File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/88/c88-1055_metho.xml
Size: 23,876 bytes
Last Modified: 2025-10-06 14:12:06
<?xml version="1.0" standalone="yes"?> <Paper uid="C88-1055"> <Title>Concretion: Assumption-Based Understanding</Title> <Section position="1" start_page="0" end_page="0" type="metho"> <SectionTitle> Artificial Intelligence Program GE Research and Development Center Schenectady, NY 12301 USA Abstract </SectionTitle> <Paragraph position="0"> A language understanding program must produce as precise a meaning representation as possible from a linguistic input. CONCRETION is the process of developing a specific interpretation by combining various levels of conceptual information. This process represents an assumption-based method of language interpretation, and departs from the traditional approach of treating multiple interpretations as independent. Concretion Mlows the language analyzer to develop a sufficiently specific representation without excessive computation or brittle interpretation rules.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="metho"> <SectionTitle> 1 Introduction </SectionTitle> <Paragraph position="0"> The ambiguity and imprecision of language are the key problems in building language understanding programs. Most systems that perform semantic interpretation \[Bobrow and Webber, 1980, Sondheimer et al., 1984, Lytinen, 1984, Hirst, 1987\] address this imprecision by offering means of selecting among alternative interpretations. The problem with these approaches is that they fail to take into account the interrelationshiPS among the interpretations, which often support or refute one another to various degrees. A better model is one in which the candidates exist not as distinct choices but as assumptions contributing to a complete meaning representation.</Paragraph> <Paragraph position="1"> The language understanding process thus gradually refines a semantic representation based on the support or refutation of each element.</Paragraph> <Paragraph position="2"> For example, consider the following potential inputs: 1. John cut the salami.</Paragraph> <Paragraph position="3"> 2. John gave a kiss to Mary.</Paragraph> <Paragraph position="4"> 3. The investor group seeking control of Warnaco...</Paragraph> <Paragraph position="5"> 4. The ~rm' command takes three arguments.</Paragraph> <Paragraph position="6"> 5. Move back to the last position.</Paragraph> <Paragraph position="7"> The examples above represent potential inputs from several disparate real and &quot;toy&quot; domains of TRUMP (TRansportable Understanding Mechanism Package)\[Jacobs, 1986b, Jacobs, 1987\]. The idea of TRUMP is to use a combination of &quot;core&quot; knowledge about language and certain specialized knowledge to produce a complete semantic interpret~ation. In each of the examples, the italicized word or phrase represents a vague, ambiguous, or metaphorical verb sense. The problem for a good semantic interpreter is to derive the real or intended sense of each phrase without excessive computation or specialized knowledge. For example, the following are reasonable paraphrases of a complete semantic interpretation of the above examples: 1. John cut the salami cut =~ sliced (NOT chopped or shortened) 27oi 2. John gave a kiss to Mary gave a kiss ~ kissed (NOT presented) 3. The investor group seeking control of Warnaco...</Paragraph> <Paragraph position="8"> seeking =~ trying (NOT searching) 4. The ~rm' command takes three argmnents.</Paragraph> <Paragraph position="9"> takes ~ requires as input (NOT moves) 5. Move back to the last position.</Paragraph> <Paragraph position="10"> Move back =# return (NOT move backwards) Each of these examples represents a clear, ordinary use of language. Yet a semantic interpreter must use a great deal of knowledge to distinguish the intended sense of the italicized phrase from other related ahd competing senses. It is simply not practical to treat this process as one of discriminating among a large set of distinct interpretations. The space of intended meanings is too large, and there are too many common characteristics of various senses. To deal effectively with the complexity of this process, a semantic interpreter must accomplish the following: 1. Identify prospective interpretations--The system must use linguistic information to select interpretations that are consistent with the input.</Paragraph> <Paragraph position="11"> 2. Use linguistic and conceptual knowledge to combine interpretations-This may result in ruling out certain candidates, or in forming new and more precise interpretations from the combination of knowledge sources. 