Local Pragmatics

1 The Problems

In the messages about breakdowns in machinery that are being processed by the TACITUS system at SRI Interna­ tional, we find the following sentence:

We disengaged the compressor after the lube oil alarm.

This sentence, like virtually every sentence in natural lan­ guage discourse, confronts us with difficult problems of in­ terpretation. First, there are the reference problems; what do "the compressor" and "the lube oil alarm" refer ·to. Then there is the problem of interpreting the implicit re­ lation between the two nouns "lube oil" (considered as a multiword) and "alarm" in the nominal compound "lube oil alarm". There is also a metonymy that needs to be ex­ panded. An alarm is a physical object, but "after" requires events for its arguments. We need to coerce "the lube oil alarm" into "the sounding of the lube oil alarm". There is the syntactic ambiguity problem of whether to attach the prepositional phrase "after the lube oil alarm" to "the compressor" or to "disengaged". All of these problems we have come to call problems in "local pragmatics" . They seem to be specifically linguistic problems, but the tradi­ tional linguistic methods in syntax and semantics have not yielded solutions of any generality.

The difficulty, as is well-known, is that to solve these problems we need to use a great deal of arbitrarily detailed general commonsense and domain-specific technical knowl­ edge. A theory of local pragmatics phenomena must there­ fore be n theory about how knowledge is used. The aim of our research has been to develop a unified theory of local

Local Pragmatics

Jerry R. Hobbs and Paul Martin Artificial Intelligence Center SRI International

pragmatics, encompassing reference resolution, metonymy, the interpretation of nominal compounds and other im­ plicit and vague predicates, and the resolution of syntac­ tic, lexical, and quantifier scope ambiguities, based on the drawing of appropriate inferences from a large knowledge base.

The TACITUS system is intended to embody that the­ ory. Its specific aim is the interpretation of casualty reports ( casreps ), which are messages in free-flowing text about breakdowns in mechanical. devices. More broadly, how­ ever, our aim is to develop general procedures, together with the underlying theory, for using commonsense and technical knowledge in the interpretation of written (and spoken) discourse regardless of domain.

The TACITUS system has four principal components. First , a syntactic front-end, the DIALOGIC system, trans­ lates sentences of a text into a logical form in first-order predicate calculus, described in Section 3. Second, we are building a knowledge base, specifying large portions of po­ tentially relevant knowledge encoded as predicate calculus axioms (Hobbs et al., 1986). Third, the TACITUS sys­ tem makes use of the KADS theorem prover, developed by Mark Stickel (Stickel, 1982). Finally, there is a prag­ matics module, including a local pragmatics component, which uses the theorem prover to draw appropriate infer­ ences from the knowledge base, thereby constructing an interpretation of the text .

To interpret a sentence, the local pragmatics compo­ nent determines from the logical form of the sentence what interpretation problems need to be solved. Logical ex­ pressions are constructed for each local pragmatics phe­ nomenon the sentence exhibits, and the proofs of these expressions constitute the interpretation of the sentence. Where there is more than one interpretation, it is because there is more than one proof for the expressions. Interpre­ tation is thus viewed as crucially involving deduction.

In Section 2, we describe the three of the phenomena we are addressing-reference, metonymy, and nominal com­ pounds. For each, we describe the expression that needs to be proved, and for the last two, we describe how cur­ rent standard techniques can be seen as special cases of our general approach.

2 Local Pragmatics Phenomena

2.1 Reference

Entities are referred to in discourse in many guises-proper nouns, definite, indefinite, and bare noun phrases of vary­ ing specificity, pronouns, and omitted or implicit argu­ ments. Moreover, verbs, adverbs, and adjectives can re­ fer to events, conditions, or situations. The problem in all of these cases is to determine what is being referred to. Here we confine ourselves to definite noun phrases, although TACITUS handles the other cases in a manner described briefly in Section 4.

In the sentence

The alarm sounded.

the noun phrase "the alarm" is definite, and the hearer is therefore expected to be able to identify a unique en­ tity that the speaker intends to refer to. Restating this in theorem-proving terminology, the natural language system should be able to prove constructively the expression

(:3 x )alarm(x)

It must find an x which is an alarm in the domain model. If it succeeds, it has solved the reference problem. 1

2.2 Metonymy

Metonymy is very common in discourse; few sentences lack examples. Certain functions frequently provide the re­ quired coercions. Wholes are used for parts, tokens for types, and objects for their names. Nunberg (1978), how­ ever, has shown that there is no finite set of possible co­ ercion functions. The relation between the explicit and implicit referents can be virtually anything.

