EPISODIC LOGIC: NATURAL LOGIC + REASONING

Karl Stratos, Lenhart K. Schubert and Jonathan Gordon

Department of Computer Science, University of Rochester, Rochester, 14627 New York, U.S.A.

Keywords:

Episodic logic, Natural logic, Implicativity, Presupposition, Entailment, Knowledge extraction.

Abstract:

There are two extreme stances in mechanizing natural language inference. One seeks to reformulate a raw

message so as to conform with the syntax and semantics of some formal logical system (such as FOL) suited

for reliable, potentially deep general reasoning. The other uses what has become known as Natural Logic—an

easy but shallow way of treating natural language itself as logic and reasoning directly on this level. Finding

the right balance between these opposing stances is one of the key tasks in advancing the ability of machines

to understand human language, and thus, for example, make inferences from text. In this paper, we provide

arguments and evidence that EPILOG, a general reasoner for the natural language–like Episodic Logic, can be

equipped with the knowledge needed for effective Natural Logic–like inference while also providing greater

generality.

1 INTRODUCTION

The beauty of Natural Logic (NLog) lies in its abil-

ity to make simple, intuitively natural inferences by

looking at the surface structure of a sentence and ex-

ploiting linguistic properties such as polarity, implica-

tivity, and factivity. Polarity refers to the fact that

certain linguistic environments are upward entailing

(positive), allowing truth-preserving substitution of

more general terms, while others are downward entail-

ing (negative), allowing substitution of more speciﬁc

terms. For example, a majority of predicates as well

as conjunction and disjunction pass the polarity of the

environment in which they occur to their operands,

while negation, conditional antecedents, and restric-

tors of universal quantiﬁers induce the opposite po-

larity in their operands. Implicativity (typically in-

volving verbs with inﬁnitive complements) and factiv-

ity (typically involving verbs with subordinate-clause

complements) interact with polarity but arise in inten-

sional contexts. For example, consider the following

news headlines

1. Vatican refused to engage with child sex abuse inquiry.

2. A homeless Irish man was forced to eat part of his ear.

3. Oprah is shocked that President Obama gets no re-

spect.

4. Meza Lopez confessed to dissolving 300 bodies in

acid.

From the Guardian, 11 Dec. 2010; The Hufﬁngton Post,

18 Feb. 2011; Fox News, 15 Feb. 2011; and Examiner.com,

22 Feb. 2011.

While such headlines may deliver messages at

multiple levels, including insinuated appraisals (e.g.,

Oprah is wrong), they certainly purport to provide

facts concerning the current state of the world. Thus,

a crucial part of understanding these headlines is mak-

ing the inferences that (1) The Vatican did not engage

with the child sex abuse inquiry, (2) An Irish man did

eat part of his ear, (3) President Obama gets no re-

spect, and (4) Meza Lopez dissolved 300 bodies in

acid.

These facts can be directly established by exploit-

ing the implication signatures a/b of the main verbs

in these headlines, where a,b ∈ {+, −,◦}. For exam-

ple, an implicative verb like ‘refuse (to)’ has an im-

plication signature −/+, indicating that in a positive

environment, ‘x refuse to y’ carries the negative im-

plication ‘not x y’, and in a negative environment it

carries the positive implication ‘x y’. Similarly a fac-

tive verb like ‘is shocked (that)’ has an implication

signature +/+, indicating that in both positive and

negative environments, ‘x is shocked that y’ implies

‘y’. The signatures of ‘be forced (to)’ and ‘confess

(to something)’ are both +/◦, indicating that these

verbs carry an implication only in positive environ-

ments. Note that the uniform signatures +/+ and

−/−, corresponding to factives and antifactives, in-

dicate presuppositional predicates. We also occasion-

ally use bracketing, e.g., +/(+), to indicate weak or

cancelable implications.

The shortcoming of this approach is that it ob-

tains little more than superﬁcial inferences. MacCart-

304

Stratos K., K. Schubert L. and Gordon J..

EPISODIC LOGIC: NATURAL LOGIC + REASONING.

