File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/86/c86-1002_metho.xml
Size: 27,480 bytes
Last Modified: 2025-10-06 14:11:47
<?xml version="1.0" standalone="yes"?> <Paper uid="C86-1002"> <Title>An Empirically Based Approach Towards a System Of Semantic Features</Title> <Section position="2" start_page="0" end_page="0" type="metho"> <SectionTitle> ATINEAVE, FRED~IKSEN, MINSKY, (~N/AK, WINOGRAD, ANDERSON & BOWER, WOODS) and last but not least </SectionTitle> <Paragraph position="0"> scme recent issues on word semantics (T. BAIIMER, W. BRENNENS~IHL, J.BALLWEG, H. FROSCH) have been taken into account in order to meet the requiremerits of a manageable syst~n of semantic features. This system is intended both to be based on a sensible theory of semantics and to satisfy the special requirements of machine translation in general and of our t~t type in particular. Moreover, it should be flexible enough to be enlarged and supplemented or changed, whenever this proves necessary on empirical evidence = this last requir~rent being made possible by the accc~plishment of the first.</Paragraph> <Paragraph position="1"> In tmying to meet the~ requirea~nts the semantic feature inventories ~t hand have been enlarged, changed, and adapted to our specific purposes m~d have been merged into one system of semantic features.</Paragraph> <Paragraph position="2"> 2. Comment on the theoretical asslmi0tions made in different machine translation systems with respect to the semantic representation with respect to the semantic representation which in EUROTRA will be implemented on the interface structure (IS) level of the source and target language it is our first and foremost aim to arrive at a coherent system of semantic features. In order not to start from nothing the above mentioned feature inventories have l~en consulted. The feature inventories developed fer these machine translation systems have different shortcomings which will be briefly comm~%ted on in the following.</Paragraph> <Paragraph position="3"> since a sufficient def:hnition of how to interpret the features is given J/% none of the proposals of the above mentioned mac/line translation systems, we will not ccm~ent here on the ~eatures the~nselves included in the propo~ils. A brief ccsment, however, is necessary on 'the general approach, which seems to imply theoretical assumptions (not explicitly mentioned, since neither a theoretical nor a practical usage-based ~101anation is given) about the organisation and processing of semantic units, for which there is no e~pirical evidence: neither natural language processing by human beings nor efficiency in automatic processing of natural language gives support to i:~ese implied assumptions. It must be mentioned, however, that this can by no means be considered to be an objective comment, since for an outsider, it is ~/~oossible to urgerstand the systematic aDtivation of these feature inventories for at least one of the following reasons: - The semantic features are not defined or at least not sufficiently defJmed in order to n~ke clear their conceptual structure and thus to make clear how they are meant to be used. This is especially true for the EUROTRA proposal, in ~lich semantic features are not defined at all. rfhis is, however, only a proposal, which has not been applied yet, but is being tested at the. mcment. But also the SYSTRAN semantic features, as well as those of JAPAN, which have been worked out rather sophistieatedly, are not cc~_nted on. The semantic features of METAL are defined, their definition, however, remains rather vague. Even when taking into consideration the ~les which are added, the reader does not arrive at a satisfactory un- dersta~. - The dependencies holding between features are. not explained. This is especially true for SYSTRAN, which only gives a list of features z~ferring to arguments. A hierarchical system consisting of two levels of semantic features is defined by METAL, which is far frc~ sufficient. JAPAN J s worked out in a more sophisticated way with respect to this problem. Both in METAL and in JA-PAN, however, relations between the dominating features are not defined. The ~ proposal gives an enumeration on the second and lowest le~ vel of the feature tree, which is just a conglon~ration of semantic information, which should be described at different levels~ in order to achieve the overall aim of linguistically consistent semantic description.