' for TO) would have necessarily to Appear at the top of any substack. How could Schank express the action-dependent ordering of the rest of the substack? Another peculiar feature of SchankTs treatment of these phenomena is his suggestion (ibid. p. 232) that any ordering of preposition functions must go &quot;from the general to the specific&quot;. Such limited data as are afforded by substacks like (24)' - (29)' indicate that it may turn out to be the other way round, because the most general (right-hand) criteria tend to appear in the paraplate at .the bottom of the substack, Much though not all, of what has been sald about Schank's approach applies to Riesbeck's (1974) implementation of it. It is again a strongly expectational system, in practice in ~iesbeck's case, and that leads to the mentioned defects of a depth-first appr~ach (if implemented without back-up, as it is),since the whole notion of &quot;preference&quot;, in case as elsewhere (Wilks 1973a), depends on following a number of possibilities breadth-first for a while before comparing them and committing the system to only one.</Paragraph> <Paragraph position="38"> Only in that way could one treat, in a general way, examples like (33) that require comparison of how high up a paraplate stack a given interpretation comes relative to others.</Paragraph> <Paragraph position="39"> ~lesbeck's system generates expectations (called &quot;requests&quot;f from main sentence verbs and uses these to anticipate the case parts of the seritence. The main difference from Schank's approach, and which makes it a curious implementation of it, is that it deals with prepositions not, as one might expect, by primitive actions seeking certain cases, but by verbs seeking actual preposition names. So,</Paragraph> </Section> <Section position="8" start_page="12" end_page="12" type="metho"> <SectionTitle> 1 t </SectionTitle> <Paragraph position="0"> for exmple, prevent&quot; he seeks &quot;from&quot; and &quot;by&quot;, jus-t as the computational linguistic systems of the fifties and sixties did, It also has less power than Schank's theory (and paraplates) in that Riesbeck specifically says (ibid. p . 95) that requests (including preposition requests presumably) are not ordered, All this seems to require that all possible preposition sequents be stored for every surface verb in the system which leaves little scope for semantic generalization (the aim of the whole exercise, presumably).</Paragraph> <Paragraph position="1"> The strong point of difference, between Riesbeck's system and the one described here, concerns the role of uninhibited &quot;expect-</Paragraph> <Paragraph position="3"> The examples presented here suggest that one cannot base a system simply on the observation that if someone uses &quot;by&quot; after &quot;leave&quot; then it will be expected to introduce an instrument, For &quot;~e left Lugano by day'' does not fit that and yet cannot be sa4d to be violating any axp,ectation, It is simply that the structural ambiguity of prepositions muse be accounted Eor in a structured and general manner like that offered by the pawaplates. The usefulness of expectationg in any field, depends on (a) the high chances of Tt being right, and (b) knowing what to do when it is frustrated, and ~iesbeck's system seems to lack both these desiderata, $ For further discussion of the issue, see (Wilks 1975c, pp. 40-45) Case extraction inferences Extraction inferences produce new template-like forms from the case content of formulas embedded in source templates: those that have been matched with text sentence f ragrnents, They are styled &quot;template-like&quot; here only to indicate that they have not been matched with input text, And hence the inferred propositions they represent have not necessarily been stated explicitly in the input text, Let us first see the effect of doing this, and then the mechanism that does it, In what follows, we extend the &quot;short form&quot; of templates (obtained by writing square brackets round English words, clustered at three nodes to show the distribution of formulas in the full template) by writing extractions as English words inside double square brackets, Let us consider (37) John fired at a line of stags with a shotgun The result of matching this with templates, applying paraplates as described above, and then performing case extractions can be written in summary form as fallows:</Paragraph> <Paragraph position="5"> a+ sha tgun 1 The extracted templates are tied by dotted lines to the source template from which they have been extracted, and the case name on the dotted line shows the case type of the extraction, The inferences cover both those that must be true (like the OBSE extraction, slnce to fire at a line of stags 1s necessarily to fire at stags) and those, like the GOAL extraction, that are only likely.