DOI: 10.5220/0003690003040310

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (KEOD-2011), pages 304-310

ISBN: 978-989-8425-80-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

ney demonstrated that their NATLOG system, an en-

tailment veriﬁer based on NLog, makes surprisingly

accurate judgments on FraCaS test instances,

but it

can only verify the given entailment; one has to spec-

ify both the premise and the conclusion (MacCartney

and Manning, 2008). Moreover,inferences are limited

to single premise sentences and have to result from

“aligning” the premise with the hypothesis and then

judging whether a sequence of “edits” (substitutions,

insertions, deletions) leading from the premise to the

hypothesis makes it likely that the premise entails the

hypothesis.Hence NATLOG can verify the correctness

of the entailment

Jimmy Dean refused to move without his jeans

James Dean didn’t dance without pants

but it would not be able, for example, to use a second

premise, ‘Jimmy Dean could not ﬁnd his jeans’ to con-

clude that ‘Jimmy Dean did not dance’. (Assume that

not being able to do something entails not doing it,

and not ﬁnding something entails not having it.)

We show that Episodic Logic (EL), a very nat-

ural representation of human language, has the po-

tential to overcome the inherent shallowness of the

NLog scheme. To demonstrate this potential, we sup-

ply EL axioms, meta-axioms, and inference rules to

EPILOG, a general EL reasoner that has been shown to

hold its own in scalable ﬁrst-order reasoning against

the best current FOL theorem provers, even though

its natural language–like expressive devices go well

beyond FOL. It has been used to solve problems

in self-aware and commonsense reasoning and some

challenge problems in theorem proving (Morbini and

Schubert, 2007; Morbini and Schubert, 2008; Schu-

bert and Hwang, 2000). Once a sentence is in EL

form, we only need a KB that contains axioms and

inference rules specifying what conclusions can be

drawn from predicates with particular signatures. The

result is a reasoning system that can not only handle

the dual-premise example above but can also perform

general logical reasoning not directly related to nat-

ural language. We point out the beneﬁts of our ap-

proach over ones based only on NLog or FOL—and

also provide an evaluation on 108 sentences randomly

sampled from the Brown corpus—in Section 4.

2 PREVIOUS WORK

In the linguistics community,a tremendous amount of

effort has been invested in the study of presupposition,

implicativity, and polarity. We do not intend to cover

all the subtleties involved in this ﬁeld of study, but we

See MacCartney’s site http://www-nlp.stanford.edu/

∼wcmac/downloads/

Table 1: The typical behavior of E, P, and I.

E P I

Project from embeddings no yes no

Cancelable when embedded – yes –

Cancelable when unembedded no no yes

give a brief discussion of the aspects directly relevant

to our work.

The Strawsonian deﬁnition of presupposition (rel-

evant to factives and antifactives) is

One sentence presupposes another iff when-

ever the ﬁrst is true or false, the second is true.

This provides a nice logical characterization that cov-

ers the case of lexically “triggered” presuppositions—

in particular, the polarity-independentexistence of the

presupposed content (Strawson, 1952). As we will see

in Section 3, this rules out an axiomatic approach to

presupposition inference.

Other important aspects of implicativity and pre-

supposition are cancelability and projection. The im-

plications of an implicative such as ‘refuse’ can be

canceled in a negative context (‘John didn’t refuse to

ﬁght, but simply had no occasion to ﬁght’), and do

not survive an embedding (‘John probably refused to

ﬁght’). In contrast, a presupposition typically cannot

be canceled (#‘John doesn’t know that he snores, and

in fact he doesn’t’), and typically projects when em-

bedded (‘John probably knows that he snores’), but

not in all cases (‘I said to Mary that John knows that

he snores’). The typical behavior of entailments (E),

presuppositions (P), and implicatures (I) are summa-

rized in Table 1 (Beaver and Geurts, 2011). A notable

attempt to regulate presupposition projection is the

classiﬁcation of embedding constructions into plugs,

ﬁlters, and holes (Karttunen, 1973). Plugs (e.g., ‘say’

above) block all projections, ﬁlters (e.g., ‘if–then’)

allow only certain ones, and holes (e.g., ‘probably’

above) allow all.