3. Assume a specific interpretation--As in the above examples, a practical understanding of the input must be somewhat more than the maximum that can be &quot;safely&quot; inferred. The system must produce some knowledge structures that are likely candidates but are not certain from the linguistic description.</Paragraph> <Paragraph position="12"> 4. Fail gracefully on contradictions--If an assumed interpretation results in a contradiction, the system must preserve those interpretations that do not conflict. If other interpretations are dependent on a conflicting one, these too must be discarded.</Paragraph> <Paragraph position="13"> The requirements above suggest a model of language understanding that progressively refines a semantic intet~preta tion, based on linguistic and contextual information, but that incorporates into each specific interpretation knowledge upon which that interpretation builds. In other words, the ultimate goal of the system is to produce the most specific consistent interpretation, and the means of achieving that goal is to treat each interpretation as an assumption.</Paragraph> <Paragraph position="14"> This assumption-based interpretation process is known as CONCRETION * \[Wilensky, 1983\]. The idea of concretion is to *This term was originally proposed by Joe Faletti. The problem of concretion was initially defined in a series of seminars conducted by Robert determine ~u~ specific a meaning as is possible from an input, while enabliltg recovery if this interpretation proves overly specific. This process is the essential element of a framework that satisfies the criteria mentioned above.</Paragraph> <Paragraph position="15"> Concretion is an important method for dealing with the problem of vagueness and imprecision as framed above. A system that pexforms concretion can successfully produce a complete interpretation without overcommitting to such an interpretation. The discussion that follows describes the concretion process as implemented in TRUMP and considers how this technique improves upon previous approaches.</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 2 Concretion </SectionTitle> <Paragraph position="0"> Concretion :b~ the process of taking abstract concepts and prodating from ~hem concepts that are more precise, or concrete.</Paragraph> <Paragraph position="1"> The motivation for this mechanism is strong in story understanding \[Norvig, 1983, Wilensky, 1983\], because understanding a story ~,eems to involve a continuous refinement of the major concepts into more specific categories. Concretion does not really involve inferencc, since often the specific meaning is quite explicit, in the text. The process of concretion is evident in understmlding simple words and phrases in limited linguistic contexts as well, as illustrated in the examples presented earlier.</Paragraph> <Paragraph position="2"> Concretion is important because it is the mechanism that allows general knowledge about language to apply at very specific levels o{ semantic interpretation. This is essential for natural language interfaces and well as text processing systems, because it allows a core of linguistic and conceptual knowledge to be used for a variety of domains, mid makes the addition of domain-specific linguistic knowledge easier. For example, knowledge about verbs such as give and take and their relation to transfer-events applies in discussing operating systems or corporate ~akeovers as well as in more general applications.</Paragraph> <Paragraph position="3"> It is hard to see how portability can be achieved without thc capability to entertain a range of inter,elated meaning representations. null A typical natural language input can test several aspects of the concretion process. In example 3, the investor group seeking control of Warnaco, A first-pass semantic analysis derives a seeking action--The investor group is the searcher, and control of Warnaco is being sought. Domain-independent conceptual knowledge suggests that looking for a state means trying to realize the state. Domain-specific knowledge produces the assmnpti(m that the phrase describes a corporate takeover attempt. An interpretation of this specificity is necessary to drive inferences and cooperative responses.