Every morpheme in a sentence conveys information that corresponds to a predication, and every predicate imposes selectional constraints on its arguments. Since entities in the text are generally the arguments of more than one predicate, there could well be inconsistent constraints im­ posed on them. To eliminate this inconsistency, we inter­ pose, as a matter of course, another entity and another relation between any two predications. Thus, when we encounter in the logical form of a sentence

we assume that what is intended is really

1 In this paper we ignore the problem of the uniqueness of the entity referred to. A hint of our approach is this: If the search for a proof is heuristically ordered by salience, then the entity found will be the uniquely most salient.

for some entity z and some relation q. The predication q(z, a) functions as a kind of buffer, or impedence match, between the explicit predications with their possibly incon­ sistent constraints. In many cases, of course, there is no inconsistency. The argument satisfies the selectional con­ straints imposed by the predicate. In this case, z is a and q is identity. This is the first possibility tried in the im­ plemented system. Where this fails, however, the problem is to find what z and q refer to, subject to the constraint, imposed by the predicate after, that z is an event.

TACITUS thus modifies the logical form to

and for an interpretation, the expression that must be proved constructively is

Find an event z bearing some relation q to the alarm.

The most common current method for dealing with metonymy, e.g. , in the TEAM system (Grosz eta!., 1985), is to specify a small set of possible coercion functions, such as name-of. This method can be captured in the presep.t framework by treating q not as a predicate variable, but as a predicate constant, and expressing the possible coercions in axioms like the following: ·

That is, if x is the name of y, then y can be coerced to x. This is the method we have implemented in our initial version of the TACITUS system.

2.3 Nominal Compounds

To interpret a nominal compound, like "lube oil alarm", it is necessay to discover the implicit relation between the nouns. 2 Some relations occur quite frequently in nominal compounds-part-of, location, pttrpose. Moreover, when the head noun is relational, the modifier noun is often one of the arguments of the relation. Levi (1978) argued that these two cases encompassed virtually all nominal com­ pounds. However, Downing (1977) and others have shown that virtually any relation can occur. A lube oil alarm, for example, is an alarm that sounds when the pressure of the lube oil drops too low.

To discover the implicit relation, one must prove con­ structively from the knowledge base the existence of some possible relation, which we may call nn, between the enti­ ties referred to by the nouns:

In our framework, we can implement the approach that hypothesiszes a small set of possible relations, by taking nn to be not a predicate variable but. a predicate constant, and encoding the possibilities in axioms like

2 Some nominal compounds can of course be treated as single lexical items. This case is not interesting and is not considered here.

For example, if a blade x is a part of a fan y, then "fan blade" is a possible nominal compound.

To deal with relational nouns, such as "oil sample", we encode axioms like

This tells us that if x is a sample of oil y, then x can be referred to by the nominal compound "oil sample".

Finin (1980) argues that one of the most common kinds of relations is one that involves the function of the referent of the head noun. The function of a pump is to pump a fluid, so "oil pump" is a possible nominal compound. This can be encoded in axioms of the pattern

That is, if e is the function of x where e is the situation of x doing something p to y, then there is a possible nn relation between y and x.

As with metonymy, in our initial version of TACITUS, it was the standard, restricted method that we implemented. This is because we wanted to make sure we were not los­ ing ground in seeking a general solution. Nevertheless, our approach allows us to begin experimenting with the gen­ eral solution to the nominal compound problem, where the implicit relation can be anything at all.

3 Order of Interpretation

DIALOGIC produces a logical form for the sentence in something like a first-order logic but encodes grammatical subordination relations as well as predicate-argument rela­ tions. It is "ontologically promiscuous" in that events and conditions are reified. A slightly simplified version of the logical form for sentence (1) is (2):

1. The lube oil alarm sounded.

"I" can be read "such that" or "where", so that a para­ phrase of this formula would be "In the past there was an event e1 which was a sounding event by a1 where a1 is an alarm and there is an nn relation between a1 and o1 such that o1 is lube oil."

In general, the logical form of a sentence is a "proposi­ tion". A proposition is a predicate applied to one or more arguments. An argument is either a variable or a "com­ plex term". A complex term is a variable, followed by a "such that" sign, followed by a "restriction". (Complex terms are surrounded by square brackets for readability.) A restriction is a conjunction of propositions.

