File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/91/w91-0212_metho.xml
Size: 30,743 bytes
Last Modified: 2025-10-06 14:12:47
<?xml version="1.0" standalone="yes"?> <Paper uid="W91-0212"> <Title>Redefining the &quot;Level&quot; of the</Title> <Section position="3" start_page="127" end_page="129" type="metho"> <SectionTitle> 2 Morphemes and &quot;Semantemes&quot; </SectionTitle> <Paragraph position="0"> We are led as readers of English, and as dictionary users, to see white space between words as indicating a tangible boundary between units of some sort. We base much of our linguistic analysis on the assumption that there is such a thing as the &quot;word level&quot;. If you go &quot;below&quot; this level of structure you are dealing with morphology or with lists of idiosyncratic information; &quot;above&quot; it is the level of syntax. Under this assumption semantics is also of two sorts: the largely classificatory operations of lexical semantics, and the compositional functions of syntactic structure-building. As a working assumption, this division of linguistic knowledge has some merit, and has allowed much useful work to proceed.</Paragraph> <Paragraph position="1"> But in fact, even on the most superficial examination of real language, it is common to find gross disruptions of this convenient pattern. Unfortunately, lexicalization is an arbitrary process even within a single language, and it is probably only the merest accident when it happens that an atomic chunk of meaning is linked bi-uniquely to an invariant chunk of phonology. The more common run of structure-meaning correspondences are much more complicated.</Paragraph> <Paragraph position="2"> For one thing, combinations of two or more surface words often function synchronically as more or less absolute semantic units. 2 These collocations may be able to be taken apart for syntactic purposes, but semantically they are not decomposable; they link to meanings which are more than, less than, or different from, the sum of the meanings of their individual parts. The commonly cited examples are phrasal verbs ('hold up', 'kid oneself' 'brim over with') and idiomatic expressions of various sorts ('house of cards', 'lie low', 'eat \[someone\] out of house and home'). The difficulties here, including locating the boundaries of true idiomaticity, are much discussed, and I will not dwell on them here.</Paragraph> <Paragraph position="3"> Second, even what are structurally single words can exhibit the kind of compositionality ordinarily expected in syntactic constructions. The most interesting category here is that of lexicalized deverbal compounds, such as clambake, carwash, ill-advised, and timesaving. Roeper and Siegel (1978), and Selkirk (1982), among others, have explored the compositional regularities found in this highly productive component of the lexicon, and found that the construction and interpretation of novel verbal compounds is constrained according to criteria whose canonical field of application is syntax.</Paragraph> <Paragraph position="4"> Both these cross-over phenomena carl be accommodated structurally by means of analytical tricks of one sort or another. We have ways of representing verb-particle combinations, for example, so that they are looked up in the same fashion as structurally single words, even while their parts can be separated by theoretically indefinite stretches of the sentence. It is also possible to install vestiges of verbal structure into the representation of 2This is true even if we leave aside such trouhlesome orthographic conventions as the fact that fully lexicalized compounds are often written as separate words.</Paragraph> <Paragraph position="5"> derivationally deverbal items, so that the regularities which they exhibit in compounding are predicted.</Paragraph> <Paragraph position="6"> A more severe problem than either of these for locating the &quot;level&quot; of the &quot;word&quot; is exemplified by these examples of perfectly legitimate paraphrase (from Talmy, 1985): be of interest to somebody = interest somebody carry out an investigation into something = investigate something Certainly we would want a natural language understanding system to arrive at the same conclusions from either variant in each pair. It is not at all clear, however, whether we should cause the single-word versions to be represented as decomposable into the phrasal expressions, or analyze the phrasal expressions as stylistically or pragmatically inspired expansions of the more succinct variants, or take some other other tack, such as just finding a way to list both variants as meaning the same thing. 