This note describes a logical system based on concepts and contexts, whose aim is to serve as a representation language for meanings of natural language sentences. The logic is a theoretical description of the output of an evolving implemented system, the system Bridge, which we are developing at parc, as part of the aquaint program. The note concentrates on the results of an experiment which changed the underlying ontology of the representation language from cyc to a version of WordNet/VerbNet.
For those interested, the system Bridge parses sentences in English using our
industrial-strength parser xle and a hand-crafted lexical functional (lfg)
grammar[
Given that our logical akr representations are intimately connected to the underlying ontology, one might expect that the change of ontology from cyc to WordNet/VerbNet would necessitate a total reworking of the system Bridge. This turned out not to be case, the re-architecture of the system was surprisingly easy and almost trouble-free. It is true that new trade-o s were made and these are some of the issues that we discuss here. But before discussing trade-o s we should explain why this change of ontology and what are the representations obtained in TIL2.
Our aim is to map free text to logical formulae based on a conceptual hierarchy that one can reason with. Our initial intuition was to try to use the concepts provided by the biggest knowledge base available cyc, and to take advantage of its reasoning component, which was familiar to some of us. Although we used cyc concepts for our rst logic, we found it useful to map the text to an abstract form of knowledge representation (akr), that could be realized as cyc or km or any other knowledge representation formalism. The design of this akr aimed for a sweet spot between ease of mapping from text to a formalism, and mapping from that formalism to standard logical representations. A happy surprise was our realization that the akr representations were already good enough for some important classes of textual inferences that we wanted to concentrate on. In general, the inferences we wanted to concentrate on were immediate, almost simple-minded, but necessary for the understanding of the text. For example, if the text says that \John managed to close the door" then we can safely infer that \John closed the door" and this kind of immediate inference is absolutely necessary to answer questions, based on snippets of text, as is the case in our primary application. Furthermore, these inferences did not seem to depend crucially on the particular ontology; they were much more dependent on the articulation of inference patterns surrounding the use of particular classes of words which appear quite often in open texts.
At the same time, we were having serious di culties completing mappings
from open texts when using the cyc system. cyc's mappings from word to cyc
concepts are very sparse, as might be expected from a knowledge base not built
to model language. We realized that having good information, very deep, about
some concepts and nothing at all about others was worse than having super cial
information about most words. Thus we decided to move to a WordNet/VerbNet
ontology, or more precisely, to the projection of WordNet/VerbNet obtained from
our own Uni ed Lexicon[
As in the previous version of the system the logic is based on the notion of
events expressed in a neo-Davidsonian style [
We rst describe the logical system, using several examples. Rather than listing all kinds of relations between contexts and concepts that the implemented system produces, we aim to give a feel for how our representations look like. Then we discuss the changes, gains and losses, caused by the change of ontology from cyc to WordNet/VerbNet. Finally we discuss methods and criteria for evaluating the coverage of the logical system obtained. We close with some ideas for further work. 2
TIL2 via examples It is traditional for logics of Knowledge Representation to be fragments of rstorder logic (fol). It is traditional for logics for natural language semantics to be higher-order intensional logics. Our logic has concepts, which make it look like \description logics", that is, fragments of fol, but it also has contexts, a possible-worlds-like construct that, we hope, is expressive enough for the needs of natural language.
Concepts, the way we conceive them, come from both neo-Davidsonian event
semantics and, somewhat independently, from description logics. Some of our
reasons for using a concept denoting analysis instead of an individual denoting
analysis when mapping noun phrases to logic are discussed in [
One of the main di erences between TIL and TIL2 consists in the type of concepts that are used. While in the previous logic TIL the basic ontology was the cyc ontology, for TIL2 the basic ontology is WordNet/VerbNet. But whatever the basic ontology, concepts in our logic are of two very di erent kinds: the rst kind of concepts are given a priori, sitting in a established hierarchy, based on the hierarchy underlying either cyc or the synsets of WordNet, considered as a taxonomy. The second kind of concepts are dynamic, created by the implemented system Bridge when we feed it an English sentence. The dynamic concepts are created and placed in the hierarchy in use, as best as we can, at run time.