</Paragraph> <Paragraph position="6"> These extracted templates are not printed out (in the way in which, MARGIE (Rieger 1974) daes) for they do not in themselves constitute testable output. Their role, as described in (Wilks 1973c, 1975a) is to be data for further inferences using common-sense inference rules, whose function is not essentially connected with case and will not be recapitulated here. In the implementation described in those references, the extractions were purely &quot;problem driven&quot;, in that they were only done when some problem of reference resolution in the text demanded that the representation be deepened, However, that was a strategic consideration in no way a consequence of the nature of extractions: they could equally well be data driven, and be executed after every matching of a sentence with templates, The difference between paraplates and extractions should be clear from (38) because it contains both an INST paraplate-imposed tie between main clause and preposition phrase templates, - and an INST extraction from the main clause template yielding an extracted template equivalent to &quot;John uses a thing(=a shotgun)&quot;, In this way, we are able Lo retain as part of the overall semantlc representation of text,and in a message form appropriate for subsequent inference, both the surface structure (paraplate INST tie) - and the underlying structure (INST extraction inference) As we saw in the last section, Schank hblds that these are one and the same thing.</Paragraph> <Paragraph position="7"> It is true, as we shall see, that in order to extract that John uses a shotgun, the extraction mechanism must consult the template tied by the )INST paraplate but, as I wguedr in detail, in the last section, this In no way establishes that the urfac semantics and the underlying sei~ antic structures are one and the same, requiring only a single representatlon, The ettraction mechanism consists of a 'specialist'l(to use Winograd's tend for each case (and for CAUSE, which is treated as a semigcase during extraction). It is called after an initial semantic representation for a text has been built up as templates tied together by para?lates and anaphora ties (See Wilks H73c, 1-975b) An extraction, resulting in a new doubJe-square-bracketted template, as in (38) above, is made for each case (or CAUSE) sub-formula at the top level of the formulas of each source template.</Paragraph> <Paragraph position="8"> Let us see how the extractions in (38) are actually obtai'zed.</Paragraph> <Paragraph position="9"> This will requlre that we give more of the content of the first source template in (38), and in particular the formula for &quot;fire+atl', (39) may be considered a semi-full-form of [ 30 hn f ired+at line( of+stags)l in that the centre node has been expanded to its formula but the other two nobs are left in &quot;short fom&quot;, Each top-level case (or CAUSE) subfornula is extracted by the appropriate &quot;case specialist1' which has access to the additional information grovided by the whole template in which that formula is embedded (and, as we shall see, to other templates as well).</Paragraph> <Paragraph position="10"> By way of illustration, we will look at the OBJE and GOAL extractions from the action formula in (39).</Paragraph> <Paragraph position="12"> The dependent of the OBJE case in (39) shows that &quot;firejat&quot; line(+of+s tags) I prefers an ANImate object, but the formula is in a template whose object is - not animate (it is &quot;line&quot;), and so we have a failed preference, However, an animate object (stags) is available as a dependent of the surface object in the template, The extraction process takes the form of filling - a new copy of the source template, and imposing the available preferred animate object, to yield: C C John f ired+at s tags] 1 ~xtracting the GOAL, case is more complex.</Paragraph> <Paragraph position="13"> As we saw earlier, in the formula (10) for &quot;break&quot;, the real dependent of a GOAL case is an assertion group, The appropriate full dependent to the left of GOAL in (39) is</Paragraph> <Paragraph position="15"> where $ is indicates the position of &quot;GOAL&quot; in (39).</Paragraph> <Paragraph position="16"> This is an assertion group as it appears in a formul J. The extraction takes the form of rewriting this in tmplate form (from SOV to SVO form) and attempti% to fill in its nodes wit11 full formulas matching the preferences: so, on rewriting the above we get:</Paragraph> <Paragraph position="18"> in which an agent and object can be inserted fromithe whole template (39) so as to match the preferences expressed (if we incorporate the above 0BcJ;EFj extraction as well). Thus, we get an extracted &quot;short form&quot; template (converting STRIK to &quot;strikes&quot; for uniformity of expression, though it really remains as the primitive): C [John strikes stags1 3 The extractions for INST and CAUSE require that we are able to look outside the source template for conPS irmation to other templates, .