Semantics A/R set

General

The semantics A/R set is the mechanism by which items are indexed in working memory according to their potential roles in lexical-conceptual structure (Jackendoff, 1990). The concept of the A/R set is taken from (Lewis, 1993) who used a structure analgous to the one described here to index syntactic relations in working memory. A more thorough discussion of the A/R set and its theoretical import for NL-soar's theory of sentence comprehension, see syntactic a/r set.

The use of the A/R set for semantics extends the mechanisms already demonstrated to be important in sentence comprehension. It does not however, constitute NL-soar's entire theory of semantics. Rather, it is used at the level of lexical semantics, providing instantiated LCS structures for further pragmatic and discourse processing. For a more detailed discussion of semantics in NL-soar, see S-model .

Implementation

The semantics A/R set follows the model of syntax in allowing only 2 non- descriminable items to be held in working memory. This is implemented by top state attributes
^semassigners S2 ^semreceivers S3 
Each of these attributes is indexed by a semantic category from the set of conceptual categories enumerated in (Jackendoff, 1990). Thus, we have the following:
(S2 ^action A20 ^amount A21 ^event E2 ^path P19 ^place P18 ^property P20 ^state S6 ^thing T1)
(S3 ^action A22 ^amount A23 ^event E3 ^path P22 ^place P21 ^property P23 ^state S7 ^thing T2)
These can be thought of as the semantics analogue of the syntactic categories ^a(djective), ^c(omplementizer),^i(nflection),^n(oun), ^p(reposition),^v(erb), ^det(erminer). Thus, for syntax, the first level of assigners is
(A1 ^a A3 ^c C1 ^i I3 ^n N1 ^p P4 ^v V1 ^det D1) 
For syntax, the value of each category unfolds as follows:
	(V1 ^max V2 ^one V3 ^zero V4)
	  (V2 ^head X6 ^spec X6 ^adjoin X6)
    	    (V3 ^head X5 ^comp X5 ^adjoin X5)
    	    (V4 ^adjoin E26)
This reflects a static populating of the A/R set according to X'-bar theory (ala Radford, 1988). Thus, each category has projection levels max , one , and zero . Each of these levels assigns particular relations: max can assign spec , head , and adjoin , head can assign head , comp , and adjoin , etc.

The implementation of the semantics A/R set is different in this regard. Syntax incorporates a notion of projecting from a zero to max level through the addition of various constituents modifying the head, with the result being a maximally-sized constituent of the same category as the head. Semantics does not have this notion, but rather assocates maximal projections in certain conceptual relations defined by the semantics of the head. Thus where bar-level is a relevant distinguishing factor for syntax, for semantics we take it to be the primitive sense of the verb.

Thus, for the sentence "the man entered" the semassigners set has the following form:

(S2 ^action A20 ^amount A21 ^event E2 ^path P19 ^place P18 ^property P20
       ^state S6 ^thing T1)
  (T1 ^entity P73)
  (E2 ^go P149)
  (S6)
  (A20)
  (P18)
  (P19)
  (P20)
  (A21)
This reflects the fact that the verb "enter" is an EVENT of type of "go" and the word "man" is a THING of type "entity".

There are two important things to note here. First, that unlike syntax, the semassigners list is not fully populated, but rather is dependent on the particular sentence currently being comprehended. This greatly reduces the computational cost and theoretical morass that would be associated with devising a static structure for semantics, while preserving the intuition that the particular set of semantics primitives is difficult to identify and of potentially infinite cardinality. Secondly, the value of these primitives is dependent on the lexical-conceptual structures for the words being comprehended. These are discussed in more detail with the discussion of the lexicon .

Having discussed the first two levels of discrimination, we now turn to the third. The reader will remember that for syntax, this was the particular syntactic roles available at each projection level. Thus, we had

	(V1 ^max V2 ^one V3 ^zero V4)
	  (V2 ^head X6 ^spec X6 ^adjoin X6)
    	    (V3 ^head X5 ^comp X5 ^adjoin X5)
    	    (V4 ^adjoin E26)
As mentioned previously, in semantics there is no notion of projections, since the zero-level items are already assumed to be maximal projections. We do however, have the discrimination of external vs. internal arguments. This discrimination is well documented in linguistics and appears as subject/object assymetries in the theoretical literature. External arguments are traditionally associated with subjects. In this implementation we maintain this association, and following the work of Relational Grammar, extend them to modifiers. The semassigners list for "the man entered" will therefore expand to:
(S2 ^action A20 ^amount A21 ^event E2 ^path P19 ^place P18 ^property P20
       ^state S6 ^thing T1)
  (T1 ^entity P73)
    (P73)
  (E2 ^go P149)
    (P149 ^external S62 ^internal S62)
  (S6)
  (A20)
  (P18)
  (P19)
  (P20)
  (A21)
The value of both the external and internal relations is the LCS for the word being indexed, here entered . In other words, the word enter assigns the relation "external argument of a go event" and the relation "internal argument of a go event." To see how these relations might get filled, we must now turn to a discussion of the semreceivers list, which has the following form:
(S3 ^action A22 ^amount A23 ^event E3 ^path P22 ^place P21 ^property P23
       ^state S7 ^thing T2)
  (T2 ^entity P75)
    (P75 ^external S62 ^internal S62 ^internal2 S62
           ^missing-semassigner-relations internal2
           ^missing-semassigner-relations external
           ^missing-semassigner-relations internal)
  (E3 ^go P151)
    (P151 ^context S62 ^external S62 ^internal S62 ^internal2 S62
           ^missing-semassigner-relations context
           ^missing-semassigner-relations internal
           ^missing-semassigner-relations external
           ^missing-semassigner-relations internal2)
  (S7)
  (A22)
  (P21)
  (P22)
  (P23)
  (A23)
Ignoring the ^missing-assigner-relations flags for the moment, we begin by notice that the semreceivers list is indexed similarly to the semassigners list. The categories are the same set as those used for assigners as are the primitives and roles. This is necessary for matching potential assigners to receivers: a receiver must be indexed for the same role, primitive sense, and category as the assigner. The major difference for receivers, however, is that they can potentially receive their role from any category and primitive sense: an ENTITY of type THING could be the external role of a GO EVENT or the internal role of an IN function of category PLACE, etc. To effect this, we allow each word to receive any role from any category.

Having said this one should now notice that only S62 is a value for any of these relations, and this is the LCS structure for the word "entered". To find the indexing for "man" we must look at semreceivers2:

(S5 ^action A26 ^amount A27 ^event E5 ^path P28 ^place P27 ^property P29
       ^state S9 ^thing T4)
  (T4 ^entity P76)
    (P76 ^external S30 ^internal S30 ^internal2 S30)
  (E5 ^go P152)
    (P152 ^external S30 ^internal S30 ^internal2 S30)
  (S9)
  (A26)
  (P27)
  (P28)
  (P29)
  (A27)
Here, S30 is the LCS for "man". It appears on semreceivers2 because of our previously stated working memory limit of 2 non-discriminable items. When the second word "entered" came in, "man" got moved to semreceivers2 because it had not yet received a role. Now that we have "entered" we can match the two by noticing that "man" can receive the "external role of a GO EVENT" and "event" can assign the "external role of a GO EVENT". This allows removing "man" from the receivers list and "fusing" it into the LCS structure of "entered". For a more thorough discussion of the fuse operator, see fuse operator . The final tree resulting from this fusion is seen below:

Maintainer: vandyke@cs.cmu.edu (Last updated 21 August 1997 by vandyke)