For example, when using the cyc ontology, for the sentence A zebra slept, we use two cyc concepts Zebra and Sleeping and two dynamic concepts zebra : 1, a subconcept of the cyc concept Zebra and sleep : 11, a subconcept of the cyc concept Sleeping. Now when the same sentence is analyzed in the WordNet/VerbNet version of the system, the dynamic concept zebra : 1 will be mapped to a subconcept of the WordNet synset corresponding to the zebra animal, but the dynamic concept corresponding to the zebra's sleeping, sleep : 11 will be mapped to two di erent static concepts in WordNet, one corresponding to the WordNet meaning of animal sleep, the other corresponding to \accommodate", as in the sentence The tent sleeps six.
The concepts in WordNet are treated by Bridge, following WordNet convention, using the synset numbers. These are not very easy to read, hence the system pretty-prints it as a head word of the synset, followed by a number. The dynamic concepts are written as the word colon a number, showing that this is simply a Skolem constant. For example a clause like subconcept(sleep : 11; [sleep 1; sleep 2]) means that the dynamic subconcept of the zebra sleep (sleep : 11) is either a subconcept of sleep 1 or of sleep 2.
The most underspeci ed concept in the WordNet hierarchy is entity, which corresponds to the concept Thing in cyc. All our concepts are subconcepts of the most underspeci ed concept. We assume that that there are no circularities nor inconsistencies1 in the given initial hierarchy, be that cyc or WordNet.
The second main di erence between TIL and TIL2 has to do with how the
concepts are related, when expressing propositions. In both systems concepts are
related via \role" assertions, but the kinds of roles available are di erent. Thus
continuing on with the same example \The zebra slept" when using the cyc
ontology, we were able to use the cyc role bodilyDoer to connect the sleeping
event concept to the zebra concept, so the representation ends up with the
two subconcept clauses plus a clause for role(bodilyDoer; sleep : 11; zebra : 1),
while the representation using the WordNet/VerbNet ontology is very similar,
but has instead role(Agent; sleep : 11; zebra : 1) The cyc role bodilyDoer is
much more speci c than the role Agent from the much more limited collection of
VerbNet roles. Our uni ed lexicon ([
The logical system described so far looks like a description logic. We have concepts Concept with their own partial order (written as subconcept(t1; t2)) and roles Role, which are binary relations on the set of concepts Concept. We write clauses that either relate concepts via subconcept relations or relate roles to pairs of concepts, like role(Agent; sleep : 11; zebra : 1). And we write collections of clauses that correspond to representations of natural language sentences and hence correspond to propositions.
But our simple logic has contexts Context, as well as concepts. There is a rst initial context (written as t) that corresponds roughly to what we take the world to be like, as far the author of the sentence is committed to. But since this circumlocution is awkward, we will usually talk about this top level context as the `true context'.
Contexts in our logic were conceived as syntactic ways of dealing with intensional phenomena, including negation and non-existent entities. They support making existential statements about the existence and non-existence in speci ed possible worlds of entities that satisfy the intensional descriptions speci ed by our concepts. The possible worlds re ect the worlds implicitly (partially) described by the author of a text. Authors statements of propositional attitudes clearly require use of intensional terms, since no existence in the real world can be implied by such descriptions. It is clear that intensional notions are required when dealing with the representation in logic of propositional attitudes. We use propositional attitudes as an example of our use of contexts.
Propositional attitudes predicates relate contexts and concepts in our logic. Thus a concept like `knowing' or `believing' or `denying' introduces a context that represents the proposition that is known, believed or denied. For example, if we want to represent the sentence Ed denied that the diplomat arrived, we will need concepts for the arriving event, for the denying event, for the diplomat and for Ed. And we will need roles that describe how these concepts relate to each other. Thus we need to say who did the `denying' and `what was denied' and who did the arriving. The content of what was denied in the denying event is the proposition corresponding to The diplomat arrived. The role corresponding to `what was denied' relates a dynamic concept, the concept of the denying event (written as deny : 4), to (the contents of) a new context. To name this new context we use its `context head'. The context head is the arriving event, so the new context is called context(ctx(arrive : 4)) (`contex-head' is one of the many roles in the system that is not a thematic role).
Contexts allow us to localize reasoning: the existence of the denying event and of Ed are supposed to happen in the true world, but the existence of the arrival of the diplomat is only supposed to happen in the world of the things denied by Ed. In particular the arrival event could be considered as not happening, if Ed is known as a reliable source. (The system takes no position as to the instantiability or not of the arrival event in the top context: the instantiability of the arriving is only stated in the context of the things denied by Ed.) In some cases (for example if the sentence was Ed knew that the diplomat arrived) we can percolate up the truth of assertions in inner contexts up to the outside context. In many cases we cannot. The happening or not of events is dealt with by the instantiability/uninstantiability predicate that relates concepts and contexts.
