Natural language understanding involves lots of hard issues such as various types of ambiguities, indeterminacies caused by ellipses or fragmental utterances, or ill-formedness. Being confronted with these difficulties, it does not seem reason able to seek for a method of logically deducing the speaker's intended meaning or plan from utterances. Instead, it is much more natural to characterize natural language understanding as an abd active process of exploring most plausi­ble interpretation which can explain given ut­terances.

In this paper, we present an abductive in­ference mechanism, called the integrated pars- ing engine, for natural language understanding. The integrated parsing engine is able to:

• make plausible assumptions at appropriate time

e reason with multiple alternatives based on different sets of assumptions

• switch the search process to the maximally plausible alternative

• detect contradiction resulting from incon­sistent assumptions and eliminate all con­flations which depends on these assump­tions

® update plausibility factor of each belief based on new observations.

Thus, the integrated parsing engine is general enough to carry out linguistic and nonlinguis tic inferences in a uniform manner, by drawing information from various sources: syntax, se­mantic, discourse, pragmatics, or real world.

In the remainder of this paper, we first de­scribe mechanisms for maintaining consistency and plausibility. We then show how these two mechanisms interact to guide the inference pro cess. Finally, we use an implemented exam­ple to demonstrate how the integrated parsing engine is used to interpret sentences by taking contextual factors into account.

The (/ME (Consistency Maintenance Engine) is a. component of the integrated parsing engine responsible for maintaining consistency among beliefs. Basic design principles of the CME is based on de Kleer's ATMS (Assumption-based Truth Maintenance Engine) [de 86]. The CM]'] maintains a set of alternative be­liefs, each of which, consists of a set of as­sumptions and their conclusions, as follows:

cnvir onment conclusions An external problem solver is assumed to exist which makes assumption, adds conclusion, and detects contradietion.

The main task of CME is to maintain alterna­tive beliefs by removing all alternatives whose net of assumptions has turned out contradic­tory. Like ATMS, the CME takes advantage of the following monotonie property:

if a contradiction is derived from a set of assumptions A, then contradiction is also derived from any set of assump­tions B such that B D A.

Thus, if contradiction is derived from a set of assumptions {0S.D}, alternative interpreta­tions depending on sets of assumptions such as {B, C, D), {A, B, £>}, {A, B, Ct D\, ... are re­moved. In addition, the CME keeps records of contradictory sets of assumptions to prevent any interpretation depending on them from be­ing considered in future.

Unlike ATMS whose control regime is bread-first, our CME uses a tree called the environ­ment tree, or the E-tree for short, to guide the search process. Each node of the E-tree rep­resents an environment, a set of assumptions. Each arc of the E tree represents that a lower node is derived from the upper node by mak­ing one more assumption. Thus in figure 1, Eq is the root node, and it represents an environ-mnet without any assumption. Nodes below &o represent environments with one or more assumption added to its parent node's envi­ronment. Thus, Ei -~ Eq U Mi} = En ~ Et U {An} = {/ii, /In}, and so on.

We assume that a set of assumptions made at the same parent node are mutually exclusive. Although this is a rather strong assumption, it makes sense in natural language understand­ing .since many assumptions being made dur­ing the natural language understanding process are mutually exclusive. Even if this is not the case, any set of assumptions can be transformed into a set of mutually exclusive assumptions by adding appropriate conditions. Although this is a cumbersome solution, it does not often take place in natural language understanding and most importantly it saves the amount of com­putation.

Note that the CME alone cannot determine which way to go when there is more than one possibility of extending the set of beliefs. This information is provided by the PME, as de­scribed in the next section.

The PME (Plausibility Maintenance Engine) maintains estimations of how plausible each en­vironment is. This information is given as con­ditional probabilities and it is kept as annota­tions to each arc of the E-tree. Thus, in figure 2, which is a slightly more precise version of fig­ure 1, pi stands for P(Ei), pij for P(Ej\Ai), Pijk for P{Ek\Ai, Aj), etc.

It follows from the property of conditional probability that if i -fi j and E, and Ej are immediate children

(a) initial B-tree Linguistic and Nonlinguistic Pioblem Solver (b) The E-tree after -iEt is observed.

E3 E4 EsFigure 3: A Sample E-tree with Annotation

if Ej is a parent node of .

Initial value of p^s are to be given from the external problem solver. The PME's role is to maintain estimation of plausibility by taking into account given observations. Currently we only take -i ./■/', the event of environment E run­ning into contradiction, as an observation. We use a Bayes' law to modify P(A) into P(A\-iE). Thus, if Ei and Ej aie brothers, (l) is further simpli­fied to:

For example, suppose it has turned out that environment e4 is in contradiction and hence ~<Ei is observed (figure 3(a)). The annotations to the E-tree are updated as in figure 3(b). Notice that the update of conditional proba­bility can be done based on local information.

