A Parallel-Process Model of On-Line Inference Processing

A Parallel-Process Model of On-Line Inference Processing

Kurt P. Eiselt

lrvine Computational Intelligence Project Computer Science Department University of California lrvine, California 92717

does ATLAST, seeks to integrate lexical, syntactic, and pragmatic inference processing [Charniak, 1983], though his model differs from ATLAST in other respects.

2.0 Background: The Theory in Brief

The theory behind ATLAST is described in detail in (Granger, Eiselt , & Holbrook, 1985], but a brief review of the theory is provided here to aid in understanding the program.

ATLAST is a direct descendant of earlier work on inference decision processes at the pragmatic Ievel. Specif­ ically, it came about as an attempt to address word-sense ambiguity problems which arose during research into dif­ ferent pragmatic inference strategies used by human sub­ jects while reading text, and the development of a pro­ gram, called STRATEGIST, which modelled that behav­ ior [Granger, Eiselt , & Holbrook, 1983; Granger & Hol­ brook, 1983]. As we worked on STRATEGIST, we observed that lexical and pragmatic inference processes appeared to have much in common. Many pragmatic inferences seemed to be triggered by individual words. This is hardly new news, of course, as there exist integrated models of lan­ guage understanding in which higher-level inferences are directly activated by input text (FRUMP [DeJong, 1979] and IPP [Lebowitz, 1980] are notable examples). We be­ lieved, though, that the relationship was even closer than described by previous modeis- that the inference decision mechanisms themselves were in some way interdependent at the very least . For example, in the text

The CIA called in an inspector to check for bugs . The secretariea had reported seeing roaches.

the first sentence alone has an unambiguous interpreta­ tion: the "hidden microphone" sense of "bugs" is more appropriate than the "insect" sense. Upon reading the second sentence, the "insect" sense is obviously more ap­ propriate, and the initial choice of word-sense for "bugs" must be supplanted. To explain this process, we theorize that "CIA" triggers a pragmatic inference about espionage which in turn influences the choice of the "hidden micro­ phone" word-sense for "bugs". Later, the word "roaches" generates higher-level inferences which suggest that the "in­ sect" sense of "bugs" is correct and that the Central Intel­ ligence Agency, in this case at least, is more appropriately

Abstract

This paper presents a new model of on-line inference processes during text understanding. The model, called ATLAST, integrates inference processing at the lexical, syntactic, and pragmatic Ievels of understanding, and is consistent with the results of controlled psychological ex­ periments. ATLAST interprets input text through the in­ teraction of independent but communicating inference pro­ cesses running in parallel. The focus of this paper is on the initial computer implementation of the ATLAST model, and some observations and issues which arise from that im­ plementation.

1.0 lntroduction

This paper describes a new theory of inference pro­ cessing developed at the lrvine Computational Intelligence Project, and an initial computer implementation of that theory. The research described here integrates inference processing at the lexical, syntactic, and pragmatic Ievels, and is consistent with the results of controlled psychologi­ cal experiments. The theory centers upon a parallel-process model of text understanding which explains inference be­ havior at the different Ievels as the result of interactions between three independent but communicating inference processes. Though there are three processes operating at three different Ievels of language understanding, there is no direct correspondence between the Ievels and the processes. lnference decisions at all Ievels are made through the com­ bined actions of the three processes running in parallel. We call this model ATLAST (A Three-level Language Analysis SysTem).

ATLAST represents a real departure from most pre­ vious models of language understanding and inference pro­ cessing [e.g., Schank, 1975; Cullingford, 1978; Wilensky, 1978; DeJong, 1979], though there are models which inte­ grate some of the Ievels of inference processing. For ex­ ample, IPP (Lebowitz, 1980) and BORIS (Dyer, 1982) inte­ grate the syntactic and pragmatic Ievels, while the model of Small, Cottrell, and Shastri [1982] integrates lexical ac­ cess and syntactic parsing. Finally, Charniak's model, as viewed as a generic employer trying to rid itself of pests rather tban as an espionage agency protecting its secrets. Tbus, tbe context generated by "CIA" determined the se­ lection of a word-sense for "bugs", while the context gener­ ated by "roacbes" resulted in an entirely new interpretation of "bugs" and a slightly modified interpretation of "CIA". Because of tbis interdependence between inference Ievels, tbeories about pragmatic infereoce mechanisms must in­ clude tbeories about lexical access processes.