There have also been many efforts to computation-

ally process these linguistic phenomena. They tend to

focus on handling monotonicity properties of quanti-

ﬁers and other argument-taking lexical items, which

ultimately determine the polarity of arbitrarily embed-

ded constituents. For instance, (Nairn et al., 2006) pro-

posed a polarity propagationalgorithm that accommo-

dates entailment and contradiction in linguistically-

based representations. MacCartney and Manning’s

NATLOG and its success on FraCaS examples showed

the potential effectiveness of a NLog-based system

that leverages these linguistic properties (MacCart-

ney and Manning, 2008). (Clausen and Manning,

2009) further showed how to project presupposi-

tions in NLog in accord with the plug–hole–ﬁlter

EPISODIC LOGIC: NATURAL LOGIC + REASONING

305

scheme. (Danescu-Niculescu-Mizil et al., 2009) ex-

ploited Ladusaw’s hypothesis—that negative polarity

items only appear within the scope of downward-

entailing operators—for unsupervised discovery of

downward-entailing operators (DEOs): lexical items

with negative polarity in their argument scope.

The main focus of this paper is not on handling all

the linguistic subtleties examined in the literature (in

particular, the projection problem of presuppositions).

Rather, it is to show how NLog-like reasoning based

on implicatives, factives and attitudinal verbs can be

incorporated into a formal reasoner, to come to grips

with some interesting problems that arise in the pro-

cess, and to argue that our approach ultimately enjoys

advantages over other approaches to inference in lan-

guage understanding.

EPILOG’s capability in NLog-like entailment in-

ference has already been partially demonstrated by

(Schubert et al., 2010). EPILOG’s inference mecha-

nism is polarity-centered, in the sense that much of

its reasoning consists of substituting consequences

of subformulas in positive environments and anti-

consequences in negative environments. In that re-

spect it rather closely matches NLog inference. For

instance, having inferred that Jimmy did not move

from ‘Jimmy refused to move’, it easily makes the

further inference that Jimmy did not dance (knowing

that dancing entails moving). But our focus here is not

on these natural entailment inferences, but on build-

ing a lexical knowledge base that will permit us to

obtain NLog-like inferences on a wide variety of text

examples involving implicatives, factives, and attitu-

dinal verbs.

3 METHOD

We have manually constructed a list of around 250

implicatives, factives, and attitudinal verbs with their

semantics. About half of the items come from (Nairn

et al., 2006) via personal correspondence with Cleo

Condoravdi at PARC. We have further expanded them

by considering their synonyms and antonyms, as well

as entirely novel items. The attitudinal verbs were

separately collected, with the goal of enabling infer-

ences of beliefs and desires. For example, if John

thinks that Bin Laden is alive, then we may reason-

ably infer that John believes that Bin Laden is prob-

ably alive; if Mary struggles to get an A, then Mary

surely wants to get an A; etc. We have also collected a

list of around 80 DEOs such as ‘doubt (that)’, which

preserve truth under specialization of the complement.

Around 60 of them came from those obtained by

(Danescu-Niculescu-Mizil et al., 2009).

We can encode lexical items into a semantic

Table 2: Some simpliﬁed axiom templates.

x dare to p ⇒ x p

x not dare to p ⇒ x not p

x decline to p ⇒ x not p

x not decline to p ⇒ probably x p

x is delighted that w ⇒ w

x is not delighted that w ⇒ w

x doubts that w ⇒ x believes probably not w

database for EPILOG by declaring the types of the

predicates and stating axioms or inference rules. In

this seemingly straightforward process, we encounter

both implementation issues and interesting linguistic

issues.

3.1 Axiomatizing Implicatives

EPILOG allows expression of very general axiom

schemas through syntactic quantiﬁcation (e.g., the

quantiﬁer ‘all pred’) and quotation (transparent to

syntactic metavariables). Thus, we could formalize

the implications of verbs like ‘manage’ or ‘dare’ in

a positive environment as follows:

(all pred p (’p imp+p)

(all pred q

(all x ((x p (ka q)) ⇒ (x q)))))).

