Brought to you by EarthWeb
IT Library Logo

Click Here!
Click Here!

Search the site:
 
EXPERT SEARCH -----
Programming Languages
Databases
Security
Web Services
Network Services
Middleware
Components
Operating Systems
User Interfaces
Groupware & Collaboration
Content Management
Productivity Applications
Hardware
Fun & Games

EarthWeb Direct EarthWeb Direct Fatbrain Auctions Support Source Answers

EarthWeb sites
Crossnodes
Datamation
Developer.com
DICE
EarthWeb.com
EarthWeb Direct
ERP Hub
Gamelan
GoCertify.com
HTMLGoodies
Intranet Journal
IT Knowledge
IT Library
JavaGoodies
JARS
JavaScripts.com
open source IT
RoadCoders
Y2K Info

Previous Table of Contents Next


2.3. THE "CONCEPTUAL" APPROACH

The real appearance of a "semantic" approach to NLP -- i.e., the awareness that the most complex types of utilization of computers in the natural language field, like, e.g., mechanical translation, could not be satisfactorily executed without some sort of "understanding" of the meaning of the original documents -- must be associated mainly with the elaboration by Roger Schank, in the late 1960s, of the theory of Conceptual Dependency (CD). This last was based on the thought that the "meaning" of a given sentence could be represented as a formal sentence or group of sentences written according to the syntax of a particular conceptual language -- two sentences identical in meaning, and independently from the particular language in which they were originally written or uttered, should have necessarily the same conceptual representation.

Conceptual dependency representations were organized around a small number of semantic primitives, which included primitive acts (11 in the 1977 formulation) and primitive states. An example of a primitive act is ATRANS, representing the transfer of an abstract relationship: "to give" corresponds to ATRANS the relationship of possession. Other examples of primitive acts are PTRANS, the transfer of an object to a new physical location, like "putting" an object somewhere; MTRANS, the transfer of a mental information between people, like in "telling;" MBUILD, used to represent the construction of new information from old, like in "imagining" or "inferring;" or PROPEL, the application of a physical force to an object. Primitive states consisted of predicates with associated numerical values in the range -10/+10: e.g., HEALTH(-10) means that a particular individual is dead; the number of primitive states is much larger than that of primitive acts.

The association of primitive acts, primitive states, and "concepts" through a small set of dependency links in the form of "case grammar" labels gives rise to "conceptualizations" which, in the original Schank's notation, are represented in a particular semantic network style, including several sorts of arrow. For example, the "conceptualization" translating the "deep meaning" of the sentence, "John gives Mary a book" will then be centered around the ATRANS deep predicate; the concept corresponding to "book" will be introduced through the "objective dependency," labeled with "o," and the recipient-donor dependency between Mary and John, on one hand, and the book, on the other, through the recipient-donor dependency "R." An example of conceptualization including a primitive state could correspond to the NL sentence, "Mary tells John that Bill is dead;" in this case, the primitive act used is MTRANS, and the argument introduced through the objective dependency is, in a simplified form, "(Bill BE HEALTH(-10))." The important point about conceptualizations is that they no more represent, as the structures built up in many of the preceding systems, some sort of surface analysis of a given sentence through the representation of the syntactic relationships among its components, but a representation of the semantic links that link together the (language independent) elements of meaning corresponding to this sentence.

Schank's theories have been severely criticized: by the linguists, because the Conceptual Dependency Theory is not a theory of the language; by the psychologists, because conceptualizations have no real correspondence with the mechanisms of human thought; by the logicians, because CD is not sufficiently formalized and it does not include the existential and universal quantifiers; by the computer scientists, because CD is too vague and, for this reason, difficult to translate into computer algorithms. It had, however (and it continues to have), a strong influence on all the work intended to deal with the "deep meaning" corresponding to surface, NL utterances [see, in the late 1970s, Sowa's Conceptual Graphs and, more recently, Zarri's NKRL (Narrative Knowledge Representation Language); see also subsection 3.5.3].

Schank's disciples (and, more in general, members of the so-called "Yale school;" Schank was for many years Professor of Computer Science at Yale) have developed, during the 1970s and at the beginning of the 1980s, several NLP systems based on CD. MARGIE (Riesbeck, Rieger, and Goldberg), e.g., was a very complex system: in a first phase, a "conceptual parser" translated from NL into a CD-like sort of representation using a specific sort of rules ("expectations") that tested first if some particular, surface linguistic clues (normally, words) were present in the original text, and then executed an action concerning how to convert the input just found into CD terms, what to look for next in the NL input, etc. The middle part of the system was an inference engine capable of deducing facts from the information coming from the original texts and translating into CD format. For example, from the CD equivalent of "John hit Mary," the system was able to deduce that "John was angry with Mary" or that "Mary might get hurt." The last part of the system was a generation module, able to convert back into English the result, in CD format, produced by the inference engine. SAM (Schank and Riesbeck; this system made use, as a front end, of ELI, a strengthened conceptual parser developed by Riesbeck from that of MARGIE) and PAM (Wilensky) worked in the context of the hypotheses on the activity of human story-understanding developed by Schank in collaboration with Abelson. For them, this activity consists in a mixture of applying known "scripts" (a script is a standardized sequence of events, represented in CD format, that describe some sort of stereotyped human activity, like going to the restaurant or visiting a doctor) and of inferring goals when an appropriate script does not exist. PHRAN (Wilensky and Arens) developed further the "conceptual parsing" techniques of MARGIE by including complex "patterns" (standard semantic structures) in the test parts of the rules; FRUMP (de Jong), a "text skimmer," used "sketchy scripts" to produce a surface analysis of stories coming off the AP news wire. We can also mention here another "conceptual" system: Wilks' MT system from English to French, making use of no syntactic analysis and based on the instantiation of templates having a basic agent-action-object structure, like "<MAN GET THING>".

All these systems were flawed in at least two respects. The first one concerns a very general problem, that of the huge quantity of "common sense knowledge" necessary for activating correctly the different inference mechanisms that are proper to any conceptual approach; no global solution has been found, until now, for this sort of problem. The second one concerns the absence, for all of the above systems, of a syntactic analysis component, considered as heterogeneous with respect to a semantic/conceptual approach. This lack implied the introduction into these systems of very ad hoc tools -- like the "recency rule," introduced to partially obviate the absence of simple syntactic relations such as "subject" or "direct object." Modern systems that follow, at least partly, the conceptual approach, like KERNEL, by Palmer et al., or SCISOR, by Jacobs and Rau, all make use now of syntactic tools (see Pereira and Grosz, 1994).


Previous Table of Contents Next

footer nav
Use of this site is subject certain Terms & Conditions.
Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details.