Tbis researcb was supported in part by tbe National Science Founda­ tion under grant IST-81-20685 and by tbe Naval Ocean Systems Center under contracts NOOI23-81-C-1078 and N66001-83-C-0255.

L€xical access is tbe process by wbich a word's mean­ ing is extracted from its written (or spoken) form. Recent researcb into lexical access bas led to the counter-intuitive cooclusion that when an ambiguous word is presented in context (i .e., a sentence or phrase), all meaoings of tbe word are initially accessed, and context is subsequently consulted to determine the most appropriate meaning [Swinney & Hakes, 1976; Taneobaus, Leiman, & Seidenberg, 1979; Lu­ cas, 1983; Granger, Holbrook, & Eiselt, 1984]. This hap­ pens regardless of tbe syntactic category of the word, or wbetber the context is biased toward one meaoing or an­ otber.

If the lexical access process does in fact work as de­ scribed above, and if individual words trigger the higber­ level pragmatic inferences, tben it is likely that the prag­ matic ioference decisioo process is much tbe same as the lex ical inference decision process. Work oo ATLAST goes under the assumption that, when more than one interpre­ tation (i.e., pragmatic inference) of an input text is possi­ ble, all possible interpretations are pursued in parallel, and those interpretations which do not fit well with the existing context are "de-activated" or inhibited.

3.0 How ATLAST Works

3.1 Memory

ATLAST is built around a high- Ievel episodic memory structure which contains two kinds of memory organiza­ tion packets (MOPs) [Schank, 1982; Kolodner, 1984]. For each word in ATLAST's vocabulary there is a MOP which represents that word. Most lexical-entry MOPs contain a one-way link to one or more word-senses directly associ­ ated with that word, and syntact ic information about the word-senses. Function words, such as "a" and "the" , are not linked to otber MOPs and serve only to aid in syntac­ tic decisions. Tbe word-senses are an example of the otber kind of MOP in ATLAST's memory: tbose which represent events or objects. These MOPs are interconnected through a network of two-way links which serve to define tbe rela­ tionships between tbe MOPs. These MOPs can be, but are not necessarily, directly linked to lexical entries.

The inference decisions in ATLAST are carried out by three primary components: the Capsulizer, the Pr<r poser, aod the Filter. Theoretically, tbese processes run in parallel. However, ATLAST is written in UCI-LISP on a DECSYSTEM-20, so the parallelism which is so important to tbe theory is necessarily simulated in its implementa­ tion. This simulation is accomplished by repeatedly cycling through tbe three processes. Thus, the Capsulizer runs for a pre-determined amount of time, followed by the Proposer, then the Filter, then the Capsulizer again, and so on. The amount of time each process is allocated is an important issue with respect to the accuracy of the model. This issue bas not yet been fully explored.

3.2 Capsulizer

Tbe Capsulizer contains the first stage of a tw<rstage syntactic analysis process similar in some respects to that described by Frazier and Fodor [1978]. The Capsulizer makes intra-phrasal, as opposed to inter-phrasal, syntac­ tic decisions about the words in the input text (again, see [Granger, Eiselt, & Holbrook, 1985] for a discussion of the theory behind two-stage syntactic analysis). As the Cap­ sulizer encounters each new word in the input text, it re­ trieves the syntactic category information associated with that word (e.g. , "this word can be used as a noun and a verb") and activates any word-senses associated with that word. The word-senses arenot used in any decisions made by the Capsulizer, tbough pointers to the word-senses are retained. The activated word-senses serve as a starting point for the search carried out by the Proposer, which is described below.