Knowledge Base

Problem Solving Engine (PSE)

Working Memory

Associative Networks Previous Topic

The Integrated Parsing Engi

The Integrated Parsing Engine (CME) Plausibility Maintenance Engine (PME)

E-true ft .

Figure 4: The Structure of a Natural Lan­guage Understanding System with the lute grated Parsing Engine as a subsystem

4 Natural Language Un­derstanding System Us­ing the Integrated Pars­ing Engine as a Subsys­tem,

The integrated parsing engine consists of the CME and the PME. The architecture of a natu­ral language understanding system with the in­tegrated parsing engine as a subsystem is shown in figure 4.

The knowledge base contains various types of information for language comprehension, in eluding lexicon, morphology, syntax, semantics, discourse, pragmatics, commonsenses, and so on. The whole system is controled by the prob­lem solving engine (PSE). The PSE can access to the knowledge base and use the integrated parsing engine as an aid to seek for most plau­sible interpretation. Input texts are analyzed in a sentence-by-sentence manner. The discourse structure is maintained as a previous topic in the working memory.

When it scans a new sentence, the PSE first initialize the E-tree with only the root node. Then the PSE repeats the following cycle:

(step 1) choose a leaf node with the high est probability as a working enviri ornent (step 2) repeatedly derive conclusions from

the library the xerox the meeting room room room

believed propositions until either (a) the goal is achieved, (b) contra diction is derived, or (c) no more conclusion is derived unless malting more assumption.

In case (a), the process halts.

In case (b), the process is passed to the PME, which modifies current es­timation of plausibility so that this fact is reflected, then an alternative of mexhuuui plausibility is chosen and is suggested to the CME.

In case (c), the process also is passed to the PME, which assigns plausibil ity to new nodes, and working envi­ronment is chosen again.

The integrated parsing engine has been writ­ten in Lisp. It is running with a small exmeri-mental grammar for Japanese. The next section shows how it works.

Suppose a dialog environment in which a pro lessor speaks to a clerk to borrow a key of some rooms (figure 5) and utters the following Japanese sentence:

The referential meaning of this sentence is ambiguous if there is more than one key in a given situation. Suppose three keys are there: keyl for a library room, key2 for a xerox room, and keyS for a meeting room.

Although sentence (3) is ambiguous in nor­mal contexts, it becomes much less so if it fol­lows sentences like:

Even if no previous sentence is spoken, sen tence (3) is acceptable in a situation where the speaker and the hearer mutually believe that the xerox room is accessed so often that "the key" is usually used to refer to key2, the one for the xerox room.

Note that the omission of the patient case does not matter in usual situations, since there is a strong default that the filler of this case is the speaker.

Now let us show how sentence (3) is ana­lyzed in a context where sentence (4) was pre­viously uttered. The task of analyzing input starts from recognizing words. Lots of ambi­guities arise in this phase. For sentence (3), *KA' might be a single word lKA' (postposi­tion marking interrogative) or a part of a longer word 'KAGI' (key). Since longer match is con­sidered to be more plausible in general case in Japanese analysis, we assign larger number of probability to the latter possibility. Following this analysis, the PSE makes the assumptions to the integrated parsing engine:

<§>word~j (take the sequence 'KA' as a word): probability 1/3.

®\nord-2 (take the sequence 'KAGI' as a word): => probability 2/3.

Accordingly, the CME extends the initial E-tree as in figure 6. Since, the environment E\ has the highest plausibility, the CME chooses it for the next environment and control is returned to the PSE.

This paper was inspired by a number of works. A massively parallel parsing by Waltz and Pollack IWP85] has demonstrated the effect of integration through a uniform computa­tion mechanism (marker passing) in context-dependent comprehension of discourse. They have pointed out the importance of non-logical, associative relation between concepts. Char-niak has pointed out the abductive nature of language comprehension. Charniak's Wimp [Cha86] uses a marker passing mechanism as a basis of abductive inference engine for lan­guage comprehension. But it is not used alone; it is augmented by a logical process called path proof. In a parser used in Lytinen's MOP-TRANS [Lyt86], a mechanism is provided to allow close interaction between syntax and se­mantics, while keeping the modularity of the system. Another thing to note is that Lytinen's integrated parser makes use of strong semantic expectation to constrain the search.

The integrated parsing engine presented in this paper takes advantages of these preced­ing works. Unlike Waltz and Pollack, and like Charniak and Lytinen, our integrated parsing engine has a hybrid architecture for logical and non-logical inferences. What is novel with our integrated parsing engine is the method of inte­grating and maintaining logical and non-logical information obtained from various source. In

We have presented an inference engine for inte­grated natural language understanding, based on a characterization of natural language un­derstanding as an abductive process. The essence of our approach is connecting con­sistency maintenance engine and plausibility maintenance engine closely enough to allow their dense interaction. Although we have shown rather "low level" issues, we believe the same idea is applicable to "higher level" prob­lems such as inferring speaker's intention and plan.