Meeting the Interlocking Needs of LF-Computation Deindexing and Inference: An Organic Approach to General NLU

1 The Organism Awaits: A Manifesto

We report here on an approach to theoretical and practi­ cal NLU that aspires to be complete and comprehensive, with respect to all the major syntactic, semantic, and pragmatic phenomena encountered in NL texts. Before launching into technical particulars, we will devote some space to motivating our endeavor, since comprehensive approaches appear to be somewhat unfashionable. For instance, the overwhelming majority of papers at recent ACL conferences have focused on specific issues in gram­ mar formalisms, parsing, discourse, corpus-based analy­ sis, etc., with scant attention to the role of the proposed theories or techniques within a larger NLU framework.

This trend toward specialization represents a striking shift from the heady days of the 70's and early 80's, when (at least to a casual observer) each year appeared to bring a more complete picture of the essential struc­ tures and processes underlying language understanding and a more impressive set of working systems demon­ strating the power of the new ideas and mechanisms. In retrospect, these efforts were fated to encounter two formidable barriers: a "knowledge bottleneck" and a "theory bottleneck." The knowledge bottleneck lies in the dependence of genuine understanding on a vast sub­ strate of general and specific knowledge, and the diffi­ culty of imparting all but trivial amounts of such knowl­ edge to machines "by hand." The theory bottleneck lies in the myriad problems of syntax, semantics, pragmat­ ics and inference that any general NLU or dialog system needs to address. AI-oriented researchers perhaps tended to underestimate the latter bottleneck, armed as they were with new insights and techniques from AI. But the brittleness and inextensibility of the NLU systems of the 70's and 80's was ultimately seen to lie as much in their reliance on programmer intuitions and domain-specific hacks to bridge gaps in theoretical foundations, as in the meagerness of their knowledge.

·The authors are grateful to James Allen and Philip Harrison for helpful comments during EL development, and Stephanie Schaeffer for staying with program development for EPILOG. This research was supported in part by the Boe­ ing Co. in Seattle (under Purchase Contracts W278258 and W288104), NSERC Operating Grant A8818, ONR/DARPA Research Contracts No. N00014-82-K-0193 and No. N0001492-J-1512, and NSF Research Grant IRI-9013160.

From this perspective, it is not surprising that the field has shifted into "divide and conquer" mode , char­ acterized by fragmentation and specialization. However, we believe that at this point , the dearth of comprehen­ sive theorizing and system-building also represents some­ thing of a loss of nerve in the AI/CL community. Divide­ and-conquer is tenable up to a point , but ultimately we need to confront language as an organic phenomenon, whose facets have no clearly discernible shape indepen­ dently of each other. Even grammar, the most accessible facet , is moot (witness the various competing syntac­ tic formalisms, and competing analyses of many specific phenomena); and logical form, context, discourse struc­ ture, knowledge representation, and inference are utterly hypothetical. But of course they are tightly interlocked, through the coupling between surface form and logical form, between logical form and the ultimate meaning representation (mediated by discourse context), and be­ tween the ultimate meaning representation and subse­ quent conversational (and other) behavior; as well as through the crucial role played by world knowledge in all of these transductions.

Compared to the 70's and early 80's, present prospects for principled, integrated NLU are greatly improved. Considerable strides have been taken in our understand­ ing of all aspects of language processing: e.g., grammar, parsing, theories of discourse structure and intention, etc. Most importantly from our perspective, new logical frameworks such as DRT, situation semantics and type and property theories provide new insights into various long-standing semantic conundrums, such as the seman­ tics of events, anaphora, properties, and attitudes.

How, then, do we propose to go about "putting it all together"? We find that the various "building blocks"

offered by various authors and schools of thought simply do not fit together in any obvious way, and leave numer­ ous gaps. What we have therefore been attempting is a creative synthesis of a modularized NLU framework, with progressive refinement of the overall abstract architec­ ture and gradual "debugging" and theoretical deepening of the entire account. Lest we lose sight of our computa­ tional objectives, we have at all times maintained a tight link between theoretical development and computer im­ plementation (the latter culminating in the EPILOG rea­ soning system) .

We feel that what we have accomplished so far vin­ dicates our methodology. The current framework con­ sists of a GPSG-like approach to grammar with a Montague-like coupling of phrase structure to logical form; "preference" -based methods of parsing and scope disambiguation; a new theory of tense and aspect "dein­ dexing"; a combined semantic representation/knowledge representation; and methods of making deductive and probabilistic inferences in both input-driven and goal­ driven modes in this representation. For good reason, the semantic/ knowledge representation, called Episodic Logic (EL), has turned out to be the centerpiece of our effort . After all, it is the choice of representation which determines how easily we can derive content from surface form, how fully we can capture the semantic nuances of NL text, and how readily we can perform needed infer­ ences.