As the Capsulizer processes the input words, it accu­ mulates the syntactic information it retrieves and makes initial decisions about syntactic relationships within the phrases of the input text. These intra-phrasal decisions, along with the pointers to tbe word-senses whicb comprise the phrases, are passed along to the Filter as "capsules" of information. The Filter then makes decisions about the syntactic relationships between the phrases (i .e., inter­ phrasal syntax). If an input word activates more than one word-sense (i.e., a word-sense ambiguity), the pointers to the multiple word-senses are all passed oo to the Filter, which will eventually select the "best" word-sense. This process is also described in more detail below.

3.3 Proposer

The Proposer gets its name from tbe idea tbat it "proposes" possible inference paths which might explain the input text. Essentially, it is a search mechanism which employs spreading activation to traverse the links between the MOPs in memory and find connections between word­ senses whicb bave been activated by the Capsulizer.

The Proposer maintains pointers to the most recently activated MOPs in memory, and to the word-senses which are the origins of the spreading activation search. Each time the Proposer is invoked, it traverses the links leading away from tbe recently activated MOPs, activates the adjacent MOPs at the end of those links, and updates its Iist of pointers. If the spread of activatioo from ooe poiot of origin intersects the spread of activatioo from some other poiot of origio, theo the Proposer has found some plausible relationship, by way of links and MOPs, between two (and possibly more) of the word-senses activated by the input text. The Proposer then passes information about this newly-discovered pathway to the Filter; in this way, the Proposer "proposes" possible inference paths for evaluation by the Filter.

Spreading activation has been employed in a number of models [e.g., Quillian, 1968; Fahlman, 1979; Anderson, 1983; Charniak, 1983; Norvig, 1983; Riesheck & Martin, 1985] . Spreading activation allows ATLAST to pursue mul­ tiple inference paths in parallel. Were this process allowed to continue unchecked, it would Iead to a combinatorial ex­ plosion of inference paths. To prevent this from happening in ATLAST, the third major process, the Filter, constantly evaluates or "filters" inference paths and inhibits pursuit of those which appear to be poor explanations of the input text . Though the idea of beginning pursuit on all inference paths instead of j ust the "appropriate" ones may seem both counter-intuitive and counter-productive, there are two ar­ guments for using this approach. One is that it would seem impossible to determine which inferences may be appropri­ ate without first evaluating all inference possibilities. The other is that this approach is consistent with experimental sturlies of human behavior [Tanenhaus, Leiman, & Seiden­ berg, 1979; Seidenberg, Tanenhaus, Leiman, & Bienkowski, 1982; Granger, Holbrook, & Eiselt, 1984].

The Proposer is implemented in ATLAST as a sep­ arate process, but from a theoretical perspective it might be more appropriately viewed as an emergent property of a human memory organization. Computer memory seems to work somewhat differently than human memory, though, so it was necessary to provide a separate process to make the spreading activation possible.

3.4 Filter

The Filter performs two functions; the first is that of inter-phrasal syntax. As capsules are passed from the Capsulizer to the Filter, the Filter makes decisions about the relationships between the phrases represented by the capsules. Inter-phrasal syntax rules enable the Filter to fill the Actor, Action, and Object slots, for example. Future work on the ATLAST program will add rules about mod­ ifying phrases, keeping track of referents across phrases, and agreement of tense, number, and gender, among other rules.

The Filter's other function is the evaluation o( infer­ ence paths. When two competing inference paths are pro­ posed (e.g., different paths connecting the word-senses of two words from the input text ), the Filter attempts to se­ lect the more appropriate path through the application of three inference evaluation metrics.

First, the Filter evaluates the inference paths accord­ ing to the specificity metric [Wilensky, 1983]. If one path is determined to be less specific than the other, the less specific path is inhibited; that is, the spread of activation from nodes on the path is stopped, and that path is no Ionger considered as a plausible explanation for the input text. Specificity is determined by the links in the path: a path which includes a "viewed-as" link (from the "view" re­ Jationship defined in [Wilensky, 1984]) is less specific than a path which does not contain such a link. In tbe example of Section 3.5, tbe CIA is a special case of a spy agency, but a spy agency can also be viewed as an employer; an inference patb wbicb describes tbe CIA only as a spy agency is more specific tban one wbicb explains it as a spy agency and an employer.