</Paragraph> <Paragraph position="4"> The concretion process is illustrated in figure 1. Each stage descril)ed above, and each knowledge source, must be distinct. The surface level semmltic analysis is essential because it derives conceptual relations that would apply also to &quot;looking for&quot;, &quot;searching for&quot;, and even &quot;pursuing&quot;; this analysis thus avoids the redundant representation of each construct. This intermediate analysis also makes it possible to use abstract conceptual roles (such as actor and recipient) to determine specific underlying roles (such as the target and suitor of a corporate takeover) (of. \[Jacobs, 1987\]). The second aspect of concretion, applying conceptual knowledge to produce Wilensky at the University of California at Berkeley. In addition to Wilensky, Faletti, and the author, participants in these seminars included Yigal Arens, Margare~ Butler, David Chin, Chuck Fillmore,' Paul Kay, Marc more specific interpretations, is necessary to refine vague terms and identify metaphorical or other non-literal constructs. The third component, using domain-specific knowledge, separates this general conceptual knowledge from assumptions that depend on an implied context~ the domain of corporate takeovers in this example.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.1 Types of Concretion </SectionTitle> <Paragraph position="0"> Concretion is the specialization of an abstract concept to a more precise interpretation in a given context. As tile examples of the previous sections illustrate, concretion can involve a combination of linguistic and conceptual knowledge, and can result in either a direct specialization or a metaphorical extension. In all cases, concretion re(tuires four ingredients: * An instantiated concept to be specialized e A linguistic or conceptual &quot;trigger&quot; e A target concept type o A conceptual relation between source and target For example, in John cut the salami, the concept to be concreted is cutting, the trigger is the combination of cutting with ediblc-obloid or some such~ the target concept is slicing (indicated by the trigger), and the relation is subcategorization, or DOMINATEs.</Paragraph> <Paragraph position="1"> Concretion is often triggered by a linguistic structure, such as a particular combination of lexical items or the use of a certain phrase structure. Figure 2 shows the concretion of the concept cutting to the concept, severing in the phrase cut the tip off. In this case, the knowledge base contains the informarion that severing is DOMINATEd by cutting. The DOMI-NATE relation, labeled D, indicates that severing is a subcategory of cutting \[Wilensky, 1984\]. Diagonal links labeled with italicized words show relationships between concepts and conceptual roles. The specialization of cutting in this instance is triggered by the linguistic relation v_part-cut_oE, representing the use of the verb cut with the particle off.</Paragraph> <Paragraph position="2"> A specific interpretation is often reached through a series of concretions, as in the seeking control example. In the case of cut, the assumed interpretation can be further specialized if the conceptual object is of a particular type, such as body_part, as 2&quot;7\].</Paragraph> <Paragraph position="3"> The doctor cut the tip off.</Paragraph> <Paragraph position="4"> shown in figure 3. In this example, the concept of amputating is reached through a combination of linguistic and conceptual clues, The doctor cut the leg off, An intended meaning is not necessarily a subcategory of an intermediate interpretation, as shown by the seeking control example. Associations between concepts that are analogous or metaphorically related are represented as VIEWs\[Jacobs and Rau, 1985, Jacobs, 1986a\], which also associate related roles. Figure 4 illustrates the application of a VIEW in the concretion process. In this example, the use of the noun kias to describe the conceptual object of a giving serves as a trigger. A general VIEW of action as transfer-event relates giving (which is a tranffer-event to kissing (which is an action), and also the recipient of the giving to the conceptual object of the action. The m label (for MANIFEST) on some roles indicates that the roles are not necessarily specializations of any more abstract relations. When used in this concretion, the VIEW derives a kissing concept with Mary as the kissee. There are two conceptual objects in a kiss--the person being kissed and the surface being kissed--and the concretion mechanism must use constraints on these roles to determine the correct role. The above examples show several ways in which concretion results in the creation of a specific interpretation using a combination of linguistic and conceptual relationships. The examples also demonstrate that different types of concretion can be combined, as in cut the leg off. In any of these cases, a slight variation in the input can negate the resulting interpretation. When this happens, the concretion mechanism retreats to intermediate structures, thus preserving as much of the semantic result as possible. In the doctor cut the leg off accidentally or the doctor cut the leg off the table, the system will preserve the severing interpretation. Each concretion, therefore, is an assumption that explicitly depends on its trigger and on other consequential relationships.