-Here, only INST receives any such confirmatim because there is no mention of any &quot;bullet&quot; or other missile that would confirm ( (THLNG MOVE) CAUSE) ) . Since the' primitive dependent on INST is THING. the &quot;INST specialist&quot; produces a copy template whose primitive action is USB (if the dependent had been another type of entity, the rule could have been different) . The INST-tied template matched to &quot;with a shotgun&quot; in (39) is accessed by the INST specialist to provide the full object formula for USE yielding the extracted template shown in (39) as: (40) [[John. uses thing 11 (=shot gun Let us look briefly at how extractions cope in a novel way with the standard ergatxve paradigm rehearsed at the beginniug of the It paper, and in particular with the interesting farms The window broket' and h he hm-er broke the window&quot;, Initially these will receive a template match so as to yidd a representation even more superficial than that given in Simmons' system. Under extraction however, appropriate representations are obtained and cued by the fact that the superficial subject formula in the templates does not, in either case, satisfy the animate agency preference of the formula (10) for &quot;break&quot; which is the action node of both tkmplates, Thus h he window broke&quot; yields miitially: (41) C the+window broke Now, on extraction, ~~~'suEJ (agent) specialist sees not only that (a) the surfqce subject (window) does not specify satisfy the (*AN1 SUBJ) preference of (lo), but (b) the same surface subject does satlsfy the (*RQJYSOB OBJE) preference of (lo), which is filled by only a dummy in the source template (41), Thus the &quot;SUBJ specialist&quot; produces a copy template with the agency preference satisfied: C C some+animate brbke U 33 while &quot;oBJE specialist&quot; correspondingly produces : [:[: R broke window11 and these are immediately conflated, On the general preference milks 1973a) principle of producing the fullest repzesentation possible, as the extraction: f [ some+animate broke windowl 3 where the agent formula (now, of course, a tlue agent, not a surface subject) is merely (THIS *AN), an extractiqn from the &quot;break&quot; formula (10) : it has not been $oafinned by the Cext, and, replaced by a full formula froma source templake, as were the objec~ and agent of (40), Let us turn finally to the sentence matched initially as: (42) [hamer broke window1 and again consider the extractions f rom4 (10) for &quot;break&quot; when at the action node of (421, We can produce the same SUBJ extraction as for (41) above, and we can again, using the &quot;INST specialist&quot;, produce an extraction with USE as its main action prrmitive, as from (39), but in this example che dumles in the extracted template can hat be filled in from an INST-tied template, as in (41), and as would be the case with &quot;John broke the window with a hammer&quot;, but from the surface subject of the source template ,itseif. Thus we obtain an extraction scheme for (42): It is, of cours'e, essential here that the extractions are applied in a fixed order, so that the lNST extraction can make use of the SUBJ extraction and know that &quot;hanrmer&quot; is no,t avaixable to be the real agent of an extraction from (42).</Paragraph> <Paragraph position="19"> 'Not all ergative verbs would receive identibal treatment, Thus, the relat'ion between: (44) John moved the stone and (45) The stone moved would not be the same as between the corresponding examples above for &quot;break1', The action formula for &quot;move&quot; in (44) would have CAUSE as head, but the action (again &quot;move&quot;) in (45) would be a different sense and so a diffezent formula, whose head would be MOVE, and which would not contain an instrument specification. So we would not expect any extraction: [ C some+animate moved stone1 I from (45), though we might well expect an extractiorl equivalent to (45) from the template for (44) There can be legitimate dispute about the limits to which case =tractions should go: so, for example, some writers (Parker-Rhodes 1975) would consider that &quot;John wrote a letter&quot; should be an inference from 1 received a letter from John , and he describes a lattice scheme in which an AGENT case is generally and legitimately inferred from a SOURCE case, That could only be achieved with the &quot;SOURCE specialist1' in the present system from a very full, speci~ic, formula for 'letterT (and extractions can come from nouns and not only from verbs) in which a petson) source was specified at the top level of the formula, Two points should be noticed, in conclusion, about this method of producing a qemantic structure appropriate to the ergative paradigm, First, the &quot;case specialists&quot; are not in general dependent upon the particular fomula in which they operate (e the word the fomula is for), nor upon the primitive that is the fonmlla head, Each specialist does, of course contain ranked side conditions but- th-y are not, except in special cases, dependent on the head of the formula, It is not the case that every line of the specialist for case X, say, is of the form &quot;if the head of the source formula is Y do 2, else ,,. &quot; This point will be important m the next section when we consider Charniak' s arguments concerning c as?.