If tbe specificity metric fails to make a decision be­ tween two competing paths, tbe Filter applies two varia­ tions of tbe parsimony metric [Granger, 1980]. Tbe first of these variations (tbe "lengtb" metric) gives precedence to the inference patb with fewer links. Failing this, the Filter applies the other variation of the parsimony metric (tbe "explains more" metric) , wbicb examines the "inter­ mediate" MOPs (tbose wbich arenot tbe endpoints) of the two competing patbs. Tbis variation tben selects the in­ ference patb tbat contains moreintermediate MOPs wbicb are intersection points with other inference patbs. In other words, tbe intermediate MOPs can be eitber endpoints or intermediate MOPs of paths otber tban tbe two being eval­ uated.

It is witb tbe Filter tbat the implementation of AT­ LAST diverges most from the tbeory. In some sense, this is to be expected, since the Filter is tbe most complex of the three processes. In theory, ATLAST should be able to evaluate and inbibit pursuit of apparently implausible inference paths almost as soon as pursuit bas begun, tbus preventing problems of combinatorial explosion. ATLAST would accomplisb this by comparing tbe multiple, possi­ bly incomplete, inference paths wbicb begin with a speci fic word-sense to the context it bas built up to that point in the processing of the input text, and determining whicb of the paths fit "best" with that context. This would be in agreement witb experimental results in lexical access re­ searcb [Tanenhaus, Leiman, & Seidenberg, 1979; Lucas, 1983; Granger, Holbrook, & Eiselt, 1984] . At tbis time, the ATLAST model can only evaluate complete inference paths (i .e., those which connect two or more word-senses activated by tbe Capsulizer) without regard to the existing context. Tbougb this simple inference evaluation mecba­ nism seems to work for sentences such as the one presented in the following example, it will not be sufficient to properly interpret longer, more complicated texts. This problern will be rectified in the near future.

3.5 An Example

What follows is actual (though abbreviated) anno­ tated run-time output from the ATLAST prototype pro­ gram. This example illustrates primarily how ATLAST disambiguates between two possible meanings of the word "bugs" in the text, "The CIA checked for bugs." In the in­ terest of brevity and clarity, we use a very short text and just enough of a knowledge base to process this example. Due to space limitations, we will concentrate primarily on the operation of the Filter. Also, we have abbreviated the names of some of the memory structures, again due to space limitations. The following legend should make the program trace more readable:

GEN-EMPLOYER · GENERIC-EMPLOYER

GET-SECRETS = GET-OTHERS-SECRETS

MPHONE = NICROPHONE

P-HEALTHY-ENVT = PRESERVE-HEALTHY-ENVIRONNENT

P-SECRETS = PRESERVE-OWN-SECRETS

PLANT-LISTEN-DEV = PLANT-OWN-LISTENING-DEVICE

REN-HEALTH-HZRD = RENOVE-HEALTH-HAZARD

REN-LISTEN-DEV · REMOVE-OTHERS-LISTENING-DEVICE

After processing "The CIA" and activating associated memory structures, ATLAST processes "checked", which terminates the noun phrase and begins a verb phrase. Cap­ sulizer sends a capsule consisting of the word-senses initially activated by the noun phrase (i.e., C-1-A) to Filter. Filter, looking for an actor for this sentence, fills the slot with this noun-phrase capsule. ATLAST then processes "for" :

Filter :

New path discovered : IPATHO Path from C-I-A to SEARCH C-I-A is special case of SPY-AGENCY SPY-AGENCY has goal P-SECRETS P-SECRETS has plan REM-LISTEN-DEV REM-LISTEN-DEV is special case of RENOVE REMOVE has precondition SEARCH

ACTION slot filled by SEARCH

The preposition "for" does not activate any new mem­ ory structures, but it does begin a modifying prepositional phrase. Capsulizer sends the verb component of the verb phrase (SEARCH) to Filter, which then assigns the capsule to the action slot.