3 Even when they are not so naturally expanded to phrasal versions, many--maybe most, depending on the granularity of the analysis-- of what we consider to be monomorphemic words have flagrantly compositional meanings; the most timeworn examples are kill (= 'cause-to-die') and other verbs in which a change-of-state (one meaning component) is brought about by a cause (the other meaning component). Accurate analysis of event descriptions containing such words would seem to require such sub-word semantic analysis, whether it is done as such or as an external listing of presuppositions and entailments of the description as a whole. The trouble, however, is sorting out some set of possible meaning components which is adequate to the compositional task and yet smaller (preferably both in cardinality and in size required for storage) than the set of English words.</Paragraph> <Paragraph position="7"> Compare, for example, the pairs kill / murder and send_away / banish. Both pairs incorporate an additional meaning element in the second word of each pair which is at least somewhat similar in both cases; it would mean something like 'with malice aforethought and with the intention of instituting a permanent solution to a perceived problem and probably done with heightened actional intensity'. It is possible that we could find other pairs of lexical expressions whose only difference in meaning is the presence of this additional element in one of them; this would be a good indication that a productive meaning component was at work. But if this component itself can only be expressed in English by means of a very long phrase, we have a considerable difficulty of representation.</Paragraph> <Paragraph position="8"> It is not that the notions of &quot;word&quot; and &quot;morpheme&quot; make no sense, where the latter is defined as &quot;the minimal meaningful unit of language&quot; (Pei and Gaynor, 1980:140); rather, both are invaluable from the vantagepoint of surface structure. But when we use semantics instead as the starting point, and we define something called a &quot;semanteme&quot;,4 say, as &quot;the minimal unit of meaning that has a sound&quot;, then it is not at all clear that the two constructs would meet in the middle and we would arrive at the same set of words again.</Paragraph> <Paragraph position="9"> Nor is this to say that the expanse of meaning encompassed by a structural word cannot be a unit. It must be true, as Jackendoff (1983, 1990) claims, that linguistic categories correspond fairly directly to conceptual categories. However his formulation of the mapping leaves the status of the structural word itself somewhat indeterminate: Every content-bearing major phrasal constituent of a sentence (S, NP, AP, PP, etc.) corresponds to a conceptual constituent of some major conceptual category. (1990:44) \[emphasis added\] We are still left with the question of the proper representation of lexical semantics in a lexicon when the borders of what is &quot;lexical&quot; are so fuzzy.</Paragraph> </Section> <Section position="4" start_page="129" end_page="130" type="metho"> <SectionTitle> 3 Argument Structure and Valence </SectionTitle> <Paragraph position="0"> The problem of the imperfect mapping between structural words and atomic or even bounded meanings extends into what might have been expected to be one of the simpler aspects of representing lexical semantics, that of the exposition of argument structure.</Paragraph> <Paragraph position="1"> Attempts to assign thematic role structure to a wide range of verbs run immediately into two major problems. One is that, as Jackendoff (1990) points out, it may be only part of the meaning of a verb which is responsible for assigning a particular thematic role to a given argument. For instance, the verb pass (in the construction pass the house) is analyzed as incorporating a Path functor, similar in its function to a surface preposition by, which is part of the meaning of the verb. What appears in surface structure therefore as the object of pass is better analyzed semantically as the object of the underlying by; this is why there is no very satisfactory thematic role label for the object of pass. It is also possible for a verb to assign a role to only part of the referent of an argument. For example, the verb climb, in climb the mountain, has as its Goal not the mountain but only the top of the mountain. In general, Jackendoff's arguments that the standard conception of thematic roles requires an accompanying commitment to lexicai decomposition are difficult to refute.</Paragraph> <Paragraph position="2"> At the same time, thematic role labels, or something like them, are necessary for the description of events whether we are willing to commit to lexical decomposition or not.</Paragraph> <Paragraph position="3"> The fact that the set of labels does not seem to be able to be cleanly inventoried at the level of the structural word does not change the fact that some such set of labels is necessary to let us individuate event types, that is, to distinguish events in which 'X verbs Y' from ones in which 'Y verbs X' (Carlson 1984). While we may not be able to arrive at a satisfactorily comprehensive set of labels without complete semantic decomposition, we must have a sufficient set to be able to extract the basic 'who did what to whom' structure from a sentence.