2 Representation: The Central Problem

In speaking of a semantic representation or knowledge representation, we take for granted that the represen­ tation should admit not only straightforward intuitive interpretation but also a formal denotational semantics. As long as we remain unclear about what sorts of things in the world our symbols can stand for, or how a puta­ tive knowledge base can conform with (or deviate from) how things actually are, we risk having the system lapse into total inconsistency and incoherence for all but triv­ ial knowledge bases. At the same time we should not from the outset demand absolute hygiene and mathe­ matical perfection from a representational logic which attempts to encompass all the semantic phenomena of NL which have occupied linguists and philosophers for decades, and for which no agreed-upon unified frame­ work exists. In fact, EL initially had little more than a tentative syntax, ontology, and type structure, but we have now gained an understanding of semantic entail­ ment, finding soundness proofs for some of the inference methods. In the meantime, however, we have never com­ promised expressive adequacy for theoretical thorough­ ness. Instead, all the while, the expressive completeness of EL has allowed us to experiment freely with the syn­ tax/logical form interface, tense and aspect deindexing, and inferences based on stories and conversations.

We recognize that attempts at comprehensive theoriz­ ing and system-design have not been completely aban­ doned. For example, Hobbs et al. [8; 9] and Charniak et al. [3; 4] have kept their sights on full understand­ ing, and in particular, have made very intriguing pro­ posals for full integration of all types of disambiguation, based on abduction. However, the overwhelming major­ ity of NLU researchers have in recent years shown a cu­ rious reserve toward full understanding, most strikingly in their approach to the core problem of semantic rep­ resentation. The existing approaches fall by and large into either of two categories, theory-free approaches or FOL approaches. In a theory-free approach, a repre­ sentation with no known formal semantics is employed. Consequently, the user of the representation is free to claim that a rich variety of concepts and ideas can be ex­ pressed in the representation- for instance, beliefs , ac­ tions, goals, habitual behavior, etc. -without ever hav­ ing to confront technical criticisms about internal consis­ tency of the representation, or what follows from given knowledge. We are not saying that such work is neces­ sarily bad- it is often quite appealing and persuasive with respect to the issues it addresses; but the avoid­ ance of the semantic issue, in any serious sense, leaves the overall framework excessively ill-defined.

FOL approaches limit the fragments of language con­ sidered to those which appear to be expressible in FOL, at least in rough-and-ready fashion. This has the ad­ vantage that FOL is well-understood syntactically and semantically, but also has the disadvantage that very little real language is easily expressible in it. Consider, for instance, the first sentence of the dialog which was processed by the initial TRAINS-implementation [1] :

"We have to make orange juice"

This simple sentence exemplifies the following interest­ ing semantic phenomena: it expresses an obligation (and puts that obligation, in part , on the hearer); it expresses joint agency; it is tensed (thereby expressing that the joint obligation is upon the hearer and speaker at the time of speech); it involves a mass noun phrase, "orange juice," which on most modern accounts denotes an ab­ stract kind; it involves a verb of creation, "make," which in combination with its kind-denoting object implies the coming-into-being of some quantity of orange juice (this coming-into-being cannot be expressed by existen­ tial quantification); and it involves an infinitive whose interpretation is arguably a reified property. While with practice one can become quite good at inventing ad hoc FOL-approximations to given sentences, it is quite im­ plausible that any reasonably simple, systematic trans­ duction from syntax to semantics would give an FOL translation of this sentence. We find this situation quite typical for real spoken and written language. So the FOL approach factors out most of language- at least for any algorithmic (as opposed to ad hoc) mapping from syntax to semantic representation. Thus, we suggest it is time to break away from the restrictiveness of FOL, and set­ tle for nothing less than NL-like expressiveness (without retreating to theory-free representations) .