While we may be prepared to make the simplifying assumption that if `X is known' than `X is true', we certainly do not want to make the assumption that if `X is said' than `X is true'. We say that the context introduced by a knowing event is veridical with respect to the initial context t, while the context introduced by a saying event is averidical with respect to the initial context. Negation introduces a context that is anti-veridical with respect to the original context. Thus we have a fairly general mechanism of contexts (these can clearly be iterated), which can represent some positive and some negative information. Similarly to McCarthy's logic we also have `context lifting rules' that allow us to transfer veridicality statements between contexts, in a recursive way.
Our representations also have a (preliminary) layer of temporal representation on them. The idea is to order events according to their times of happening and with respect to some generic time `Now'.
A few words on related work: Clearly our goals and motivations are very
similar to the SNePS project[
Changing the ontology allowed us to talk about ambiguity-enabled or packed representations. While we could, in principle, do the same with the cyc ontology and we did so, to a limited extent, in practice we simply didn't have the di erent concepts for each word. For many words we did not have a single concept associated to it, for very few we did have more than one. So we were restricted to what the ontologists in cyc thought the meaning of a given word was. (Of course we are now constrained to the meanings that the lexicographers at WordNet think one should have, but the pool is much bigger. So we do not have the problem of \missing concept for skolem", by and large). Thus a sentence like \Ed arrived at the bank" will not be assigned simply one of possible meanings of \bank" (river bank or nancial institution). Actually it will map to any of the ten possible meanings of bank in WordNet. Also \arrive" will be mapped to two di erent meanings, the physical reaching of a destination and the somewhat metaphoric, succeed in a big way. But instead of having twenty di erent representations for the meaning of the sentence, sharing the concepts 'Ed', 'arrive', 'bank' and the VerbNet roles for 'arrive', we have a single representation packing all of this as subconcept(arrive : 4; [arrive 1; arrive role(Experiencer; arrive : 4; Ed : 1) role(Cause; arrive : 4; bank : 15) subconcept(Ed : 1; [male 2]]) subconcept(bank : 15; [bank 1; : : : bank
One bad side of this is that we are forced (to begin with, at least) to use the very uninformative VerbNet roles. Thus in the example above we end up with one sensibly named role role(Experiencer; arrive : 4; Ed : 1) and one not so sensibly named role(Cause; arrive : 4; bank : 15). We have discussed ways of augmenting the number of roles of VerbNet (from less than twenty) to a reasonable number, presumably much less than the 400 that cyc has, but have found that a daunting task, so we are still exploring possibilities. Roles in our system are supposed to support inference and at the same time are supposed to make the mapping from language feasible. For the latter purpose (mapping from language feasibly) VerbNet roles are well-suited, but they are too underspeci ed to help with inference. The quest is on to nd a collection of roles that keeps feasibility of the mapping, but improves the inferential capabilities.
While the mechanism that implements the packing of representations could be used with the cyc ontology too, the actual details of the previous implementation, which looked at noun concepts before verb concepts (given cyc's more extensive coverage of nouns) made packed representations the exception rather than than the norm. In any case packing makes more sense when using WordNet/VerbNet where we do have many concepts for each word.
Another feature of our use of the new ontology is that it does not enforce \sortal restrictions". Using cyc we could make sure that in the sentence Ed red the boy the verb ` re' was used with the meaning of what cyc calls DischargeWithPrejudice, while in Ed red the cannon it was used with a ShootingAGun meaning. With the new ontology we do not weed out even the worst clashes of meanings. But a single representation covers a multitude of meanings. We take this as a shortcoming that we plan to address in future work. 4
The reason for introducing event concepts was the fact that they make some inferences that can be complicated in other semantical traditions very easy. For example it is obvious how to obtain Ed arrived in the city from the sentence Ed arrived in the city by bus. This inference corresponds simply to conjunction dropping in our logic. But of course there is much more to textual inference than simply dropping conjuncts.
To test textual inference our system provides a method for detecting entailment/contradictions, called \qa" for the application in question answering. When given two passages \qa" tells us whether the second passage is entailed by the rst one or not. Simple subconcept/superconcept reasoning is handled. In addition we support some pre and post condition reasoning. So Ed arrived in the city does entail that A person arrived in the city, since Ed is a person. Similarly Ed arrived in Rome should entail that Ed arrived in a city, as `Rome' is a city, but given that the proper names in WordNet are somewhat sketchy, we do not use this facility.