</Paragraph> <Paragraph position="4"> However, finding a boundary between the semantic contribution of the verb itself and that of its expected arguments is often practically impossible. Boguraev and Levin's contention that semantic class determines syntactic behavior, while arguable on other grounds, surely stems from the accurate observation that verbs which describe similar situations naturally tend to govern similar numbers of arguments in similar ways. 5 Par-SUnfortunately, the arbitrariness of lexicalization rears its ugly head here also, and the notion of &quot;semantic class&quot; necessary to make this linkage between syntactic behavior and meaning may be somewhat circular; that is, in doubtful cases will we decide whether an item is a change-of-state verb on the basis of whether it undergoes the causative/inchoative alternation? Carlson and Tanenhaus (1988) would not have been able to construct their experiments comparing the effects of thematic structure ambiguity with those of semantic ambiguity, if there did not exist verbs of the same semantic class which do not share the same possibilities for thematic structure alternation. An example pair from their study is load / filh load / fill the truck (with bricks) load / *fill the bricks ticularly where it is true that a whole semantic class of verbs exhibits similar syntactic behavior, it becomes extremely difficult to separate argument structure as an independent phenomenon from the core meaning,Sf there is such a thing, of the assigning verb. What would it mean to 'eat' without 'eating something' ?--it seems pointless even to consider such a question.</Paragraph> <Paragraph position="5"> Even the definition of individual thematic roles is susceptible to influence from the meanings of expected fillers for those roles; this fact, along with the the difficulty of discovering how these roles are assigned to arguments by word-level functors, may help to explain the persistent lack of consensus on the proper set of thematic roles, as exemplified in Wilkins (1988) and elsewhere. Jackendoff faces one aspect of this problem by removing Agent from the set of primary thematic roles; rather, Agent and Patient are derived role types, designated on a separate Action Tier and superimposed on basic thematic structure through an interaction of the semantics of the verb and the nature of the entities filling potential Agent and Patient role slots. That is, a sentence with kill always implies a killer, but it has an Agent just if the killer is an entity capable of Agent-like activity, for instance a person.</Paragraph> <Paragraph position="6"> Syntax too has an effect on the definition of thematic roles, and that in turn has a pervasive effect on verbal semantics. Talmy (1985) discusses the fact that subjecthood, because of its frequent association with Agenthood, &quot;may tend to confer upon any semantic category expressed in it some initiatory or instigative characteristics.&quot; Consequently a Stimulus expressed as Subject may be seen to be acting purposefully or at least actively upon the Experiencer, while where tile Experiencer is Subject, &quot;the mental event may be felt to arise autonomously and to direct itself outward toward a selected object.&quot; Verbs of experience whose canonical syntactic pattern includes either Stimulus or Experiencer as Subject (for example, please and admire, respectively) are thus under the influence of this effect in one direction or the other; this may be one source of the ambiguity in the sentence 'He pleases me' and hence in the verb please.</Paragraph> </Section> <Section position="5" start_page="130" end_page="133" type="metho"> <SectionTitle> 4 Senses of words </SectionTitle> <Paragraph position="0"> This blurring of what we would rather see as a clear division between the semantic contributions of functor and argument also complicates the proper treatment of word senses.</Paragraph> <Paragraph position="1"> Compare the following uses of the verb introduce: 1. We then introduced a catalyst into the solution.</Paragraph> <Paragraph position="2"> 2. She wants to introduce a little culture into his life.</Paragraph> <Paragraph position="3"> 3. I introduced Pete to green chili / green chili to Pete. 6 4. I introduced Joe to Pete / Joe and Pete.</Paragraph> <Paragraph position="4"> Only load undergoes the locative alternation.</Paragraph> <Paragraph position="5"> Talmy (1985) also gives numerous examples of verbs in which shared semantic class does not guarantee identity of valence. An example is the triple emanate, emit, and radiate. Emanate must have Figure as subject, emit must have Ground as subject, and radiate can have either. Likewise steal, rob, and rip off, which all have Agent prominently in focus as subject, differ in their treatment of Figure and Ground. Steal makes Figure the object, rob makes Ground the object, and rip off can put either one in that position.</Paragraph> <Paragraph position="6"> 5. Ronco has introduced several fascinating new products.