We now briefly review EL and its role in a comprehen­ sive, modular approach to NL U. EL is a first-order logic that is very expressive, formally interpretable, and easily derived from surface utterances, yet allows efficient in­ ference . As mentioned, it is based on a Montague-style coupling between syntactic form and logical form, while incorporating from situation semantics the idea that sen­ tences describe situations (events, states, episodes, eventualities, etc.). 1 Moreover, all of this is implemented in at least a preliminary way. The EPILOG system [17], the computer implementation of our logic, makes quite com­ plex inferences, e.g., with utterances from the TRAINS domain [1], telex reports for aircraft mechanical prob­ lems in the ARMS (the Boeing Commercial Airplane Re­ liability and Maintainability Project) domain [16], and excerpts from Little Red Riding Hood story [19]. The in­ ferences in these experiments are made quite straightfor­ wardly, despite the richness of our logic, or-we would argue- because of it, and the knowledge they employ is uncontrived (it corresponds quite directly to English sentences, and each individual piece of knowledge is for­ mulated at a maximally general level, rather than being particularized to the needs of a specific story) .

In the rest of this paper, we first describe EL syn­ tax and inference rules, using examples from the ARMS application domain, and then discuss how one can get episodic logical form from English input .

3 The Episodic Logical Form

The most important feature of EL is its NL-like expres­ siveness . This makes it easy to derive EL-translations of English sentences , while also providing a basis for concise, easily understood inferences. Its syntax allows lambda abstraction, restricted quantifiers, propositional attitudes, predicate and sentence modifiers, kind forming operators, nominalization operators, action abstraction, DRT-like anaphoric variables, generic conditionals, and other non-standard constructs. Most importantly, how­ ever, it makes use of episodic variables in the representa­ tion of episodic sentences, making implicit temporal and causal relationships between situations explicit.

To give an idea of the syntax, we show below an ELF representation of the sentence "A large bird bumped into the left wing of (airplane) VB7, causing it to get a crack" (with certain simplifications). 2

[(3el : (e1 before u!] [(The x1: [[x1 ((attr left) wing)] A [x1 part-of VB7]] (3 Yl : [Y! ((attr large) bird)] [Y! bump-into x!])) ** e!]) A (3e2 : [e1 cause-of e2][(3zdzl crack][x1 get z!]) ** e2])]

This sentence introduces three episodes: u1 , the utter­ ance event of the above sentence itself; e1 , an episode of "large bird Yl bumping into the left wing x 1"; and e2 , an episode of "wing x1 getting crack z1." Note the clause [e1 cause-of e2] that shows the causal relationship between the two episodes. The occurrence of e1 outside its quanti­ fier scope is allowed thanks to the parameter mechanism of EL (more on this later). '**' is an episodic modal operator, such that (<II ** 77] means formula <II charac­ terizes (or, completely describes) episode 7J . A weaker form of this modal operator is '*'· [<II * 77] means that <II partially describes (or, is true in) 7J . This is similar to the f= ("support") relation of situation semantics [2; 6], except that we are relating sentence intensions (par­ tial mappings from situations to truth values) , rather than "infons," to situations.

1 We use the term "episode" to emphasize the transient nature of the situations that commonly occur in narratives.

2 Below, attr is an operator that transforms a predicate into an attributive predicate modifier. Note that we use infix notation for readability and restricted quantification of the form (Q a : ~'II), where Q E {V, 3, The, Few, Most, ... }, a is a variable, and restriction ~ and matrix W are formulas. Read­ ers are referred to [10; 14] for details of EL syntax.

The following examples illustrate attitude predicates as well as kind abstraction, action abstraction , and event abstraction. That, K, Ka, and Ke below are nominal­ ization operators: That maps sentence intensions to propositions, 3 K maps predicates into "kinds" of prop­ erty, Ka maps predicates into kinds of action or attribute , and Ke maps sentence intensions into kinds of event.

• The mechanic doubted that the crack indicated danger.
• The mechanic tried to repair the crack. (3e5 : [e5 before u5] ((The x3: [x3 mechanic] (The y3: [Y3 crack] (x3 try (Ka (repair YJ))])) ** e5])
• The inspector suggested that an expert examine the hydraulic seam.

(3e3 : (e3 before u3] [[y2 indicate (K danger)] ** e4]))]))

[(The x2: (x2 mechanic] (The y2: (y2 crack] [x2 doubt (That (3e4: [e· same-time e3] ** e3])

(3es: (es before us] ((The x4: [x· inspector) (The y4: [Y4 ((attr hydraulic) seam)] [x· suggest (Ke (3z4: [z4 expert] [z4 examine y4]))])) ** es])

The DRT-like treatment of indefinites is illustrated by the donkey sentence, "Every inspector who found a crack longer than 2 inches immediately reported it" :

(3e7: [e7 before u7] ((Vx5 : ([x5 inspector] A (3es: ([es before u7] A [es during e7]] ((3y5: [(y5 crack] A [(length-of Y5 inch)> 2]] (x5 find y5]) ** es])] (3eg: (eg imm-after es][(x5 report Y5] ** eg])) ** e7]).