</Paragraph> <Paragraph position="20"> Secondly, this way of describing the ergatwe paradigm is naturally different from Fillmore's because, as has been noted, he is primarily concerned with restrictions of the generation (=productron !) of surface sentences by means of his ordered Subject Selcztiozl Rule: &quot;IPS there is an agent it is the subject, else if there is an experiencer it is the sub,iect, else if there is an instrument .... and so on&quot;. This way of l~king at things assumes that John must in effect, already be marked as an agent, hammer as an Instqent or Object and so on, Simmons' analysis paradigms for the same example are, as we saw at the beginning, an analytic version of thzt rule, The difference between those two approaches and the present one is twofold, First, that the present system is preferential in that John does not have to be.myrked as an agent: no confpsion would be caused here, for example, if he was used as an instrument, Secondly, and more importantly, the construction here of the ergative representation follows not only from the application of the case specialists&quot;, but from the general rule of preference (Wilks 1973b) that as full (or maximally redundant) a representation should be produced as possible. TRis is required, in addition to the case specialists, to produce the extraction null CCsome+animate broke window31 from, the less r&undnmt ergative, (411, On this groudY I wuu1d argue that the phenomenon is dealt with here using a more general rule, and moreover, a geperal rule that drives this whole preference semantics system of analysis.</Paragraph> <Paragraph position="21"> ILL d recent paper, Charniak (t975) has argued that A1 systems that use case notation in their analysis of natural language are ndt really making use of case, and the present system is one of those he cr%ticises in this connection, Argunients along the same general lines are to be founa in Welin (1975). I think Ch-rniak's arguments are important and stimulating, though in sode ways misguided. They have the paradoxical sidk-effect of showing ;hat not even Fillmore 1s really making use of case. I shall briefly sur,marise seven aspects of ~harniak's complex argument, an$ make some reply, (i) The pa$ibi~n Argument, Charniak distinguishes between case notation and position notation (ibid. pp. '3~4). If we have an event of Jack opening the door with a key then, for Chamiak, a case form of that would be: (PREDICATE EV OPEN) (AGENT EV JACK) (OBJECT EV DOOR) (INSTRUMENT EV KEY) whereas OPEN(JACK, DOOR, KEY)' would be pure position notation. He argues that many A1 systems claim to be using case notation but in fact are using only position notbtion: dqing places arbitrarily to pass particular case arguments to a predicate, (ii) The evacuation argument. Charniak argues that the standard benefit of case --- the prdduction of, say, the ergative paradigm for 11 open&quot; from a case representation as in (i) above, together with a subject selection rule --- ean be obtained equally well with the position notation, (ibid. p. 17), Iq other words, the whole business can be evacuated of case altogether, with the case names remaim- null ing perhaps as mnemonic variable names&quot; (ibid. p. 27).</Paragraph> <Paragraph position="22"> (iii)The linearity argument.</Paragraph> <Paragraph position="23"> Charniak argues that the e;xpression of meaning in terms of case is non-linear. Schematically, if it was linear it would look like: while the meaning of &quot;see&quot; would have EXPERIEHCER where the above has AGENT, However, in A1 systems it tends to be don-linear, like:</Paragraph> </Section> <Section position="9" start_page="12" end_page="12" type="metho"> <SectionTitle> MEANING (LOOK) =MEANING (OBSERVE ,AGENT) </SectionTitle> <Paragraph position="0"> or, in other words, the meaning of MTANING(AGENT) cannot be computed independently of knowing that the agent is the agent of OBSERVE.</Paragraph> <Paragraph position="1"> (iv) Theproduction argument, Charniak argues that thenotion of case for Fillmore is es sentially connected with its productive generative role in controlling surface grammaticality, and that since AI systems are preoccupied with analysis, they are not making use of case, (v) The case-content argument, Charniak argues that we are never told what it is to be an agent as such, This is a subargument of (iii) above that one cannot compute MEANING(AGENT) independently, (vi) The surface analysis argument. Charniak demands that the &quot;semantics-based systems&quot; in A1 aJso do surface case analysis, of the type found in illm more's papers. He argues that some such analyses, like the formula gi-ven earlier for &quot;drink&quot;, qctually exclude the surface forms, but I showed earlier that this is not so.