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.2 Related Research </SectionTitle> <Paragraph position="0"> Most language analyzers do not really perform concretion.</Paragraph> <Paragraph position="1"> Unification-based systems \[Pereira and Warren, 1980, Pereira and Shieber, 1984, Gawron et al., 1982\] tend to refine semantic representations by adding semantic FEATURES, represented as variables with assignments. Some of the systems that use a KL-ONE knowledge representation \[Bobrow and Webber, 1980, Sondheimer ct el., 1984\] perform a similar function, but use specific interpretation rules to place concepts in more specific categories, rather than to attempt an algorithm for combining lexical and conceptual knowledge. Hirst's\[Hirst, 1987\] &quot;Polaroid Words&quot; are described in a manner similar to concretion: the &quot;words&quot; gradually develop into a complete representation. However, each word sense is still independent; a &quot;polaroid word&quot; cannot, for example, confiate two word senses. Lytinen's MOPTRANS\[Lytinen, 1984\] includes a specialization mechanism that selects the most specific applicable frame, but like the Polaroid Words, it does not take into account the common assumed meaning. None of these approaches allows interpretations to be mapped or refined into more precise interpretations. Other related research addresses the problem of concretion from a different perspective. Norvig's work \[Norvig, 1986\] concentrates on conceptual concretion as inference, independent of the linguistic issues. Zernik and Brown \[Zernik and Brown, 1988\] model language as a reason maintenance process, unlike TRUMP, which incorporates principles of reason maintenance within a more traditional linguistic framework.</Paragraph> </Section> </Section> <Section position="4" start_page="0" end_page="272" type="metho"> <SectionTitle> 3 The Concretion Algorithm </SectionTitle> <Paragraph position="0"> The discussion that follows describes the details of the concretion mechanism of the TRUMP semantic interpreter.</Paragraph> <Section position="1" start_page="0" end_page="272" type="sub_section"> <SectionTitle> 3.1 When to Concrete </SectionTitle> <Paragraph position="0"> Concretion is performed according to the following rules: * When a linguistic description produces a new conceptual structur G concrete it.</Paragraph> <Paragraph position="1"> New pieces of conceptual knowledge are continually derived as the linguistic input is processed. When a new concept is produced, the new knowledge can interact with existing knowledge to produce a more specific interpretation. ~'or exm~lple, in &quot;John cut the salami,&quot; a cutting with John as cutter is later concreted to a slicing, when the concept of salami is produced.</Paragraph> <Paragraph position="2"> o When a grammatical structure is completed, try to concrete the concepts to which it refers.</Paragraph> <Paragraph position="3"> When a verb phrase is completed, for example, there may be a li:~t of concepts to which the verb phrase potentially refi;rs: Since the verb phrase can be enveloped by grammaritaL structures that necessarily refer to the stone con.cepts, the conceptual knowledge produced fl'om the verb phrase must be combined with any concepts produced fl'om these enveloping structures. In &quot;With a knife, John cut the salami&quot;, the meaning of the main clause is concreted using tim prepositional phrase.</Paragraph> <Paragraph position="4"> Smnmary: Perform concretion whenever new conceptual information,, might result in a more specific semantic interpretation. null</Paragraph> </Section> <Section position="2" start_page="272" end_page="272" type="sub_section"> <SectionTitle> 3deg2 Holy ~o Concrete </SectionTitle> <Paragraph position="0"> The concretion p~'ocess is performed by taking two concepts and combining their conceptual content to produce a more specific concepL Generally, this results in filling out specific roles of the derived concept with more general role fillers. Concretion can a\]E:o result in deriving a non-literM interpretation, as in the &quot;give a kiss&quot; and &quot;take argmnents&quot; examples.