This says that if a predicate p (e.g., ‘dare’) has pos-

itive implicativity in a positive environment (denoted

(’p imp+p)

, then whenever a subject x stands in

relation p to a kind of action (ka q) (e.g., ‘to dance’;

the ‘ka’ operator reiﬁes an action or attribute predicate

into a kind of action or attribute), then x does the ac-

tion q. If we now add the axiom

(s ’(’dare imp+p))

we will in principle enable the desired positive infer-

ence for ‘dare’.

This approach may be elegant, but it suffers from

O(n

) runtime with respect to proofs of length n

for a KB of size k in the current implementation of

EPILOG, since it may retrieve and attempt to match

numerous formulas containing matchable variables

and metavariables at every step in backward chaining.

(Inferential retrieval is geared toward completeness

rather than efﬁciency). A solution is to expand gen-

eral schemas like the above into verb-speciﬁc ones,

like the following for dare:

(all pred p (all x ((x dare (ka p)) ⇒ (x p))))),

(all pred p (all x ((not (x dare (ka p))) ⇒ (not (x p)))))).

A partial list of informal English templates for such

logical axioms is shown in Table 2.

3.2 The Presupposition Problem

As noted in Section 2, presuppositional inferences

can be made without regard to the polarity of the an-

tecedent. Now suppose that we try to capture this be-

havior for a presuppositional verb like ‘know’, via

KEOD 2011 - International Conference on Knowledge Engineering and Ontology Development

306

meta-axioms stating that both knowing w and not

knowing w entail w:

(all wff w (all x ((x know (that w)) ⇒ w)))),

(all wff w (all x ((not (x know (that w))) ⇒ w)))).

But this is absurd, because if both the truth and the fal-

sity of a premise lead to the conclusion that w holds,

then w simply holds unconditionally, and we will be

able to derive it even if no speciﬁc “knowing that”

premises are available. Similar comments apply in the

case of antifactives such as “pretending that” (which

in combination with axioms for factives makes EPI-

LOG conclude that any claim is both true and false).

What this indicates is that we need to carefully dis-

tinguish the assertion of a proposition in a given con-

text from its truth. It is the assertion of a “knowing

that” proposition or its negationin a context,that justi-

ﬁes adding the object of “knowing that” to the context.

The truth or falsity of a “knowing that” proposition—

one of which always obtains for any proposition in a

bivalent semantics—is no basis for inferring its pre-

suppositions.

In Natural Logic, this particular issue does not

arise, because conclusions are always based on explic-

itly available sentences, not on general logical con-

siderations. (For example, we cannot derive ‘John is

alive or he is not alive’ from an empty KB in NLog.)

But in EL, we need to avoid the above pitfall. We do

so here in a way that is adequate for top-level occur-

rences of (anti)factives or their negations by formu-

lating implicative rules as inference rules rather than

axioms, where the premises must be explicitly present

for the conclusion to be drawn. (Note that the above

issue is analogous to the fact that in logics of neces-

sity (Hughes and Cresswell, 1996), the necessitation

rule p/ p, with the premise p restricted to being a

theorem of the logic, cannot be recast as an axiom

p ⇒ p, as this would trivialize the logic, rendering

all true formulas necessarily true.) Fabrizio Morbini,

the designer of the current EPILOG, implemented a fa-

cility with which one can easily create such inference

rules. In particular the rule for know can be written

with the function

store-prs-ir

, which takes a list

of arguments, the premise, and the conclusion to gen-

erate an inference rule at compilation

(store-prs-ir ’(((w wff) (x)) (x know (that w)) w)),

(store-prs-ir ’(((w wff) (x)) (not (x know (that w))) w)).

These rules are insufﬁcient for arbitrarily embedded

occurrences of ‘know’, such as ‘John probably does

not know that he snores’. What we need more gener-

ally is a projection mechanism; this could in principle

be expressed with a meta-axiom concerning embed-

ded occurrences of ‘know’ (etc.):

(((w wff v wff x term))

(’(x know (that v)) projectibly-embedded-in ’w) v)

, where

projectibly-embedded-in

is procedurally decidable.

One fortuitous side effect is that the use of store-

prs-ir leads to faster inference than would be obtained

with axioms with similar content, because it reduces

the amount of work by blocking one direction of rea-

soning.

