---
title: "Language Processing Via Canonical Verbs and Semantic Models"
authors:
  - "Gary G. Hendrix"
  - "Craig W. Thompson"
  - "Jonathan Slocum"
doi: ""
url: "https://ijcai.org/proceedings/1973/34"
pdf: "https://ijcai.org/proceedings/1973/0034.pdf"
venue: "Proceedings of the Third International Joint Conference on Artificial Intelligence"
year: 1973
pages: "262-268"
track: "Natural Language: Systems"
keywords:
---

# Language Processing Via Canonical Verbs and Semantic Models

Gary G. Hendrix, Craig W. Thompson, Jonathan Slocum

*Proceedings of the Third International Joint Conference on Artificial Intelligence*, pages 262-268.

## Abstract

A natural language question answe ring system is presented. The system's parser maps semantic paraphrases into a single deep structure characterized by a canonical verb. A modeling scheme using semantic nets and STRIPS-like operators assimilates the sequence of input information. Natural language responses to questions are generated from a data base of semantic nets by "parsing" syntactic rules retrieved from the lexicon.

## Full text

## Introduction

In natural language processing, the problems of representing factual material, making inferences, solving problems and answering questions may be sig­ nificantly reduced if identical meanings expressed by diverse surface structures can be represented by a single conceptual construct. A major cause of surface structure diversity is the existence of a wide variety of "surface verbs" for describing basically the same situation. For example, "buy ," "sell," and "cost," etc., all describe essentially the same event. Schank (1) and others have posited the existence of "deep" or "canonical" verbs which unify in the deep structure meaning common to possibly many surface verbs used in ordinary speech. Given a canonical event such as EXCHANGE, the participants (e.g. the BUYER, the SELLER) are ordered in the surface structure, but dependent upon the choice of surface verb and voice (active or passive). This paper describes a language processing system which employs semantic network depictions of canonical events as its fundamental representation of meaning.

The language processor consists of three basic modules as shown in Figure 1. The Parser is used to map English surface structures into a semantic repre­ sentation utilizing canonical verbs. The sequence of events presented to the language processor over sev­ eral sentences is woven into a unified knowledge structure by the Modeling System, which builds a STRIPS robot-like model of the situations described. Again, as in the Parser, the canonical event provides the basis of representation. When a question is pre­ sented to the system, the Generator is used to produce English output directly from the canonical semantic net built (sometime in the past) by the Modeling System.

FIGURE 1

![Image](https://ijcai.org/proceedings/1973/fulltext/34/image_000000_61875c212154ad3fcaffa3cafea640632a17cb1c5a2bbf307eb18238bc43a053.png)

A Schematic of the Language Processing System

The mapping from English into a canonical con­ stituent structure is accomplished by a variant of the Augmented Finite State Transition Network system des­ cribed by Woods (2). Since the operation of the sys­ tem is generally well known, attention will be con­ centrated on the mechanism which allows for direct translation into canonical structures. The key to this operation is the inclusion of certain information in the Lexicon (Table 1). The property list of each surface verb form (infinitive, past, etc.) contains a reference to a central GENSYM-ed atom which serves as the lexical entry for all forms of a given verb. For readability, these "central" forms will be desig­ nated by L-verb, where "verb" is the usual infiniti~e form. The property list of each central verb conta~ns two special entries (in addition to various others): the attribute CANON-VB indicates the canonical verb (event) associated witq the given surface verb; the P-RULES attribute relates the surface noun phrases and prepositional phrases used in conjunction with the given surface verb, to the deep case structure asso­ ciated with the corresponding canonical verb. This attribute is a list of triples; the first entity in each triple indicates the manner in which an event participant is specified in the surface sentence. I f the participant is specified simply as a noun phrase , the first entry of the triple is "OK." If the parti­ cipant is specified by a prepositional phrase, then the corresponding preposition appears in place of OK. The third entry in the triple names the deep case relation associated with the event participant. For example, consider :

## JOHN BOUGHT THE CAR FROM MARY

In Table 1 it may be seen that FROM is associated with the deep case SELLER with respect to the verb BUY ·. The second entry in each triple is a list of semant~c classes: a participant must belong to one of the semantic classes specified by the second triple entry in order to satisfy the deep case relation specified by the third triple entry. This semantic class infor­ mation is used to help disambiguate the deep cases of sentence components which otherwise appear similar. (This process is also aided by the ordering of the triples.)

```
L-BUY WORDCLASS CANON-VB INF SG3 PAST -EN -ING PRULES GRULES L-COST WORDCLASS CANON-VB INF GRULES L-PAY WORDCLASS CANON-VB INF GRULES L-MAN WORDCLASS s PL SPOSS PLPOSS MKR L-WHEN WORDCLASS PI MKR EXCHANGE SURF-VB VERB EXCHANGE BUY BUYS BOUGHT BOUGHT BUYING ( (OK (HUMAN ORGANIZATION) BUYER) (OK (PHYSOBJ) THINGBT) (FROM (HUMAN ORGANIZATION) SELLER) (FOR (MONEY) THINGGIVEN) (AT (PLACE) LOC) (IN (PLACE) LOC) (OK (DAYPART) TIME) (IN (DAYPART) TIME) (BUYER ACTIVE THINGBT (FROM SELLER) (FOR THINGGIVEN) ) (THINGBT PASSIVE (FROM SELLER) VERB EXCHANGE COST (FOR THINGGIVEN) ) ) CTHINGBT ACTIVE (BUYER) THINGGIVEN) ) VERB EXCHANGE PAY (BUYER ACTIVE (SELLER) (THINGGIVEN) (FOR THINGBT) ) ) NOUN MAN MEN MAN'S MEN'S (HUMAN) QWORD WHEN (DAY DAYPART) (L-BUY L-SELL L-PAY L-COST)
```