Note that the clauses we construct satisfy the usual monotonicity patterns, both in positive and in negative form. Thus Ed arrived in the city by bus entails that Ed arrived in the city. But Ed did not arrive in the city entails that Ed did not arrive in the city by bus, while Ed did not arrive in the city by bus does not entail that Ed did not arrive in the city.
We have also implemented the transformation of nominal deverbals with their
respective arguments into verb-argument structures. The work is described in
[
We have done signi cant work on exploring how certain linguistic expressions
support classes of context-lifting rules. Using the context structure of our logic
we support inferences associated with kinds of verbs with implicative behavior.
In our uni ed lexicon, the classes of such behavior are marked. This work is
discussed in [
More complicated are the verbs that only show their implicative behavior either in positive or negative situations. For example we have positive implicatives like the verb \force (to)" takes positive polarities and produces positive polarities (e.g \Ed forced Mary to paint" ! \Mary painted"), but if \Ed didn't force Mary to paint" we cannot tell whether Mary painted or not. While \refuse (to)" only works to produce negative polarity (e.g. `Mary refused to sing" ! \Mary did not sing"). There are also negative implicatives like \attempt (to)" and \hesitate (to)" which again only work for a negative polarity, but produce a positive one (\Ed didn't hesitate to leave" ! \Ed left", but if \Ed hesitated to leave" we cannot tell whether he left or not).
Finally we have factives and counterfactives, examples being \forget (that)" ("Ed forgot that Mary left" ! "Mary left" and \Ed didn't forget that Mary left" ! \Mary left" and \pretend that" (\Ed pretended that Mary left" ! \Mary didn't leave" and \Ed didn't pretend that Mary left" ! \Mary left"). And the neutral class, where we cannot say anything about the veridicity of the complement (\Ed said/expected that Mary left"). Further work is in progress to mark implicative behavior of verbs that do not take sentential complements.
One of our di cult design decisions was over the treatment of copula. It was clear that one needed to have trivial inferences like \Ed is a clown" contradicts \Ed is not a clown". But the mechanism used to infer that should be also able to cope with answering yes to \Ed, the clown, slept" implies that \A clown slept" and several other similar and not so similar inferences. 5
From the beginning we faced the problem of measuring the `quality' of our
representations. One can try to measure that by manually inspecting the
representations themselves and checking that the arguments provided by the system
correspond to our intuitions. But this is not very e cient nor objective. We can
also try to measure the faithfulness of the representations by checking whether
the system can answer correctly questions, using these representations. We have
tried this indirect method in the aquaint pilot kb Evaluation and the
measuring is quite di cult because question/answer pairs usually have to deal with
several logic-linguistic issues at once. We devised pairs of question/answers that
try to focus on a particular speci c problem at a time. Thus we have test-suites
checking mostly deverbal nouns, or anaphora resolution or coordination of
sentences, etc. But besides being time consuming and laborious, it is not clear that
this would measure adequacy or faithfulness of the representations in a fair way.
At the moment, it seems to us that the best that can be done is to try to look
at textual entailment, as originally proposed by the pascal rte but modify it
to deal with the issues that we consider important, like the implicative behavior
of lexical items and especially the need to distinguish between entailment of a
negation from not being able to draw a conclusion [
This note only starts the discussion of the kinds of inferences that we expect to
be able to make using our logic of concepts and contexts. On the positive side we
have an implemented system Bridge that it is easy to modify as it relies on a heavy
duty rewriting system (the transfer system[
On the negative side, much work remains to be done to get the system working as well as we want it to. First we still have a long way to go as far as improving the representations is concerned. Amongst the issues we have not discussed here are how to deal with noun-noun compounds, how to deal with contexts introduced by adjectives and adverbs and how to deal with temporal modi ers and temporal interpretation in general. We have done some work on these problems and hope to describe that work elsewhere.
We have said nothing about how to deal with lexical entailments such as Ed snored implies that Ed slept. We are not sure whether this problem should be addressed by creating enriched representations (maybe the concept of `snoring' must include a \concurrent" necessary condition of `sleeping') , or whether such inferences should be handled in the entailment/contradiction algorithm. So we are back to the previous trade-o between easiness of mapping and easiness of reasoning.