</Paragraph> <Paragraph position="7"> 6. Joe introduced the coyote into Santa Fe stores.</Paragraph> <Paragraph position="8"> As the definitions and examples in the ~llowing (partial) dictionary entries suggest, it is common to think of each of these sentences as utilizing a different sense of the verb introduce.</Paragraph> <Paragraph position="10"> <ex>introduce a Bill before Parliament</ex> <def>bring (sth) into use or into operation for the first time</def> <def>cause (sb) to be acquainted with (sth)</def> <ex>introduce nee ideas into a business</ex> <ex>Tobacco was introduced into Europe from America</ex> <ex>The teacher introduced his young pupils to the intricacies of geometry</ex> <ip>introduce sb (to sb)</ip> <def>make (persons) known by name (to one another), esp in the usual formal way</def> <ex>introduce two friends</ex> <ex>He introduced me to his parents</ex> <ex>The chairman introduced the lecturer to the</Paragraph> <Paragraph position="12"> &quot;(often fo111 by to) to present (someone) by name (to another person) or (two or more people to each other)&quot; ,2 ). &quot;(follJ by to) to cause to experience for the first time&quot; ,1 ).</Paragraph> <Paragraph position="13"> &quot;to introduce a visitor to beer&quot; ). &quot;to present for consideration or approval, espl before a legislative body&quot; ,I ).</Paragraph> <Paragraph position="14"> &quot;to introduce a draft bill in Congress&quot; ). &quot;to bring in; establish&quot; ,1 ).</Paragraph> <Paragraph position="15"> &quot;to introduce decimal currency&quot; ). &quot;to present (a radio or television programme, etcl)verbally&quot; ,1 ).</Paragraph> <Paragraph position="16"> c_DEF(\[ &quot;introduce&quot;,1,1,6 \], &quot;(folll by with) to start&quot; ,I ). c_SAMP(\[ &quot;introduce&quot;,l,l,6 \], &quot;he introduced his talk with some music&quot; ). c_DEF(\[ &quot;introduce&quot;,l,l,7 \], &quot;(often fo111 by into) to insert or inject&quot; .I ).</Paragraph> <Paragraph position="17"> c_SAMP(\[ &quot;introduce&quot;,1,1,7 \], &quot;he introduced the needle into his arm&quot; ). oo * ooo However the example sentences (1)-(6) seem to illustrate not separate senses but a natural continuum of meanings, in which the contribution of introduce itself does not perceptibly change. In sentence (1), introduce comes closest to what we would probably call its core meaning, something like 'bring / put something into a new place'. Sentence (2) relocates this action in a different kind of space; both the Theme 'something' and the Goal 'new place' are abstract in this case. 7 In Sentence (3), the Goal location is some sort of recognition space, associated either with the recognizer or the subject matter; the effect is that Pete now knows about green chili. Obviously this sense extension for introduce only works if at least one of its arguments is an entity which is capable of recognizing, i.e. a person. Moreover, where the introduce-ee is also a person as in (4), (or where the single Object of introduce refers to more than one person) the introduction becomes reciprocal; the &quot;actional content&quot; of the verb, in Talmy's words, has been doubled. In (5), the meaning extension present in (2), (3), and (4) is further specialized so that the implicit Goal location is not only a &quot;recognition space&quot;, but is understood to be the recognition space belonging collectively to the individuals making up a market. We make this additional extension, however, not because we are now using a different sense of the verb introduce, but simply because of what we know about products. That is, our understanding of the type of space in which the introduction is effected is governed by both the core meaning of the verb--we know introduce means putting something into some space just as we know that eating means eating something--and the semantic content of its first argument.</Paragraph> <Paragraph position="18"> In (6) the inference from 'product' to 'market' is reversed. In the most plausible reading of the sentence, the coyote refers generically to a decorative motif or to some kind of typically salable item, not to an actual animal. This is true because we know that Santa Fe stores represents a particular kind of market, and so we prefer an interpretation in which the thing introduced is something which can typically be marketed. If we try to decide then what the senses of introduce are, as the dictionary model would suggest that we should do, we can take several tacks. We could just take a dictionary's word for the number and definitions of separate senses (leaving aside for the moment the problems of mapping between the different sense divisions of different dictionaries), but then we have clearly missed a generalization about the extensions of meaning possible for the verb introduce and others like it. We could distinguish separate senses on the basis of the use in some of the examples of into and in others of to, but that distinction cross-cuts what are clearly more important semantic distinctions.