## TABLE 1

Some lexical structures

Since semantic class information will be useful in determining deep cases, each noun entry in the lexicon contains a special attribute (MKR) on its property list; this specifies the semantic classes to Which the noun belongs. Interrogative pronouns and adjectives (QWORDS) also have the attribute MKR.

The application of the transition network par­ sing grammar to an input sentence produces two (three, in the case of questions) intermediate structures. For example :

WHO BOUGHT A CAKE AT THE NEW BAKERY FROM THE BAKER. will cause the . production of the intermediate struc­ tures:

```
Verbal constituent: (CANON-VB EXCHANGE MODAL (TENSE PAST MOOD INTERROG CASE AFFIRM) ) NP-PP constituents : (OK (PHYSOBJ) (TOK L-CAKE DET INDEF NBR S)) (AT (PLACE) (TOK L-BAKERY DET DEF NBR S MOD (AGE L-NEW))) (FROM (HUMAN) (TOK L-BAKER DET DEF NBR S)) QWORD constituent: (OK (HUMAN) (TOK L-WHO))
```

The Verbal constituent indicates the canonical verb underlying the sentence as well as the sentence modal­ ity. The NP-PP constituents list specifies the NP and PP components of the input sentence in the order in which they appeared. Each such component is repre­ sented by a triple: the first entry in the triple indicates the syntactic form (PP, or OK in the case of NP) in which the component appeared; the second entry indicates the semantic classes associated with the component (copied from the lexicon); the third entry is an even-order list describing the nature of the component (as given in the input sentence). The appearance of a QWORD in an interrogative sentence will cause a QWORD constituent to be produced: this is a triple having the same form as a (single) NP-PP constituent.

The final stage in mapping English sentences into canonical structures consists of correlating the entries in the NP-PP constituents list, and the QWORD constituent if present, with the PRULES associated with the surface verb, Basically, for each NP-PP con­ stituent a search is made through the first entries of successive PRULES for a preposition or OK which matches the first entry of the NP-PP constituent; when a candi­ date is found, the rule's second entry is matched against the second entry of the constituent: if the two lists (sets) of semantic markers have a non-null intersection and the deep case indicated by the third entry of the PRULE triple is not already assigned, then that deep case is associated with the constituent as shown below. Once the NP-PP constituents have been assigned their appropriate cases, the QWORD constituent (if any) is considered; the matching process is essen­ tially that described above, except that all the unas­ signed cases which match are collected in a list to be paired with the attribute ARGS. The resultant struc­ ture is assigned the pseudo deep case label Q. Thus the final structure ("canonical structure") produced by the parser for the indicated input sentence is:

(CANON-VB EXCHANGE MODAL (TENSE PAST MOOD INTERROG

```
CASE AFFIRM)
```

SELLER (TOK L-BAKER DET DEF NBR S) THINGBT (TOK L-CAKE DET INDEF NBR S) LOC (TOK L-BAKERY DET DEF NBR S MOD (AGE L-NEW)) Q (ARGS (BUYER) TOK L-WHO) )

The purpose of the parser is to provide a simple, expressive "front end" to the modeling system; the ver­ sion described is (so far) limited to (possibly con­ joined) simple, active sentences without embeddings. Work is in progress to extend the parser to include passives and several types of embedded sentences. Nat­ ural extensions include: (1) more complex PRULES, which explicitly express the order of NP's and PP's in the constituents list; (2) a more intelligent correla­ ting algorithm, which is able to recognize NP comple­ ments and relative clauses; (3) PRULES for nominalized verbs and nouns modified by PP's; and (4) a richer variety of syntactic and semantic markers.

A State-of-the-World Graph

![Image](https://ijcai.org/proceedings/1973/fulltext/34/image_000001_84f53d052f3ca3fcce600d73e586c64df12daead0ac169b1433387f9cd1ce041.png)

## The Modeling System

The purpose of the modeling system is to inte­ grate the collection of facts or concepts presented to the system through the parser. Three distinct types of concepts must be encoded: (1) objects; (2) rela­ tionships between objects; and (3) changes in rela­ tionships with respect to time. The modeling system uses an encoding procedure analogous to that used in motion pictures, recording changes in relationships with respect to time by a sequence of "still photo­ graphs." The "still photos" of the modeling system are graphs (nets) representing the state of the world at an instant in time. Nodes in the graph stand for objects, and arcs represent static (binary) relation­ ships between objects. (No time relationships are included.)