4 SOME RHETORIC AND SOME

RESULTS

Having created a lexical knowledge base as described

above, we can perform the top-level inferences al-

lowed by our implicatives, factives, and attitudinal

verbs. In particular, we can go back to the open-

ing examples in this paper. Given the following EL-

approximations to the news headlines for use in EPI-

LOG (where we have ignored the role of episodes,

among some other details),

(s’(Vatican refuse

(ka (engage-with Child-sex-abuse-inquiry)))),

(s’(some x (x (attr homeless (attr Irish man)))

(x (pasv force)

(ka (l y (some r (r ear-of y)

(some s (s part-of r) (y eat s)))))))),

(s’(Oprah (pasv shock)

(that (not (Obama get (k respect)))))),

(s’(Meza-Lopez confess

(ka (l x (some y (y ((num300) (plur body)))

(x dissolve y))))).

EPILOG returns the correct answers to each of the fol-

lowing queries in a small fraction of a second:

(Vatican engage-with Child-sex-abuse-inquiry), [NO]

(some x (x (attr homeless (attr Irish man1)))

(some r (r ear-of y) (some s (s part-of r) (x eat s)))) [YES]

(Obama get (k respect)), [NO]

(some y (y ((num 300) (plur body)))

(Meza-Lopez dissolve y)) [YES]

Note the conformity of these LFs with surface se-

mantic structure. They are also close to the outputs

of an existing parser/interpreter—when it works cor-

rectly, which is not very often, mostly because of

parser errors and the lack of a coreference module.

The greatest shortcoming of the current work remains

that we cannot yet fully automate the conversion of

natural language into EL. Does this defeat the whole

purpose of our approach—easy and effective infer-

ences on the lexical level, within a more general in-

ference framework? We argue that it does not by high-

lighting the advantages of our approach over purely

FOL- or NLog-based reasoners.

4.1 Advantages Vis-

a-Vis FOL

The weaknesses of FOL as a representationfor natural

language are well-known. In particular intensionality

EPISODIC LOGIC: NATURAL LOGIC + REASONING

307

(including, but not limited to, attitudes), generalized

quantiﬁcation (‘most people who own cars’), modiﬁ-

cation (‘unusually talented’), and reiﬁcation (‘his ab-

sentmindedness’, ‘the fact that he snores’) can at best

be handled with complex circumlocutions. It is of-

ten claimed that a more expressive logic suffers from

higher computational complexity. But this is false, in

the sense that any inference that is straightforward in

FOL is just as straightforward in a superset of FOL

(as was shown in the EPILOG references cited earlier).

In fact, a richer, language-like representation can fa-

cilitate many inferences that are straightforwardly ex-

pressible in words, but circuitous in a more restrictive

representation.

Another common misunderstanding is that any

logical representation demands absolute precision and

disambiguation to be usable. However, it should be

emphasized that we can be just as tolerant of impre-

cision and ambiguity in EL as in NLog (although in

both cases there are limits to how much can be tol-

erated without adverse effects; when told that John

had gerbils as a child, we probably do not wish

to conclude that he ate, or gave birth to, small ro-

dents). The language-like syntax and tolerance of im-

precision of EL allow us to easily handle modality

and vague, generalized quantiﬁers. At the same time,

it supplies a solid framework for accumulation of

context-independent, modular knowledge, which can

then be used for both superﬁcial and deep reasoning.

4.2 Advantages Vis-

a-Vis NLog

4.2.1 Multiple Premises

Because EPILOG is a logical system that stores its

knowledge in a KB available throughout its lifespan,

it can trivially handle inferences requiring multiple

premises. Consider the following contrived, but illus-

trative inference example. From the sentence ‘John is

surprised that Mary declines to contribute to charity’,

we wish to be able to derive that ‘Mary is probably

not very altruistic’ based on the world knowledge ‘If

someone declines to donate to charity, that person is

probablynot very altruistic.’ Given the premises in the

EL-approximations,

(John surprised

(that (Mary decline (ka (contribute-to (k charity)))))),

(all x ((x decline (ka (donate-to (k charity)))) ⇒

(probably (not (x (very altruistic))))))),

and also the knowledge

(all x (all y ((x donate-to y) ⇒ (x contribute-to y)))),

EPILOG correctly answers the query:

(probably (not (Mary (very altruistic)))) [YES].