This notation can be translated into a notation using four-part quantifier structures by successively applying the following transformation:

Since interpretation involves solving a number of prob­ lems, or proving a number of expressions, . a question is raised. In which order should we try to solve them? A naive answer would be to try to solve them "from the in­ side out". Before trying to find the lube oil ·alarm, we should try to find the lube oil the alarm is an alarm for. Before checking that the lube oil alarm obeys the selec­ tional constraints imposed by "sound", we should learn as much as we can about the lube oil alarm; in particular, what "the lube oil alarm" refers to.

This means that given the logical form (2), we should solve the local pragmatics problems in the following order:

1. Find the reference of o1, the lube oil. Prove

1. Given that, find the implicit relation nn en­ coded in the nominal compound. If o1 was resolved to 0, then prove

1. Given that, find the reference of a1 , the alarm. If nn is the relation for, then prove

1. Given that, check the predicate-argument con­ gruence of sound applied to a1 . If a1 was re­ solved to A and sound requires its argument to be a physical object, then prove

Unfortunately, this order will not always work. Infor­ mation relevant to the solution of any of these local prag­ matics problems can come from the solutions of any of the others. This can be seen particularly in problems 2 and 3. Sometimes we will need to have resolved the reference to have enough information to solve the nominal compound problem, and sometimes conversely.

Thus, in a more sophisticated approach, we would con­ struct a single expression to be proved, encoding what is required for all of the local pragmatics problems. For sen­ tence (1), the expressio_n would be

The conjuncts of this expression could be proved in any order. The inside-out order is only one possibility; it can be taken as an indication of how much work the theorem­ prover should devote to proofs of the various conjuncts, and how early. The difficulty with this approach, however, is that it is difficult to get partial results in cases of failure.

We are currently using a comprimise between these two orders-a fail-soft, inside-out order. As we proceed inside out, at each step the theorem-prover is given the full ex­ pression built up to that point. However, the expression has as an antecedent the instantiations of what was proven in earlier steps. Thus, in step 4 in the example, the ex­ pression is

3 Quantifiers other than existentials are ignored in this paper. For the treatment we intend to give them, see Hobbs (1983).

Those prior instantiations consistent with higher con­ straints will be proven immediately from the antecedent, and new proofs will need to be discovered only for those which are inconsistent. 4

4 Other Issues

Syntactic Ambiguity: Within this framework we have also implemented a treatment of the most common syntactic ambiguities. Attachment ambiguity problems are translated into constrained reference problems by in­ troducing an existentially quantified variable representing, say, the subject of the prepositional phrase, constrained by a disjunction listing the possible attachment sites.

Given and New: A problem that arises in the casreps in a severe form, since determiners are often omitted, is the problem of determining whether an entity referred to is definite or indefinite, that is, given or new. In

Metal particles in oil sample and strainer.

"oil sample" is indefinite (or new) while "oil strainer" is definite (or given). This problem occurs in general dis­ course as well, especially for nonnominal reference. Our approach is to assert the existence of each entity, but with a cost determined by the type of reference-in increasing order, indefinite, nonnominal, bare, or definite. If this as­ sertion (called a referential implicature) is required in the proof that constitutes the interpretation, the entity is new; if not, it is given.

Minimality: Axioms can be assigned a cost, depending upon their salience. High salience, low cost axioms would then be tried first. Short proofs are naturally tried before long proofs. Thus, a cost depending on salience and length is associated with each proof, and hence with each inter­ pretation. Where, as usually happens, there is more than one possible interpretation, the better interpretations are supported by less expensive proofs. A second criterion for good interpretations is to favor the minimal solution in the sense that the fewest new entities and relations needed to be hypothesized. For example, the argument-relation pat­ tern in nominal compounds, as in "oil sample", is minimal in that no new implicit relation need be hypothesized; the one already given by the head noun will do. In metonymy, the identity coercion is favored for the same reason, and shorter coercions are favored over longer ones. These ideas at least give us a start on the very difficult problem of choosing the best interpretation.

4 This technique is due to Mark Stickel.

Acknowledgements

The authors have profited from discussions with Mark Stickel, Doug Edwards, Bill Croft, Fernando Pereira, Ray Perrault , Stu Shieber, and Mabry Tyson about this work. The research was funded by the Defense Advanced Re­ search Projects Agency under Office of Naval Research contract N00014-85-C-0013, as part of the Strategic Com­ puting Program.