</Paragraph> <Paragraph position="2"> (vii) The case-inf erence argument. Charniak argues that a system using case for semantic representation should provide case infe.zencesll. The meaning of case CASE would then be &quot;the set of inferences me tan Wke &bout X, knowing only that X is in case In reply, I think Charniak misjudges where the heart of his argument lies by giving undue prominence to rather bizarre arguments like (i) and (ii), whereas the serious ones are those like (iv) and (vi) that draw attention, as this paper has tried to do from a quite differ-</Paragraph> <Paragraph position="4"> ent point of view, to the need to relate the use of case in semantics-based understanding systems&quot; to its original Fillmorean role as a classifier of superficial formsllike preposition and other case constructions in English. I would paraphrase Charniak's Wrnent on this</Paragraph> </Section> <Section position="10" start_page="12" end_page="12" type="metho"> <SectionTitle> 1 t </SectionTitle> <Paragraph position="0"> point, perhdps unfairly, as claiming that case&quot; takes its meaning from surface phenomena and therefore any attempt to locate it only in a deep t! semantics, wholly divorced from such phenomena, is to make case&quot; meaningless. My answer t6 that, and it is a serious charge, is the presentation, here and elsewhere, of a system that - both parses surface English by means of case - and expresses its underlying semantics in the same notation, The third aspect of CharniakTs arguments, like (v) and (vii), is his concern with the precise content of the cases, This bothers everyone, including Fillmore, who had beer. unable to provide it, and so if the production of such definitions is made a necessary condition of using case then Fillaore, like everyone else, fails the test.</Paragraph> <Paragraph position="1"> Let Be, ip conclusion, reply to the arguments above in a little more detail.</Paragraph> <Paragraph position="2"> (i> is, I think, simply confused, The two (case and position) are simply equivalent forms of information, provided one knows that the &quot;first argument is the agent&quot; etc. etc. Charniak himself seems to see this when he admits (ibid. p, 4) that one could have case notation without case, and case without case notation, (ii) shows only the triviality that any cornputatidla1 system could have all its top level notions replaced by gensyms, like WQ5.567, and function in the same way, This would not however show that the notions being programmed were not essential to the meaningful function and interpretation of the system. In any system doing propositional logic, the notion AND can properly be replaced by a truth-table vector 1000, but that does not show that the notion AND represents is vacuous, or a mere mnemonic convention.</Paragraph> <Paragraph position="3"> (iii) does not establish that systems must be wholly linear, in Charniak's sense. However, it would clearly be a defect in a system if, every time an inference about AGENT was made, the system had to look and see wl4at verb it was the agent of, or what the primitive action coding the verb was, Some examples of Schank's that Charniak discusses do show that this access to the head primitive is somerlmes made, and Charnzak then argues that the case is ipso facto dispensible, and the inferehces arabatter seen as inferences from that7-head Grimitive aqtion, But, as T pointed out. when discussing extraction, the I I case specialistsi' in,the present system, at least, do not generally inake such reference) back co the acr;fun head, So, for example, the &quot;GOAL-specialist&quot; did not have to find out, before operating, what the action was that a given GOAL was a goal of.</Paragraph> <Paragraph position="4"> (iv) and (vi) These arguments are partly true but their points are aet by presenting a system, like the preference semantlics one, that ilses case both as a semantic representation and a parsing mechanism (the paraplates) , (v > It is true that no system, Fillmore's fncluded, has given precise definition of the cases beyond the most general indications, such as that agency is concerned with initiating force etc. This criticism is true far beyond case: it applies equally to notions of thinghood, causation and movement etc, It is just a fact that basic semantic notions arelvague,and perhaps necessarily so.