Proposer, looking for intersections among the "wave­ fronts" of spreading activation, finds a connection, or in­ ference path (IPATHO) , between C-1-A and SEARCH, and notifies Filter. Filter knows of only one inference path at this time, so there is no basis for comparison and evaluation of inference paths yet. ATLAST then moves on to "bugs" :

Filter :

New path discovered: IPATH1

Path from C-I-A to SEARCH

C-I-A is special case of SPY-AGENCY SPY-AGENCY can be viewed as GEN-EMPLOYER GEN-EMPLOYER has goal P-HEALTHY-ENVT P-HEALTHY-ENVT has plan REN-HEALTH-HZRD REN-HEALTH-HZRD is special case of RENOVE RENOVE has precondition SEARCH

New path discovered : IPATH2 Path from C-I-A to INSECT

C-I-A is special case of SPY-AGENCY

SPY-AGENCY can be viewed as CEN-ENPLOYER CEN-ENPLOYER has goal P-HEALTHY-ENVT P-HEALTHY-ENVT has plan REN-HEALTH-HZRD REN-HEALTH-HZRD has role-filler INSECT

New path diacovered: IPATH3

Path from C-I-A to NPHONE

C-I-A is special case of SPY-ACENCY SPY-ACENCY has goal GET-SECRETS GET-SECRETS has plan PLANT-LISTEN-DEV PLANT-LISTEN-DEV haa role-filler NPHONE

New path discovered : IPATH4 Path from C-I-A to SEARCH C-I-A is special case of SPY-ACENCY SPY-AGENCY has goal CET-SECRETS GET-SECRETS has plan PLANT-LISTEN-DEV PLANT-LISTEN-DEV has role-filler NPHONE NPHONE is role-filler of REN-LISTEN-DEV REM-LISTEN-DEV is special case of RENOVE REMOVE has precondition SEARCH

New path discovered: 'IPATH5 Path from SEARCH to NPHONE SEARCH is precondition of RENOVE RENOVE has special case REN-LISTEN-DEV REN-LISTEN-DEV has role-filler NPHONE

New path discovered : IPATH6

Path from SEARCH to INSECT

SEARCH is precondition of RENOVE RENOVE has special case REN-HEALTH-HZRD REN-HEALTH-HZRD has role-filler INSECT

New path discovered : IPATH7

Path from NPHONE to C-I-A NPHONE is role-filler of REN-LISTEN-DEV REM-LISTEN-DEV is plan of P-SECRETS P-SECRETS is goal of SPY-ACENCY SPY-AGENCY has special case C-I-A

Parsimony : IPATH7 explains more than IPATH3

Specificity : IPATH4 more specific than IPATH1

Parsimony : IPATHO shorter than IPATH4

Capsulizer reads the ambiguous word "bugs", which results in the activation of two word-senses: INSECT and MPHONE. Proposer's search has uncovered several new in­ ference paths. When two different inference paths connect the same two word-senses, Filter applies inference evalua­ tion metrics to the two paths to determine which of the two provides the better explanation of the input text. The rejected paths are de-activated until later text results in activating that path again. Finally, ATLAST encounters the end of the text :

Filter:

OBJECT has competing slot fillers:

INSECT vs . NPHONE

Specificity: IPATH7 more specific than IPATH2

Parsimony : IPATH5 explains more than IPATH6

Lexical ambiguity resolution :

NPHONE vs. INSECT

All paths through INSECT de-activated Ambiguity resolved: KPHONE selected OBJECT slot filled by KPHONE

Capsulizer sends to Filter a capsule containing the word-senses activated by the prepositional phrase. Filter determines that the capsule contains the object of the ac­ tion SEARCH, and that this object is ambiguous. Filter attempts to resolve this ambiguity by applying the infer­ ence evaluation metrics to the remaining aetive inference paths. Because MPHONE and INSECT are now known to be competing word-senses, Filter treats IPATH7 and IPATH2 as competing inference paths. That is, although IPATH7 connects MPHONE to C-1-A and IPATH2 con­ nects INSECT to C-1-A, the two different paths are evalu­ ated as if they connected the same two word-senses because INSECT and MPHONE were activated by the same lexical entry ("bugs"). Forthissame reason, IPATH5 is evaluated against IPATH6. This evaluation results in the two remain­ ing inference paths containing INSECT to be de-activated, so Filter resolves the ambiguity in favor of MPHONE. Be­ low is the active memory structure after all processing has ended, followed by the pointers into the structure.