</Paragraph> <Paragraph position="4"> 3. A proposal for a EUROTRA semantic feature rule system 3. i. Necessity of a semantic feature rule system Let us now put forward our conception of the ~o system~ ef ~tic features with respect to its fom~alization. We have two gr~, one describing &quot;SITUATION&quot; features, the other one describing &quot;~TITY&quot; features. Neither of the two systems is strictly hierarchically organized. The hierarchical principle, however, which always defines a refinement of the doafinat~/~g feattzt~, prevails. Particularly the most general semantic features, such as the &quot;ENTITY&quot; features &quot;CONC~ETE/&quot;ABSTRACT&quot;, &quot;CC~JN-TABLE&quot;/'9~ASS&quot;, and &quot;NAIURAL&quot;/&quot;ARTIFICIAL&quot;, and the &quot;SITUATION&quot; features &quot;OONCREI~&quot;/&quot;ABSTRAC~&quot;, &quot;STA-TIVE&quot;/&quot;DYNAMIC&quot;, and &quot;H/NCIUAL&quot;/ &quot;DURATIVE&quot;/&quot;ITE-RATIVE&quot;, respectively, form pairs or triplets of semantic features. One feature of each of these alternations obligatorily occurs, and the descendents, which specify them, form disjunct sets.2 3.2. The basic formalis~a Let us now comment informally on our present conception of how the semantic features which we consider necessary so far are related to each ot2ner. We use three operations holding between semantic features in our C//rammar: l) Hierarc/Iy is the overall relation defining the derivation of the features.</Paragraph> <Paragraph position="5"> 2) Alternation relates a set of features, only one of which applies.</Paragraph> <Paragraph position="6"> 3) ~ relates semantic features obligatorily occurrJ/~g together. This type of ralationship is of course in the minority.</Paragraph> <Paragraph position="7"> The basic idea is to describe these relations by a context-free ~ule system, where the rules can for example be of the following, folul~:</Paragraph> <Paragraph position="9"> The hier~ly here is repz~sented by t/~e sign &quot;=&quot;, the alternation by the sign &quot;/&quot;, and the disjunction by the sign &quot;*&quot;. The interpretation of the rule is the following: The feature on the left handside of the rtlle dc~ minates the features appearing o11 the right handsidedeg A, B, C, and D establish a refinement of X.</Paragraph> <Paragraph position="10"> Mere precisely, in this example X is specified by a pair of features, the first ~nent of which can be either A or B and the .~econd is either C or Do ~he subordinate feattlre~ on the right llandside of the tulle can get supe~rdinate feature~ themselve~ o11 the next level lower dowel in the hie/-archy. ~e terr~inal featuz~s~ t/iat is those features which are not defined for accepting a/~y ~ulbordinat~ features, are rep ~resented by the rules</Paragraph> <Paragraph position="12"> \[~t us exemplify this with the feature &quot;OODNTABIE&quot;: (3.3) CODNTABLE~ = CATEGORY 1 * CATEGORY 2 * DEFINITION ~RY 1 = INDMDUATIVE/P~E/~ CATEGORY 2 = CAUSE/RESUI~</Paragraph> </Section> <Section position="3" start_page="0" end_page="11" type="metho"> <SectionTitle> DEFINITION = MEASURE/SOCIAL </SectionTitle> <Paragraph position="0"> By this we mean that the feature &quot;COUNTABLE&quot; is represented by three features which always occur together (marked by the operator &quot;*&quot;). Each of these three features again dfmd/lates a collection of features only one of which is selected (marked by the operator &quot;/.). The hierarchical relationship itself is implied in the left-to-right-handside associations (marked by the operator &quot;=&quot;)deg Here it is essential to note that every semantic featxlre is only defined once by one ruleo If more than one description exists, all of them are cc~bined by &quot;or&quot;. As the &quot;and&quot; relation by definition is prior to the &quot;or&quot; relation, brackets have to be placed around the alternative expression J.n the opposite case, that is when the &quot;or&quot; relation is prior to the &quot;and&quot; relation.</Paragraph> <Section position="1" start_page="0" end_page="10" type="sub_section"> <SectionTitle> 3.3. The introduction of attributes </SectionTitle> <Paragraph position="0"> So far we have introduced a formal instrtwaent with which we can describe the relations be~ features which are formally possible. In order to describe the actual relationships between features, this formal\[ instrument still has to be restricted.