Note existential variables y5 and e8 that occur outside their quantifier scope. The interpretation of such free variables is much like that of parameters in DRT. Such anaphoric variables also play a role in generic condition­ als, which, when used with probabilistic inference rules, allow us to express unreliable generalizations such as the following (with some simplifications).

[(3e [(3x: [x crack] ((length-of x inch) < .25]) ** e]) --.75,e,x [(-.[x dangerous])* e)]

A crack shorter than 1/4 inch usually is not dangerous .

The subscript '.75' in the above conditional indicates a lower bound on statistical probability, and e, x are controlled variables. The rule says, roughly, that in at least 75% of the cases in which the antecedent is true, the consequent will also be true.

3 In EL ontology, propositions subsume possible facts (see [10; 12; 14]).

We need to remark here that unlike Davidsonian [5] event variables, our approach allows for episodes in­ volving quantification, such as "Most planes were more than 10 years old, which indicated that the airline com­ pany had been in financial trouble," or episodes involv­ ing negation, such as "The mechanic did not report the crack; as a result, he was fired ." Note that such com­ plex episodes frequently occur as causal antecedents, or are anaphorically referred to or quantified over, in the ARMS (and other) applications.

Our treatment of temporal and locative adverbials views these as providing conjunctive information about the described episode, much as in [7]. For instance, "Air­ plane VB7 executed an emergency landing in Chicago yesterday" is represented as

(3elo: [e1o before u1o] [[[e1o during (yesterday-rei-to u1o)] 1\ [e1o in-loc Chicago] 1\ (3x6: [x6 ((nn emergency) landing)] [VB7 execute x6])] ** elO]).

Above, nn (standing for noun-noun modification) is an operator that transforms a predicate to a predicate mod­ ifier (cf., the nn operator of Hobbs et al. [9]).

Space limitations prevent a discussion of formal se­ mantics (but see [10; 14]), but we should remark that unlike situation semantics, EL is based on an ontology that allows possible situations. These are much like "par­ tial possible worlds," in that symbols are assigned partial extensions (and antiextensions) relative to them.

4 Inference Rules in Episodic Logic

The main inference rules we have developed are Rule In­ stantiation (RI) and its dual Goal Chaining ( GC), which resemble forward and backward chaining rules in expert systems. There is also another class of goal-directed methods that consists of standard natural deduction rules such as proving a conditional by assuming the an­ tecedent and deriving the consequent, or proving a uni­ versal by proving an arbitrary instance of it . Here we only discuss RI and GC .

Rule Instantiation (RI)

Rule Instantiation, which is heavily used in input driven inference, allows arbitrarily many minor premises to be matched against arbitrarily deeply embedded subformu­ las of a rule. It subsumes modus ponens and modus tal­ lens , but can also instantiate generic conditionals. In the unit probability version, with just one minor premise ("fact"), the RI rules are: 4

where u unifies <I>, \If . R stands for "Rule," and F for "Fact." T and j_ are truth and falsity respectively. The

4 R. Waldinger (personal communication) pointed out to us that these rules are essentially the same as those of Trau­ gott [21], except that we entirely avoid skolem functions.

+ and - signs are intended to indicate positive and neg­ ative occurrence of the embedded <I>, 'lT formulas being unified. 5 Unification is defined in a way that allows sub­ stitution for explicitly quantified, "matchable" variables. A variable in a rule or fact is matchable if it is bound by a positively occurring universal quantifier or negatively oc­ curring existential quantifier. For instance, substitution of w for x in a positively embedded subformula (Vx: [x P] [x Q]) yields [[w P] --+ [w Q]], and the same substitu­ tion in a negatively embedded subformula (3x: [x P] [x Q]) yields ((w P] 1\ [w Q]]. The first rule is sound if~ contains no unmatchable free variables which are bound in F as a whole. The second rule is sound if <I> contains no unmatchable free variables which are bound in R as a whole. So in particular, the first rule is sound ifF contains only constants and top-level universal (hence matchable) variables. (See [18] for soundness proof.)

As mentioned, the rules also work for generic condi­ tionals , and we illustrate this case. Consider the follow­ ing generic conditional:

If an aircraft less than 3 years old has a crack, usually the crack is not due to corrosion.