Proceaaing completed

Active memory structure :

Path from NPHONE to C-I-A NPHONE is role-filler of REN-LISTEN-DEV REN-LISTEN-DEV is plan of P-SECRETS P-SECRETS is goal of SPY-AGENCY SPY-AGENCY has special case C-I-A Path from SEARCH to NPHONE SEARCH is precondition of RENOVE RENOVE has special case REN-LISTEN-DEV REN-LISTEN-DEV haa role-filler NPHONE

Path from C-I-A to SEARCH C-I-A is special caae of SPY-AGENCY SPY-AGENCY haa goal P-SECRETS P-SECRETS has plan REN-LISTEN-DEV REN-LISTEN-DEV ia special case of RENOVE RENOVE has precondition SEARCH

Pointers to memory structure:

Actor: C-I-A

Action: SEARCH

Object : NPHONE

3.6 An Observation on the Ordering of lnference Metries

While testing the ATLAST program, it became ap­ parent that the order of application of the pragmatic in­ ference metrics affected ATLAST's eventual interpretation of the input text. As mentioned earlier, ATLAST applies its specificity metric first, followed by the "length" metric, and then the "explains more" metric. For the example of Section 3.5, this ordering of the inference metrics results in the interpretation that the CIA was looking for hidden mi­ crophones. On the other hand, if the order of application of the two parsimony metrics is reversed, ATLAST arrives at a different, nonsensical interpretation.

Though this observation does not Iead us to any mean­ ingful conclusions at this time, it provides an example of how ATLAST can serve not only as a "proving ground" for theories, but also as a source of new and interesting ideas worthy of further investigation.

4.0 Open Questions and Future Work

The initial implementation of ATLAST raised a myr­ iad of implementation issues, many of which are yet to be resolved. More importantly, the implementation again raised some open questions which have been encountered by other researchers.

One question has to do with the timing of the three in­ ference processes running in parallel. We do not yet know how much work each of the three processes should do in a given cycle, though we have made arbitrary initial deci­ sions. For the Proposer in particular, there are issues which have been addressed by some of the previous models uti­ lizing spreading activation [Quillian, 1968; Fahlman, 1979; Anderson, 1983]: How far does activation spread? Does activation decay with time? Is there reinforeerneut when paths intersect? Though we do not have answers to the questions now, ATLAST is designed to allow us to change timing parameters easily, possibly enabling us to "tune" the model for cognitive accuracy as work proceeds.

Another question is concerned with the content of AT­ LAST's memory. Currently, ATLAST runs with a high­ level abstraction of episodic memory: the relationships be­ tween the MOPsare fairly weil defined, but the details of the episodes themselves are almost non-existent. Thus, in­ formation is stored in the links, not in the nodes. The eventual addition of lower-level detail to the episodes will require the application of yet unknown qualitative, as op­ posed to quantitative, inference metrics.

Additionally, there is the issue of memory organiza­ tion. Whenever researchers assume that specific concepts are organized in specific ways in human memory (i.e., "this MOP is connected tothat MOP by this relationship"), it is nothing more than an educated guess. Currently, AT­ LAST's metrics depend more on the specific organization of memory, rather than on the content of memory, for correct operation. If the memory bad been organized differently, so that there were a different number of links between certain MOPs, for example, ATLAST's interpretation of the input text would have been different. This particular realization of the metrics is not necessarily inaccurate, nor does the metrics' reliance on a particular organization of memory in­ validate ATLAST, any more than similar educated guesses invalidate any other models of human understanding. This issue does remind us, however, that our implementation de­ cisions can have as great an impact as our scientific theories on the perceived accuracy of our cognitive models, and that we should remain aware of where theoretical issues end and implementation issues begin.