</Paragraph> <Paragraph position="19"> On the basis of the thematic structure that the verb introduce instantiates, which is more or less the same for all six examples, we might say that it has only one sense, s 7This is a legitimate and extremely common meaning extension, from physical space to abstract space; in fact it is exactly the sort of regularity that a LKB should include among its lexical rules. SThe one problematical construction here would be the one with a plural Theme and no explicit Goal, as in 'I introduced Joe and Pete.' In this case, as noted above, the introduction is reciprocal and therefore the thematic structure would apparently have to be copied over into another, simultaneous action. Likewise it would decompose in every case to something like</Paragraph> </Section> <Section position="6" start_page="133" end_page="133" type="metho"> <SectionTitle> CAUSE (GO (X (FROM (outside-some-space))(TO (inside-some-space)))) </SectionTitle> <Paragraph position="0"> Under this analysis it is the characteristics of 'X', the locatum, and of the respective entities which serve as Goal that differentiate senses; sentence (1) makes reference to a real thing and a physical space, (2), (3), and (4) refer to recognizable entities and individual cognitive space, and (5) and (6) refer to something which can be sold and a particular kind of collective cognitive space. In (5), naming a product allows us to infer a market as the implicit Goal, while in (6) knowing that the Goal is a market lets us know that the coyote is to be construed as a product. What seems to be happening is what Cruse (1986) calls &quot;modulation&quot; of a word sense by context, except that in this case the &quot;context&quot; is an integral part of the argument structure of the verb.</Paragraph> <Paragraph position="1"> In any case, it is clear that splitting the verb introduce into a set of separate senses at the word level will be arbitrary in one way or another.</Paragraph> </Section> <Section position="7" start_page="133" end_page="135" type="metho"> <SectionTitle> 5 Words as Worlds </SectionTitle> <Paragraph position="0"> The ideal, but computationally outlandish, solution to this problem would be to represent every word in English (not just each functor, but every word) as the union of all the situations in which it could potentially participate, so that all the combinatorial potential of each word would be an inherent part of its inventory of meaning(s). Under this system, a structural word would just be the entry point into a rich representation which might or might not observe that bit of structure further, being built instead around the way the word works in combination with others. All lexical categories, not just the prototypical functors, would have combinatorial preferences, as McCawley (1986) suggests. Semantic composition of functor and arguments in a sentence then would consist of unifying these representations in the most cognitively perspicuous way, in accordance with the bounds imposed by syntax, to create a single coherent scene. In the process of unification, gaps would be filled, construals or views chosen in accordance with constraints instantiated by specific lexical interactions and general cognitive conventions, and aspects of the unifying representations which are contradictory or not currently in focus would fall away. A new semantic combination constructed in this way would be acceptable in inverse proportion to the strain which constructing it placed on these general conventions, and selectional restrictions would simply be statistically based thresholds of strain.</Paragraph> <Paragraph position="1"> Various efforts have been made toward constructing a realistic version of this ambitious model. Raskin (1986) discusses a system he calls Script-based Semantics, which captures many of the ideas described above. In his system, the appearance of a word in a sentence being analyzed evokes both a set of syntactic senses and one or more scripts, where a script is defined as &quot;a large chunk of semantic information surrounding the word or evoked by it ... a cognitive structure internalized by the native speaker, \[representing\] knowledge of a small part of the world. ~ These scripts are represented as graphs with lexical nodes and semantic relation arcs, thus bearing a strong resemblance to tile Conceptual Structures of Sowa (1984). Under both schemes a set of combinatorial rules then unifies graphs in such a way that ambiguity between word senses disappears and a complete representation for a sentence is constructed.</Paragraph> <Paragraph position="2"> What is not clear in either system is where the graphs for individual word senses come from. If we were to use the dictionary definitions given earlier as a source of the script(s) or graph(s) for the verb introduce, in combination with the following entries for product, we would still have absolutely no basis for combining them to mean what we want them to mean in examples (5) and (6). , (from OALD3E) 7296955, . .