Suppose now the translation of a telex contains the fol­ lowing:

[((VB7 aircraft] 1\ ((age-of VB7 year)= 2] 1\ (C4 crack] 1\ (C4 located-on VB7]] ** El] The two year old aircraft VB7 has a crack.

Then, RI matches these formulas against the antecedent of the conditional, unifying VB7/x, C4/y and El/e, and derives (aided by "specialists" in EPILOG):

where the superscripted number, .8, is a lower bound on epistemic probability. (That is, the system concludes that the crack C4 is not due to corrosion, with minimum degree of confidence .8.) For this simple example, ei- · ther version of RI yields the correct inference. This kind of inference, based on unreliable generalizations, allows evidence for explanations or predictions to be weighted, much as is done in expert systems.

Goal Chaining (GC)

Goal chaining, which dominates goal-driven inference, is a pair of very general chaining rules. Chaining from rule consequents to antecedents is a special case. The following are the GC rules in the unit probability case:

where u' "antiunifies" <I> , \If (i.e., with positive existentials and negative universals in G regarded as match­ able). R stands for "Rule," and G for "Goal." The first rule is sound if <I> contains no unmatchable free vari­ ables which are bound in R as a whole. The second rule is sound if \l1 contains no unmatchable (e.g., top-level universal) variables which are bound in G as a whole.

5~ is positively (negatively) embedded in another for­ mula if the only operators embedding <I> are among {...,, ..... ,/\, V, V, 3}, and the number of occurrences of · , the antecedent scope of ..... , and the restriction scope of V among the operators embedding ~is even (odd).

For example, consider a rule

[(3x1: [x1 ((attr hydraulic) joint)] (3el [(3xdx2 crack] [x2 located-on x1]) ** ei])) --+ [(3ede2 same-time ei] [[(age-of XI year) > 5] ** e2]) V (3ede3 before ei] [[ x 1 receive (K ( (at tr excessive) pressure))] * * e3 ])]] If a hydraulic joint has a crack, either it is more than 5 years old, or received excessive pressure.

Suppose, now, we want to know if a particular hydraulic joint, Joint6, received excessive pressure. Then the goal can be posed as

? (3e: [e before El] [[Joint6 receive (K ((attr excessive) pressure))]** e]),

with a constraint that E1 is at the same time as the ut­ terance time. Since the goal matches the second disjunct of the consequent of the rule, with unification Joint6/ x1 , El/e1 , and e/ e3 , we replace the matching part of the rule with ..L, and negate the whole formula. Then we get a new goal

? (· [[(3xdx2 crack] [x2 located-on Joint6]) ** El] ---+ [(3ede2 same-time El] [[(age-of Joint6 year) > 5] ** e2]) V .L]]), which, after simplification, becomes ? [[(3xdx2 crack] [x2 located-on Joint6]) ** El] 1\ (·(3e2:[e2 same-time El]

[[(age-of Joint6 year) > 5] ** e2]))].

The new goal says, "Joint6 has a crack and is not older than 5 years ." This makes sense because one way of proving "Joint6 received excessive pressure" is to prove that "Joint6 is not older than 5 years and has a crack." As with the previous RI example, either version of GC gives us the same new subgoal.

The general version of GC allows arbitrarily many sub­ sidiary knowledge base facts to be invoked in the pro­ cess of chaining from the given goal to a subgoal. For instance, in the above example, knowledge base fact [Joint6 ((attr hydraulic) joint)] led to immediate success .

For both RI and GC , the second rule remains to be im­ plemented . Yet, the kinds of EL inferences described so far are carried out efficiently by the EPILOG system [17], a hybrid reasoning system combining efficient storage and access mechanism, forward and backward chaining, an agenda-driven control structure, and multiple "spe­ cialists" for taxonomies, temporal reasoning, etc.

5 Computing Episodic Logical Form

As we have emphasized, a crucial feature of EL with re­ spect to the goal of building general NLU systems is the ease with which EL-representations are derived from sur­ face syntax. This is made possible by its NL-like expres­ siveness.

The initial translation from phrase structure to the preliminary indexical logical form (LF) is accomplished with GPSG-like syntactic and semantic rules; the final nonindexical episodic logical form (ELF) is obtained by simple recursive deindexing rules . Such a transformation is essential because, to be useful for inference, a situa­ tional logic need to be nonindexical. Our deindexing al­ gorithm systematically handles tense, aspect, and many temporal adverbials and their interaction, and brings the context information into the logical form, removing con­ text dependency.