</Paragraph> <Paragraph position="1"> In order to keep the ru/e system cc~pact, we intro~ duced attributes which are intended to describe ~mportant o~x~zurz~nce restrictions existing between features in disjunct bra/ic/%es. The existe/3c~ of a feature activates an attribute called the derived.</Paragraph> <Paragraph position="2"> This attrik~te effects the restriction of a rule application in a disjunct part of the grammar.</Paragraph> <Paragraph position="3"> There the attribute is called the inherited. In the rule syst~n attri~\]tes appear on the left l~hndsJ.de of the rule if they are derived, on the right har~side if they are inherited. We derive a feature's attribute llke that: (3.4) x\[\] ....</Paragraph> <Paragraph position="4"> An attribute always gets the name of the semantic feature which causes the attribute, so the derivation can be marked by an eanpty pair of square brackets.</Paragraph> <Paragraph position="5"> The derived attribute appears in the right-handside context as h~herited attribute e.g. like that (3.5) Y .... A\[ x\] .., With the above mentioned example this would look as the following:</Paragraph> <Paragraph position="7"> The inherited attribute can also be assigned to a feature expression. In this case it ~ould apply to every feature within this expression. Moreover, instead of a single attribute, an expression of at= tributes can appear. In an attribute expression the above mentioned operators &quot;or&quot; and &quot;and&quot; can appear and in addition the n egatien operator &quot;not&quot; (represented by the sign \). With the introduction of at= tri~mtes the generation mentioned above has to be ~odified: the rule (3.4) states that the feature X is derived and has to be registered so that it can be used in the relevant disjunct feature context which may also be dcminated by Y as described in rule (3.5).</Paragraph> <Paragraph position="8"> We have therefore to extend the above mentioned example (3.6) by the following rule: (3.7) DIREC21ON&quot; -- SOCrRCE \[ CAUSE \]/GOAL \[ RESULT \] with the definition of these ~tles it was proved that on the one hand the formalism is powerful enough to ~present all the above mentioned phenomena and on the other hand it is still simple enough so that changes necessary in later stages may be accomplished without too much oost.</Paragraph> <Paragraph position="9"> Our rule syste~l is based on the definition of the semantic features. So far we have defined 87 f~tures for the description of &quot;ENTITIES&quot; and 87 leatures for the description of &quot;SITUATIONS&quot;. 3.4deg The use of the formalis~, Our grammar is intend~ to be a generating system.</Paragraph> <Paragraph position="10"> It will be used as inl~it for an automatic proceduredeg For every lexical unit this can be used to generate the list of semantic features which semantically describe the lexical unit sufficiently.</Paragraph> <Paragraph position="11"> Our notion of sufficiency arises frc~ our goal of autcmatic disambiguation. The autcmatic procedure leads the lexicographer through the system in the right way, so that the correct list of semantic features is generated for each lexical entry. This procedure m~es use of the rule system in order to produce menus which show the alternatives valid in each actual state. In general the list of sea~nntic features which descrJl~s a lexical unit conta~s only the terminal features which are generated, since the dominating r~n-terminals can be deduced.</Paragraph> <Paragraph position="12"> %his is, however, not valid for features and their derivates, which appear on more than one right handside of a rule (named critical rule), i.e. the resulting terminal features do not give an unambiguous specification of the lexical unit. In these cases we add the non-terminal feature from the left handside of the critical rule to the feature list, which then gives us an adequate feature spectr~.</Paragraph> <Paragraph position="13"> It is, however, possible to take into account redundancies by. taking other dc~tinating features into the list as well. This would possibly lead to a more efficient translation process.</Paragraph> <Paragraph position="14"> This output of our g~i~erating systmu will be the il~put to our dictionary and in later stages of the translation process, precisely on the interfacestructure level, is intented to be used for di~mbiguation and other strategic purposes in the process of semantic analysis and synthesis.