<ent h=product> <def>sth produced (by nature or by man)</def> <ex>farm produces</ex> <ex>the chief products of Scotland</ex> <def>(maths) quantity obtained by multiplication</def> <def>(chem) substance obtained by chemical reaction</def> * degdeg, degdeg</Paragraph> <Paragraph position="4"> .,.,, .</Paragraph> <Paragraph position="5"> , &quot;something produced by effort, or some mechanical or industrial process&quot; , 1 ). , &quot;the result of some natural process&quot; ,I ). , &quot;a result or consequence&quot; ,I ). , &quot;a substance formed in a chemical reaction&quot; ,1 ).</Paragraph> <Paragraph position="6"> \], &quot;the result of the multiplication of C/.o or more numbers, quantities, etc&quot; ,I ). \], &quot;another name for intersection ,1,3&quot; ,I ). Hand-building such scripts on a sense-by-sense basis for large-scale efforts, on the other hand, would require not only massive redundancy but also an enormous amount of sophisticated-yet-tedious lexical analysis. Not only that, but by pegging these scripts or conceptual graphs to individual word senses, we return to the problem of knowing how to arbitrarily distinguish those senses.</Paragraph> <Paragraph position="7"> The most promising answer to many of these problems can be found in the approach of collocational semantics, in which the &quot;meaning&quot; of a word consists of sets of observed uses of that word. Practitioners of this approach have sometimes apologized for it (see for instance Pazienza and Velardi, 1989) as a practical but unprincipled substitute for &quot;real&quot; semantics. But in fact the collection, classification, and normalization of collocational patterns may constitute the most realistic way to sidestep various serious difficulties involved in other methods of building a large inventory of lexical knowledge; it might be the way to discover what it &quot;means&quot; for a text to talk about the introducing a product. Collocational semantics provides an objective methodology for detecting functional equivalence between expressions, which is really what we want when we attempt to encode synonymy, and for pinpointing functional distinctions, which is what we are trying to do when we attempt to distinguish word sense meanings. Above all, this approach offers promise precisely because it does not depend crucially on the level of the structural word for the definition of semantic units.</Paragraph> <Paragraph position="8"> As for the operation of a language-understanding system including such a lexical store, obviously allowing the appearance of each word in a sentence to immediately evoke everything we know about its uses would be a computational nightmare. Instead, what is required is a mechanism for incrementally introducing lexical information into sentence analysis in such a way that just enough, and no more, is present at any given stage, after the fashion of ltirst (1987) with his &quot;Polaroid&quot; words. Under such a methodology several stages of lexical lookup would be necessary. Syntactic patterns, morphologically-derived categorizations, and minimal thematic structure, as discussed above, might be available in the usual way at the first level. The matching of partially disambiguated, intermediate sentence graphs constructed at the first level against an inventory of highly schematic &quot;entry&quot; graphs, and then against successive layers of more completely specified model graphs, would funnel the analysis through the next levels of lookup. 9 A system of semantic classes, along with a set of redundancy rules encoding systematic relations between and alternations possible for members of those classes, would constitute yet another level of lexical knowledge; this level would be important in the organization of the lexicon, but might be accessed directly during lookup only ill the case of the appearance of previously unknown words.</Paragraph> <Paragraph position="9"> How would we construct this kind of multi-level lexical knowledge base? The first level described above can and should be constructed on the basis of information from machine-readable dictionaries. The next set of lookup levels described, in which the structural word is no longer the primary entity but just a handle for indexing graphs, can be built on the basis of knowledge assembled from large volumes of text; the methodology employed by Smadja and McKeown (1990) would be one of many techniques possible for obtaining and organizing the various levels of model graphs. Our idea is that the graded levels of specificity would be constructed by means of successive generalization operations over the most specific set of graphs, extracted more or less directly from text. The final level of knowledge--that is, what Boguraev and Levin have found missing from standard LKBs-must still be installed by linguists. Dictionaries do contain some clues, and certainly the ability to systematically and automatically pull together usage instances can help; but the real work must be done by people.</Paragraph> </Section> class="xml-element"></Paper>