</Paragraph> <Paragraph position="1"> The coacreLion process often will Nil, for example if the same role is: filled by different concepts. In &quot;Mary was given a letter to BIll&quot;, this hlocks the possiMlity that Bill is the recipie'nt and tlms also resolves the attachment of the prepositional phrase (tho. correct interpretation is A letter to Bill was given to Mary). In other cases, concretion maps conceptual roles into new roles or subsumes roles altogether. In &quot;give a kiss&quot;, the role of &quot;kiss&quot; as conceptual objecl, disappears entirely in the concrete concept kissing. However, any modifiers of &quot;kiss&quot; become rol('s of the new concept, so &quot;John gave Mary a quick kiss&quot; is interpreted as &quot;John kissed Mary quickly&quot;. In this case the relation..dfip between the literal giving concept and the concrete kissin,,! is called a VIEW.</Paragraph> <Paragraph position="2"> The concretion mechanism keeps track of the linguistic sl, rnctures vpon which each concept is dependent. Titus if the parsing pro(;ess later discards a linguistic structure that has violated a constraint, a bookkeeping mechanism can discard any concepts that hinge upon that structure. For example, the phrase &quot;'the command sent the message&quot; has two parses, one in which the command, like 'rm', is sending a message, and the other in which tim command is receiving the message.</Paragraph> <Paragraph position="3"> (One parse is a complete sentence, the second is a complex uoun phrase.) In the garden path sentence, &quot;the command :;ent the m(ssage halted&quot;, the semantic interpretation of the command playing the role of sender must be discarded for syntactic reasons. Tiffs resembles dependency-directed hacktracking \[Doyle, 1979, DeKleer, 1986\], but is accomplished simply by eonstanL)y throwing away concepts that are no longer valid interpret aft, ms.</Paragraph> <Paragraph position="4"> The input to tim concretion algorithm in TRUMP is a concept Cot,t to bc concreted, and a new piece of information (2 ...... also expressed as a concept. Concept C,,~ may come fl'om a newlj interpreted piece of linguistic information, a lexical or conceptuM specialization rule, or a conceptual inference.</Paragraph> <Paragraph position="5"> The process operates as follows: o If C ..... is a snbcategory of Cold, form a concept of the .same type as b, and merge Col~'s roles with C,~eto's roles.</Paragraph> <Paragraph position="6"> This results in a concreted concept of the more specific type, with a's roles converted to roles of the new type.</Paragraph> <Paragraph position="7"> (r) If Cold is a VIEW of C ..... proceed as ifC ..... were a subcategory of Cold, except use ROLE-PLAYs from the VIEW.</Paragraph> <Paragraph position="8"> This is the metaphor application process: The result of concretion is a new concept of the type of b, but the roles must be filled according to tile same VIEW that produced b.</Paragraph> <Paragraph position="9"> * if a role at any ~tagc is filled by two different tokens, fail. Concretion does not allow conflict in the filling of ROLE-PLAYs. if a VIEW results in the application of a ROLE-PLAY that potentially fills more than one role, conflict may be avoided. For example, in &quot;All gave a punch to the jaw to t'h'azier&quot;, the two &quot;to&quot; phrases are allowed, but are automatically excluded from describing the: same conceptual role.</Paragraph> <Paragraph position="10"> If concretion results in the farther specification of a VIEW, re-apply the VIEW.</Paragraph> <Paragraph position="11"> This is another bookkeeping process. If a concept has already been concreted by a VIEW and then is further specified by concretion, the same VIEW is automatically applied. Thus, if a transfer-event is interpreted during concretion to produce an execute-operation and the object role of the transfer-event is later filled, the input role of the execute-operation is also filled.</Paragraph> <Paragraph position="12"> * If concretion results in violating a constraint, undo all concretions dependent on the concreted concept.</Paragraph> <Paragraph position="13"> Since concet)ts that have already been concreted can continue to have their roles filled by concretion, it is possible that a violated constraint may eliminate many specific interpretations. These interpretations are treated as assumptions dependent on other concretions; thus keeping track of tile dependencies allows assumed interpretations to be easily terminated based on new information.</Paragraph> <Paragraph position="14"> Sunrmary: Produce the most specific conceptual interpretation of the input, with the appropriate roles filled, taking care to avoid conflicting interpretations.</Paragraph> <Paragraph position="15"> Many complex data structures and implementation issues are involved in the details of the above process. However, this presentation should suffice to describe how concepts are refined during semantic interpretation by applying conceptual and metaphorical knowledge.</Paragraph> </Section> </Section> class="xml-element"></Paper>