</Paragraph> <Paragraph position="15"> It follows that all lists of features which can be generated by the grammar make up the set of all possible semantic descriptions which may describe concepts referring to our object world.</Paragraph> <Paragraph position="16"> Moreover, this rule system may be used in synthesis in order to find out dependencies holding between semantic features. We think that this will be necessary for semantic generalizations in the target language. ~lis dc~k%in, however, has not yet been worked out.</Paragraph> <Paragraph position="17"> A side effect of the autc~atic processing of the rule system is the generation of a graphic representation which has a i~eelike form.</Paragraph> <Paragraph position="18"> The graphic representation of this rule system has proved to be very useful. In this graphic representation the axiem is the root and every rule is represented by a subtree as shown in Figure i.</Paragraph> <Paragraph position="19"> If a ~Kx~e generates a derived attribute, an ~y pair of brackets appears as a subscript of the feature denoting the node. If the node inherits another attribute or an attribute expression, the corresponding names or expressions are represented as superscripts of the feature. Figure 2 represents the attributes and their effect in an evident way.</Paragraph> <Paragraph position="20"> It makes clear that the order of the disjunctive nodes (branching underneath the &quot;and&quot; node) is most essential for the application of the attributes.</Paragraph> <Paragraph position="21"> This representation in the form of a tree has proved to be a very transparent way of illustrating the structure of the set of semantic features.</Paragraph> <Paragraph position="23"> The arcs represented by asterisks correspond to a disjunctive expression, the arcs represented by a solid line correspond to a conjunction.</Paragraph> <Paragraph position="24"> Figure 1 The feature &quot;(X~JNTABLE,' as an exar~le of disjunctive branching</Paragraph> <Paragraph position="26"> The dotted line which relates the features &quot;CAT 2&quot; and &quot;DIRECTION&quot; represents that both features are dominated by the same node higher up in the hierarchy. The dashed line shows the derivation and inheritance of attributes.</Paragraph> <Paragraph position="27"> Figure 2 Illustration of the derivation and inheritance of the features &quot;CAUSE&quot; and HRESULTV! 4. The linguistic motivation for the specific make-up of ottr rule system of semantic features Now that we have given a description of the for ~ realism that we have made use of in order to describe the existing relationships between semantic features, let us explain at this point, why semantic features are organized like that, since it is less the formalism, but rather empirical evidence and linguistic knowledge by means of which we arrive at exactly this organization.</Paragraph> <Paragraph position="28"> Although it is essential to know that there are 9o inherent but only context-dependent features ~, apart frc~ the features &quot;NAqXJRAL&quot; and &quot;ARTIFXCIAL&quot;, the basis of our system of semantic features is an as objective as possible a definition of each feature itself. This definition is based on the crite~ ria of prototypicality by means of which we abstract fro~ our experience. Moreover, the criteria of prototypicality resulted in disjunct and alternative feature sets, which are described by our generative rule system. This means that the whole system is based on how we categorize concepts. The general process of refinement into different sub-features on which our systems are based depends on the principle of focusing diff~t areas of the superconcept and thus imaging different subconceptsdeg As one general characteristic of the system we stated above that the alternative branchings are in the majority, since in most cases the system defines a refinement of strpex~rd/nate features into subordinate features. As the other general characteristic we stated the disjunctive branchinq of the root node. We can explain this &quot;and&quot; relationship between the dominating features of the system by how we conceive of our environment. According to gestalt psychology this proceeds at least according to the following two principles holding for the perception of &quot;~glTITKS&quot; and &quot;SI~JATIONS&quot; respectively. These two principles correspond to the &quot;and&quot; relationships dominated by the root.