For example, the logical form of sentence

"A mechanic repaired the crack yesterday"

is easily computed using the following (somewhat simpli­ fied) lexical and phrase structure rules , annotated with corresponding semantic rules. (We use a feature system hierarchically organized as a tree; see [10].)

DET[indef] +- a; 3 DET[def] +- the ; The N +- mechanic; mechanic N +- crack ; crack NP +- DET N; <DET' N' > V[past, ..NP] +- repaired; <past repair> VP +- V[..NP] NP ; (V' NP') NP[def-time] +- yesterday; Yesterday ADV[ep-mod] +- NP[def-time]; (during NP') ADVL[post-mod] +- ADV[ep-mod] ; WAx ((adv-e ADV') [x P]) VP +- VP ADVL[post-mod]; (ADVL' VP') S +- NP VP; [NP' VP']

Applying these rules gives us the initial, unscoped logical form ULF ( 1) shown below. Angle brackets in ( 1) indicate unscoped expressions. Notice the indexical term Yester­ day and the indexical sentence operator past in (1). adv­ e (standing for "episode modifying adverbial" ) is an op­ erator that maps 1-place predicates over situations into sentential operators. (2) is a scoped LF; past has been raised so that it has a scope over the whole formula. The deindexing step from LF (2) to ELF (3) is carried out by the tense-aspect deindexing rules of EL that use tense trees as components of a discourse structure. Note that the deindexing step introduces an episodic variable through tense operator past into ELF (3). Also in (3), the indexical constant Yesterday is disambiguated to yester­ day relative to the utterance episode, and crack x 1 is resolved to CrackS (the latter kind of reference resolu­ tion is not currently dealt with in our system). ,y, is an extension operator that maps 1-place predicates over sit­ uations into sentence intensions. Intuitively, v 1r means 1r is true of the current situation, which is e1 in this case. Thus, (3) is identical to ( 4), which may be obtained via meaning postulates.

(1) ((adv-e (during Yesterday)) [<3 mechanic> <past repair> <The crack>]) (2) (past ((adv-e (during Yesterday)) (The x1: [x1 crack] (3yi: [YI mechanic] [YI repair x1])))) (3) (3ei :[e1 before ui] [[v (during (yesterday-rei-to u!)) 1\ (3yl :[y1 mechanic][y1 repair CrackS])] ** ei])

Our deindexing mechanism is compositional in that operators past, futr, perf, etc., contribute separately and uniformly to the meanings of their operand formulas, driving the generation and traversal of tense trees in deindexing. As an example, we show the Past-rule be­ low.

T denotes a tense tree, e1 a "new" episode symbol (i.e., not yet used in T). Embr, Lastr, /T, OT, etc., are easily computed functions on the tense tree (see [11; 20] for details). The recursively deindexed <I> is taken to characterize the new episode er, which is predicated to be at or before ( "befr") the embedding episode Embr, e.g., the utterance event (depending on the aspectual class of <I> and other factors). Tense trees also provide the "points of orientation" ( cf., [15; 22]), such as the ref­ erence point for a perfect episode or the immediately pre­ ceding past episode in a succession of simple past-tensed sentences, and the "orients" predication captures this.

As mentioned, the information in the past operator and in the time adverbial are essentially interpreted con­ junctively as in [7] . For detailed discussion of the in­ terpretation of temporal and other adverbials, see [10; 13] . Although we do not yet have grammars and LF rules for large subsets of English, the compositional ap­ proach of EL has been proven practical by a start-up im­ plementation that has successfully handled dialoe;ues in a planning system in the TRAINS domain [1]. See l10] for some of our grammar fragments . The ARMS application used a grammar and interpretation mechanism internal to the Boeing Co.

6 Concluding Remarks

We think there is cause for optimism about the possibil­ ity of constructing theoretical and computational frame­ works for full NL U. Our own efforts in that direction have led to a rather well-integrated conception of syntax, LF , knowledge representation, context, and inference, and of the interfaces linking these into an organic whole. The conception is not yet complete or fully "debugged," but it is sufficiently far along to have provided a basis for diverse start-up implementations. Unlike most past im­ plemented NLU and inference systems, these implemen­ tations strenuously avoid cutting corners in syntax, LF computation, and most of all, knowledge representation and inference. Thus, we have reason to regard the the­ oretical framework and the implementations as a solid and extensible basis for further work toward the ulti­ mate goal of general NLU.