Figure 2 is an example of a state-of-the-world graph (SWG). Node names beginning and ending with asterisks (depicting GENSYM atoms) represent particu­ lar objects in the real world. Other node names are words in the system's lexicon. Consider the node labeled *JOHN>'<. >'<JOHN"' is an element of L-JOHN and L-GROCER. L-JOHN is the set of all Johns, a subset of L-MAN, the set of all men. L-GROCER is the set of all grocers. Hence, *JOHN>'< is a node representing a par­ ticular man who is a John who is a grocer. This John owns a particular house, "'JHOUSE"'. Further, John is at that house, owns an old, red car and is fat.

of still photographs of motion pictures, is defined. The sequence of events causing changes in the state of the world is recorded on a time line (see Figure 3). Each node on the time line corresponds to an event. Beginning with some initial state, the state of the world following any event may be determined by consid­ ering the sequence of events along the time line. Associated with each event are a list of relationships which held before the event took place but which are no longer valid after the event (the delete list) and a list of relationships which did not hold (or at least were not known to hold) before the event took place but which are valid after the event (the add list). Thus, each node on the time line models a set of changes in the state of the world produced by the occurrence of some event. When known, a canonical verb and a set of parameters which describe the event causing the changes are also associated with a time line node. Not all event parameters need be known ("John bought something at the bakery."), but the more parameters which are known, the better the event is specified.

To model changes in relationships with respect to time, a sequence of SWGs, analogous to the sequence The language processing system begins operation with a SWG depicting only lexical information. As des­ criptions of events are given as inputs, nodes are added to the time line. With each new event, certain relations (those on the event's delete list) are de­ leted from the current SWG and new relationships (those on the add list) are added, transforming the old pic­ ture of the state of the world into a new picture show­ ing conditions after the event. To answer questions about relationships in the past, the "movie" may be run backwards by considering the sequence of events in

![Image](https://ijcai.org/proceedings/1973/fulltext/34/image_000002_fff2ba5e17b41668d34633c059f3480d28baff98e25b58d7639e45c6e0a248b9.png)

## canonical verb

is thfhgraph representing

the j state of the world

- is thfhnode representing the j event

## FIGURE 3

Abstract Diagram of the Time Line

Time Line Record of Two EXCHANGE Events (Note, "intermediate" nodes deleted for simplicity)

![Image](https://ijcai.org/proceedings/1973/fulltext/34/image_000003_5f8df35aeebedc281838e5635857d5c14db2b05f5da3e4fe82ea103475a5e984.png)

reverse order and deleting the adds and adding the deletes. Once a question is answered, an analogous forward process resets the SWG to reflect current conditions.

Questions regarding the nature of events them­ selves (as opposed to questions concerning relation­ ships which are altered by events) are answered by consulting the record of events on the time line .

As an example of a time line node, consider how the event "John traded his car to Tom for Tom's boat" would change the SWG of Figure 2. The relationships (OWNS 1<JOHN''' "'CAR1') and (OWNS 1<TOW< '"BOAT"<) would no l onger be true. (OWNS 1<JOHN>'< >'<BOAT"<) and (OWNS >'<TOW< "'CAR''') would become true. Hence, a time line node like the left node of Figure 4 would be produced.

## Events as operators

Since an event transforms one state of the world into another, it is reasonable to think of events as operators. Pursuing this notion, a canonical verb becomes the name of an event procedure (or operation) which transforms the state of the world (an implicit parameter) and a set of explicit event parameters into a new state of the world. The event procedures to be presented here closely parallel the operators used by the STRIPS robot (3) .

An integral part of any STRIPS operator is the

"precondition list" which the robot uses in determin­ ing a legal sequence of operations. Since our system receives a chronological accounting of events, the necessity of precondition lists for determining event sequence is eliminated. There is, however, a related problem: suppose that part of the input to the lan­ guage processor is "John bought a clock at the hard­ ware store. Then J ohn bought a cake at the new bak­ ery." Note that J ohn was at the hardware store when he bought the clock. The next event has him buying a cake at the new bakery. Before the event at the bakery could take place (as a precondition to the bakery event), J ohn must have gone (perhaps by a com­ plex route) from the hardware store to the bakery . Thus, the necessity of satisfying preconditions clearly remains a problem even when the event sequence is given. The precondition problem has become the problem of determining the nature of unspecified intermediate events which set the stage for the accomplishment of specified events.

The nature of these unspecified intermediate events may, to a large extent, be determined by the preconditions of the specified events. Returning to the example, before "John bought a cake at the new bakery" some event must have occurred which deleted the relation that he was at the hardware store and added the relation that he was at the bakery . (In general, EXCHANGE events must be preceded by events which bring the participants in the event to the place