</Paragraph> <Paragraph position="29"> On the one hand the definition of concepts depends on whether our oonceptualization of &quot;D~ITITES&quot; or &quot;SIZ~IKTIONS&quot; is based more or less directly or indirectly on our sensory perception. ~he former case in which concepts are abstracted directly on the basis of sensory perception holds for &quot;CONCRETES&quot;, the latter case ~f //xlirect cor~ization holds for &quot;A~STRACTS&quot;.</Paragraph> <Paragraph position="30"> In the case of &quot;CONCRETES&quot; no higher order information processing takes plaoe, because there are no parts for which an elaborate substructure has to be reconstructed. Moreover the peroeptual properties remain fairly constant between exemplars, so that they are easy to reidentify. Just the opposite holds for &quot;A~STRACT,' conoepts.</Paragraph> <Paragraph position="31"> On the other hand we either define concept~ by their outlines or by their inner configuration, u In the former case, in which ~ definition depends on the more or less sharp entlines of the &quot;ENTITIES&quot; or &quot;SITUATIONS&quot;, we conceptualize &quot;COUN-TABLES&quot; in &quot;the case of &quot;ENTITIES&quot;. In the case of &quot;SITUATIONS&quot; we conceptualize &quot;PERFECTIVE&quot; &quot;ACTIONS&quot; or &quot;EVICTS&quot;, that means, either &quot;SITUATIONS&quot; for which a terminal phase is expected, which holds for &quot;A~~&quot;, or &quot;SITLrATIONS&quot; which are just in ~he terminal ~hase, which holds for &quot;AfXTfEVF24ENTS&quot; and &quot;EVENTS&quot;. u In this case the bou~azy of the cencept can be defined in ~ of a terminal ~int or phase of the situation.</Paragraph> <Paragraph position="32"> In the latt~ case in which the ~/tlines are indistinct, we define &quot;~NTIRT_ES&quot; by means of their inner configuration as different subcategories of &quot;MASS&quot;. Correspondingly we define &quot;SITUATIONS&quot; as different subcategories of &quot;IMPERFFL~IVES&quot;, if the situation is focused without reference to its terminal point or phase, that means as either ',ACTI-VITY&quot; or &quot;PROCESS&quot; respectively or as &quot;STATIVE&quot;. The third 'Uand&quot; relationship of our rule systems cannot be explained by the same cognitive principle both for ,'F/qTITIES&quot; and &quot;SITUATIONS&quot;, though it is obligatory for both. Only the obligatoriness of the situational &quot;and&quot; relationship can be made evident by cognitive principles. This third &quot;and&quot; relationship of &quot;SITUATIONS&quot; is the perception of their procedur~l c/qaracteristic~, which is precisely the &quot;AETIONSAEI'&quot;o Depending on whether it is &quot;VUNCILrAL&quot; or &quot;DURATIOn&quot; or &quot;~&quot;, the 'IAktionsart&quot; cc~ bines in a definite way with aspect, which can either be &quot;P~RFECTIVE&quot; or &quot;IMPERFECrlI~&quot;. Now, both &quot;PERFECIXV~;&quot; and &quot;~&quot; can take the sub-category &quot;CAUSATIVE&quot; whereas the other subcatogoties of both aspects branch into disjunct feature sets, the refinement being defined by the &quot;is&quot; relationship and by the iltheritance of attributes.</Paragraph> <Paragraph position="33"> Here the manifold branching of the &quot;PERFECTIVE&quot; aspect into &quot;M\[~ATIVE&quot;, &quot;INONOATIVERE&quot;,&quot;REVERSATIVE&quot;, &quot;~TIV~\]&quot; and also &quot;CAUSATIVE&quot; and the inheritarK~ of the semantic featttre~ &quot;AC~T&quot;, &quot;AC-COMPLISHMENI~ 'o, and 0'EV~qT deg' are remarkable, whereas the &quot;IMPERFEClUS~&quot; aspect, apart frcm the possibility of taking the subcategory &quot;CAUSATIVE&quot; only inherits the features &quot;PROCESS&quot;, &quot;ACTIVITY&quot; and &quot;STATIVE&quot;. 'P~is is the reason for sympathizing with GAIIDON (1964.140f.), who pleads for considering the ',PERFECTIVE&quot; as the unmarked aspect since it &quot;represents our normal scheme of arranging our perceptions&quot;. In using the &quot;IMPERFELTIVE&quot; we create an artificial stability by stopping the procedure of situations and thus making them timeless, whereas the procedural arrangement within time is usually considered as the urmmrked case of &quot;SIEUATIONS&quot;.