</Paragraph> <Paragraph position="5"> It is certainly no ground for lament, and does not prevent us speaking precisely in everyday life. Provided, therefore, that a programed system can perfarm precisely using these vague notions, we can demand no more. The fallacy is to imagine that underneath the vague notions used in thought, speech and writing, there must lle precise criteria for their use.</Paragraph> <Paragraph position="6"> This is a persistent fallacy that has received much attention in Twentieth Century philosophy.</Paragraph> <Paragraph position="7"> In the present system, as in many others, specifications (selectional or pteferefitial) are given of the agents of particular actiqns [human for &quot;think? a$imate for &quot;see&quot;, and: so on), One could argue that &quot;agents as such&quot; are, if anything, the union df all such agents, but that will not, and cannot, yield an intensional definition of the notion. Extractian operating on such specif icatidns (the &quot;SUBJ- null specialist&quot;) yields Pnferences such as that some animate does so and s?&quot;' but these do not meet Charniakt s utll~and in (vii) above. Iteis worth notifig-at this point that Charniak puts his demand (or test for &quot;using zase&quot;) in such a way (in points (v) and (vii) above) that it is hard to see what could satisfy it: what result of computation could we conceivably expect as the value of MEANING(AGENT)? I would maintain that extractions, as described above, are most certainly &quot;case inPS erences&quot;, but are not &quot;the set of inferences one can make about X, knowing only that X is in case CAS~Iylbid. p. 19) because that set is not the sort of thing one wants to know in a language understanding sys tern. The &quot;case specialists&quot; in ogreration produced inferences taking information from the appropriate points in the representation; but from knowing merely that John is an agent nothing interesting follows or could follow! (And from that fact nothing follows to the effect that agency is vacuous, as Charniak seems to think), The most that could follow would be that John does somethfng or instigates sortne art.</Paragraph> <Paragraph position="8"> We -could hatdly infer at that point that John was animate, because that is presumably information generally attached to &quot;John&quot;, and, in any case, some verbs (like &quot;calculate1') may have a selection or preference restriction on their agent that is more general than animate, so that being animate could not follow from being an agent.</Paragraph> <Paragraph position="9"> To infer anything useful, as in the extractions earlier in the peaper, we have t~ look not just at what John is the agent of (which would make the whole thing circular for Charniak, by argument kl), but potentially at all the information available in the representation-. It is, perhaps, significant that agent is the case that makes Charniak's point best. The other extractions answer it better, for example in the way the I' INST-specialist&quot; creates an extracted Pemplate with action USE; or the way the &quot;GOAL-specialist&quot; extracted an inPS erence &quot;John strikes stags&quot; in a manner independent of the fact that the source action was &quot;fire at&quot;, for the &quot;GOAL-specialist&quot; did not consult the source verb name or its head primitive action CAUSE. Or, again, the way in which common-sense inference rules (Wllks 1973c, 1975a, 197513, but not discussed in this paper) would operate on - any such extracted GOAL template and infer that the corresponding agent wanted the GOAL (whatever it was). And an inference rule like this latter can operate only if-it can locate the corresponding agent, Sob when we extracted the GOAL equivalqrit to &quot;~ohn strikes stags&quot;, we tsould have (in order to operare the new inference rule) to be able to find that John was the agent of the extracted template if we whnted t~ inPS er further that John wanted to strike the stags. The agent (John) wiIl of course be the formula in the first position of the extracted template, and Charniak might then say that that does nok require agency but only access to a first template position. This would be a reversion to the bad position argument (i), and the answer is simply that the mechanism goes to the first positiou because that is where the agent has been put; in just the way that dollars are written before the decimal point and cents afterwards yet that fact does not allow us to dispense with the notions of dollars and cents in financial calculations. However, the important point here is that the &quot;seeker of the agent&quot; would seek it generally: it would not matter whaE the action (striking) in the extracted template was. The real point behind Charniakls position&quot; argument (i) is that no AI systems have such general routines, and hence an agent for one verb could be stored quite differently from that for another because there are no Such routines that would ever find this out! And that</Paragraph> </Section> class="xml-element"></Paper>