Obviously, much work remains to be done on AT­ LAST. The current implementation has been applied only to short texts. In the future, we will process Ionger texts, and different types of texts, in order to discover additional rules for inference processing. The ATLAST model pro­ vides a framework for testing theories, as weil as for making predictions which can be verified experimentally.

5.0 Conclusion

5.1 Summary

To some extent, ATLAST is a unification and refine­ ment of ideas from previous models of human inference processes at the lexical, syntactic, and pragmatic Ievels. Yet, while ATLAST shares common features with each of these models, in many ways it is different from each of these same models. The features which distinguish the AT­ LAST model from others are discussed in greater detail in [Granger, Eiselt, &, Holbrook, 1985]. A brief summary of those features follows:

• ATLAST unifies inference processing at three distinct Ievels: the lexical, syntactic, and pragmatic Ievels.
• The Separation of intra-phrasal and inter-phrasal syn­ tactic analysis enables ATLAST to process texts which humans understand and to make the same mistakes a human understander makes.
• The use of a spreading-activation memory model al­ lows ATLAST to pursue competing inference paths simultaneously until syntactic or semantic informa­ tion suggests otherwise. Previous models of inference decision processes either left a loose end or chose a default inference when faced with an ambiguity [De­ Jong, 1979; Granger, 1980; Lebowitz, 1980; Granger, 1981; Dyer, 1982; Wilensky, 1983] .
• The concurrent operation of ATLAST's Capsulizer, Proposer, and Filter permits pragmatic interpreta­ tions to be evaluated independently of syntactic deci­ sions. This parallel organization also allows imme­ diate evaluation and inhibition of competing infer­ ence paths, thus minimizing combinatorial explosion effects.
• ATLAST conforms to the results of controlled exper­ iments on human subjects.

5.2 Final Comment

This paper describes how ATLAST attempts to un­ derstand only a five-word sentence. At first glance, this hardly seems like progress when one considers, for instance, that earlier systems understood hundreds of newspaper sto­ ries; in fact, it might even appear that work in naturallan­ guage understanding is going backwards, at least from a performance perspective. What is really indicated by this phenomenon, though, is that we are becoming more aware of the great quantity of knowledge and the complexity of the processes which language understanders, both human and otherwise, must bring to bear in understanding even the simplest text . In this light, we should not measure the validity of any model of understanding in terms of how many stories it understands, how many words are in its vocabulary, or how fast it runs. More appropriately, we should ask such questions as: Is the model extensible? Does it compare favorably with experimental data? ls it learn­ able? Does it make testable predictions? In other words, cognitive models should be evaluated on the robustness of the theory which they embody. Only when that metric is satisfied will the engineering issues become relevant. From this perspective, it is safe to say that ATLAST is a step in the right direction.

6.0 Acknowledgments

Special thanks to Rick Granger and Jen Holbrook, whose efforts have made this work possible.

7.0 References

Anderson, J .R. The Architecture of Cognition. Cambridge, MA: Harvard University Press, 1983.

Charniak, E. Passing markers: A theory of contextual in­ fluence in language comprehension. Cognitive Science, 7, 171-190, 1983.

Cullingford, R.E. Script Application: Computer Under­ standing of Newspaper Stories. Ph.D. thesis. Research Report #116. Department of Computer Science. Yale Uni­ versity, New Haven, CT, 1978.

DeJong, G.F. Skimming Stories in Real Time: An Experi­ ment in lntegrated Understanding. Ph.D. thesis. Research Report #158. Department of Computer Science. Yale Uni­ versity, New Haven, CT, 1979.

Dyer, M.G. ln-Depth Understanding: A Computer Model of Integrated Processing for Narrative Comprehension. Cam­ bridge, MA: MIT Press, 1983.

Fahlman, S.E. NETL: A System for Representing and Using Real- World Knowledge. Cambridge, MA: MIT Press, 1979.