</Paragraph> <Paragraph position="34"> with &quot;~rrl~ES&quot; the third &quot;and&quot; relationship which branches frcm the root of our grmmmu: is the alternation between &quot;NA_~URAL$&quot; and &quot;ART~FICIAIS&quot;. We have thus shown how on the basis of empirical work two sFstems have grown independently of one another, one for &quot;ENTITY&quot; features and one for ,'S~ON&quot; features, which both have the same numbet of disjunctive ar~ descending from the root node. And what is even more interesting and corroborates our systems is the fact that two of the three disjunctive arcs of both systems can be explained by the same cC/~/nitive principles, which also are obligatory in the process of oonceptualization. null</Paragraph> </Section> <Section position="2" start_page="10" end_page="11" type="sub_section"> <SectionTitle> Notes </SectionTitle> <Paragraph position="0"> 1. I want to express special thanks to Angelika MUELIER-v.-BROCHOWSKY for programming the granmmr and for valuable suggestions.</Paragraph> <Paragraph position="1"> 2. This conclusion is not our private impression. A look into the literature on semantic feature networks shows that they are generally organized like this: the daRinating ncx~es of the network are related by disjunction, whereas the features lower down in the network a~i~ rather related by alternation; that is, they are more strictly hierarchically organized (cf. e.g. WOODS 1975) 3. Especially in order to cope with the manifold semantic problems when coding lexical units one cannot ignore this fact. BARSALOU (1982) has tested and verified the existence of two types of concepts: context-independent and context-dependent concepts associated with verbal expressions. The results of his investigation make him conclude, that context-dependent properties have a major role in the definition and c~tablishment of meaning, as they are also responsible for changes in the accessibility of context-independent properties (cf. ebd. p.92).</Paragraph> <Paragraph position="2"> 4. This definition of &quot;CONCRETE&quot; ~atches the GIBSONIAN theory of &quot;di~.x~t&quot; perceptiondeg 5. This principle again holds for &quot;ENTITIES&quot; and &quot;SITUATIONS&quot; respectively. Among &quot;ENTITIES&quot;, there are e.g. tables, books, knives, wars for which we can image rather definite and clear outlines, by means of which they are limited against their envirorm~nt, either as &quot;OONCRETES&quot; by a definitely shaped limitation of ~m~terial or as &quot;ABSTRACTS&quot; by the limitation of a definite phase structllre of a &quot;PROCESS&quot; or &quot;ACTION&quot;. In English the possibility of pluralization indicates that thus cencept~lized entities are &quot;COUNTABI~.~,,. Among situations there are &quot;DYNAMIC&quot; &quot;SITUATIONS&quot; like She wrote a letter yesterday or The avalanche rolled down the mountain, which are also imaged as having a definitely limited phase structure, that means as a &quot;PROCESS&quot; or &quot;ACTION&quot; occurring in a definite order and ending in a definite, i.e. expected way. ~lis should explain how we :hnage &quot;CCLRqTABLE&quot; &quot;ENTITIES&quot; and &quot;DYNAMIC&quot; situations by the same cognitive principle.</Paragraph> <Paragraph position="3"> The opposite of such a definite and sharp limitation towards the environment is the imagination of an amorphous mass, which is less precisely defined for its inner configuration and thus not at all for any definitely shaped limitation. 'ITLis is the case with '%~ASS&quot; entities like the &quot;CONCRETE&quot; sL~0stances water and o~ or abstract &quot;SI~JA~&quot; like information, inflationdeg This is also the case with &quot;ACTIVITIES&quot; and &quot;PI%0CKNS~&quot; like Yesterdaff sh__@e Dainted or The mast was shaking in the wind and even more so wit2~ &quot;STA~IVES&quot; like During the week she e~ u_D a_t seven or This mast shakes ~ the wind.</Paragraph> <Paragraph position="4"> 6. Refer to LANGACKER 1984. For the differentiation of &quot;ACTION&quot; &quot;SITUATIONS&quot; into &quot;ACTIVITY&quot;, &quot;ACCOMPLISHMENT&quot;, and &quot;ACHIEVEMENT&quot; refer to VENDLER who has introduced this classification. For the distinction between &quot;PROCESS&quot; and &quot;EVENT&quot; cfo e.g. BRANSFORD & McC/iRRELL for their criteria. See also LYONS 1977.483 and MILleR & JOHNSON-IAIRD 1976o85ffo 7. Refer to IANGAClq~R \]984</Paragraph> </Section> </Section> class="xml-element"></Paper>