The main differences between syntax and semantics are the model built ({\em situation model} versus utterance model) and the A/R set that indexes that model. For syntactic comprehension, the phrase structure rules of X-bar theory gave a well-defined set of categorical primitives and grammatical roles around which to structure the Syntactic A/R set. A similarly well-defined theory for semantics is more difficult to find. The design space for indexing semantic categories ranges from the minimalist approach of Schank's early Conceptual Dependency theory (Schank, 1972) to using the context-specific meaning of each word in the sentence.
In NLS we take a middle-of-the-road approach by adopting the categorical primitives suggested by Jackendoff's Lexical-Conceptual Structure (LCS) \cite{Jack90}. In his development of LCS theory, Jackendoff strives to maintain parallels to X-bar theory. The three generalized rules of X-bar theory, given in Section \ref{syntax}, are designed to expand the core category to various constituent levels. NLS indexes each of these constituent levels and allows the syntactic relations that are appropriate at each level. For semantics, we index not the constituent levels, but levels of abstracted meaning. As category primitives we adopt the eight basic categories suggested in \cite{Jack90}: ACTION, AMOUNT, EVENT, PATH, PLACE, PROPERTY, STATE, THING. The intermediate level, and semantic analog to the {\em X'} level in syntax, is the primitive sense, or {\em psense} of the word. Psenses are determined by the word's top-level hypernym in WordNet \cite{Miller90}. Thus, the verb {\em suspect} in our example from Section \ref{syntax} would be indexed first as an ACTION and then by its primitive sense, REASON. NLS identifies these classifications from the information found in the semantic lexical entry for the word.
In addition to category and psense, the last avenue of access to the semantics of the word is its context-specific (CS) sense. The particular relations that a word can assign at this level are found in its semantic lexical entry as well. The word {\em suspects}, for example, has an external role for the agent of the action and an internal role for the patient of the action. These external and internal roles are indexed in the semantic A/R set at the CS level under SUSPECT. Constraints on the LCS and CS structures that might fill these roles can be gained from the CS level itself or from its psense or category levels. Thus, there are three levels of meaning potentially available for use during semantic comprehension and higher-level pragmatic interpretation: ACTION, REASON, and SUSPECT.
Now that we have chosen a semantic theory, we have to worry about how to represent the semantic structure of an utterance in Soar. We could represent a using either logics or a model (Lewis, section 2.1.1, 1993). Because knowledge encoded in models can be easily extracted using match-like processing, a model representation has the advantage of being computationally efficient and that is the main reason why we chose a model as the representation of syntax. Therefore, given an utterance, we can now create a model of the syntactic structure of the situation and we call this model the situation model (henceforth, the s-model).
Realization of the u-model in NL-Soar
An example of the s-model for the sentence the man likes the horse with the red apple. is shown below:
The attached structure is the S-model; the unattached item represents alternative senses of the word. For details about how the model is constructed, click here.
The relations between items in the s-model correspond to the roles (E)xternal (for subject), (I)nternal , and Internal2(I2) and (P)roperty indexed off a particular category and psense value. Internal(2) roles are called internal because they are inherited from the psense of the verb. Thus the word GIVE has an internal and an internal2 because its psense TRANSFER entails transfering a THING to a PLACE. Words could also have embedded internal roles, such as the word enter . It would have an internal role of category PATH(TO) which itself has an internal role of PLACE(IN). In this situation we find that the internal role of enter is a path with a psense already defined and which itself has an internal role of category place. It is the internal role of the PLACE which must be filled by a fuse is the role of PLACE. This is shown in the figure below:
In addition to these roles, the category, psense, and other attributes like ^animate or ^countable constitute the full situation model.
See the Semantics A/R set to see how this s-model is organized in the working memory.
Back to the problem-space hierarchy.
This page written by Julie Van Dyke (vandyke@cs.cmu.edu) August, 1997