Frazier, L., &, Fodor, J.D. The sausage machine: A new two-stage parsing model. Cognition, 6, 291-325, 1978.

Granger, R.H. When expectation fails: Towards a self­ correcting inference system. Proceedings of the First An­ nual National Conference on Artificial Jntelligence, Stan­ ford, CA, 1980.

Granger, R.H. Directing and re-directing inference pursuit: Extra-textual influences on text interpretation. Proceedings of the Seventh International Joint Conference on Artificial lntelligence, Vancouver, B.C., Canada, 1981.

Granger, R.H., Eiselt, K.P., &, Holbrook, J .K. STRATE­ GIST: A program that models strategy-driven and content­ driven inference behavior. Proceedings of the National Con­ ference on Artificial Intelligence, Washington, D.C., 1983.

Granger, R.H., Eiselt, K.P., & Holbrook, J.K. Parsing with parallelism: A spreading-activation model of inference pro­ cessing during text understanding. In J .L. Kolodner and C.K. Riesheck (eds.), Memory, Experience and Reasoning. Hillsdale, NJ: Erlbaum, 1985 (to appear) .

Granger, R.H., & Holbrook, J.K. Perseverers, recencies, and deferrers: New experimental evidence for multiple in­ ference strategies in understanding. Proceedings of the Fifth Annual Conference of the Cognitive Science Society, Rochester, NY, 1983.

Granger, R.H., Holbrook, J.K., & Eiselt, K.P. Interaction effects between word-level and text-level inferences: On­ line processing of ambiguous words in context . Proceedings of the Sixth Annual Conference of the Cognitive Science Society, Boulder, CO, 1984.

Kolodner, J .L. Retrieval and Organizational Strategies in Conceptual Memory: A Computer Model. Hillsdale, NJ: Erlbaum, 1984.

Lebowitz, M. Generalization and Memory in an Integrated Understanding System. Ph.D. thesis. Research Report #186. Department of Computer Science. Yale University, New Haven, CT, 1980.

Lucas, M. Lexical access during sentence comprehension: Frequency and context effects. Proceedings of the Fifth An­ nual Conference of the Cognitive Science Society, Rochester, NY, 1983.

Norvig, P. Six problems for story understanders. Proceed­ ings of the National Conference on Artificial Intelligence, Washington, D.C., 1983.

Quillian, M.R. Semantic memory. In M. Minsky (ed.), Semantic Information Processing. Cambridge, MA: MIT Press, 1968.

Riesbeck, C.K., & Martin, C.E. Direct Memory Access Parsing. Research Report #354. Department of Computer Science. Yale University, New Haven, CT, 1985.

Schank, R.C. Conceptual Information Processing. Amster­ dam: North-Holland, 1975.

Schank, R.C. Dynamic Memory: A Theory of Reminding and Learning in Computers and People. New York: Cam­ bridge University Press, 1982.

Seidenberg, M.S., Tanenhaus, M.K., Leiman, J .M., & Bi­ enkowski, M. Automatie access of the meanings of ambigu­ ous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14, 489-537, 1982.

Small, S., Cottrell, G., & Shastri, L. Toward connectionist parsing. Proceedings of the National Conference on Artifi­ cial Intelligence, Pittsburgh, PA, 1982.

Swinney, D.A., & Hakes, D.T. Effects of prior context upon lexical access during sentence comprehension. Journal of Verbal Learning and Verbal Behavior, 15, 681-689, 1976.

Tanenhaus, M., Leiman, J ., & Seidenberg, M. Evidence for multiple stages in processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18, 427-440, 1979.

Wilensky, R. Understanding Goai-Based Stories. Ph.D. thesis. Research Report #140. Department of Computer Science. Yale University, New Haven, CT, 1978.

Wilensky, R. Planning and Understanding. Reading, MA: Addison-Wesley, 1983.

Wilensky, R. Knowledge representation- A critique and a proposal. Proceedings of the First Annual Workshop on Theoretical Issues in Conceptual Information Processing, Atlanta, GA, 1984.