This kind of reasoning, requiring both the NLog-style

superﬁcial inference and multiple-premisederivations

We know that we have hydrogen in water.

→ We have hydrogen in water.

The second of the above pair of sentences is a reasonably

clear and plausible conclusion from the ﬁrst sentence.

1. I agree

2. I lean towards agreement

3. I’m not sure

4. I learn toward disagreement

5. I disagree

Figure 1: The survey on the Brown corpus inferences.

Table 3: The frequency of the ratings. Lower numbers are

better; see Figure 1.

Rating Frequency Count Percent

1 502 75%

2 114 17%

3 31 5%

4 14 2%

5 3 0%

Table 4: The frequency of words in the sampling.

Word Count Word Count

think 25 suppose 4

know 15 appear 3

say 9 show 3

guess 7 tend 3

try 4 20 others 27

beyond the scope of NLog veriﬁers, is at the heart of

commonsense reasoning in our daily lives.

4.2.2 “Anywhere, Anytime”

After temporal deindexing and author/addressee dein-

dexing, EL formulas are usable for inference “any-

where, anytime”, whereas English sentences are not.

For instance, the deindexed form of John’s assertion

‘Yesterday I managed to propose to Mary’ would be

that ‘John asserted at about 1pm June 14/11 that John

managed to propose to Mary on June 13/11’. This fact

could be used in any context, at any time, e.g., to make

the implicativity-based inference that ‘John conversa-

tionally implied at about 1pm June 14/11 that John

proposed to Mary on June 13/11’. By contrast, the

English sentence is false from virtually anyone’s per-

spective but John’s (because of the use of ‘I’), and

even for John will become false by June 15/11 (be-

cause John didn’t propose to Mary ‘yesterday’ rela-

tive to June 15); likewise a conclusion like ‘I conver-

sationally imply that I proposed to Mary yesterday’

becomes false, even from John’s perspective, very

shortly after John’s utterance (because he has moved

on to saying and implying other things).

KEOD 2011 - International Conference on Knowledge Engineering and Ontology Development

308

Table 5: Average score of each judge on the inferences.

“Corr.” is the average pairwise Pearson correlation.

Judge 1 Judge 2 Judge 3 Judge 4 Judge 5 Corr.

1.33 1.21 1.56 1.40 1.23 0.13

Automatic deindexing of tense and temporal ad-

verbials is quite well-understood in EL (Schubert

and Hwang, 2000), and tense deindexing (as well

as quantiﬁer scoping) are performed in the exist-

ing parser/interpreter. Speaker/addressee deindexing

is also handled in a limited way. However, adverbial

deindexing remains unimplemented, and in any case

improving statistical parser performance and imple-

menting coreference resolution are more urgent needs.

Despite this incompleteness in the implementation

work, it is clear that systematic deindexing is feasi-

ble, and that only deindexed formulas (or else ones

permanently tagged with their utterance contexts) are

usable “anywhere, anytime”.

4.3 Evaluation on the Brown Corpus

We have randomly sampled 108 sentences from the

Brown corpus containing the relevant implicative,

presuppositional, and attitude predicates in our KB,

and run forward inferences on their EL approxima-

tions. The NL-to-EL conversion was done by manu-

ally correcting the ﬂawed outputs from the current EL

interpreter. EPILOG produces 133 distinct premise-

conclusion pairs when the approximated EL formu-

las are loaded. The EL-to-NL (verbalization) direc-

tion is completely automated. To evaluate the plau-

sibility/usefulness of the inferences, ﬁve people (stu-

dents and researchers at two sites) judged their quality

on a 1–5 scale; the survey question is shown in Figure

As seen in Tables 3 and 5, the ratings are very

high overall, afﬁrming the robustness of inferences

rooted in the well-studied linguistic properties we

made use of. The highest-rated inferences tend to be

those where the premise and conclusion are contentful

and easily understood, and of course the conclusion

is viewed as obvious from the premise; e.g., ‘The sol-

diers struggle to keep open a road to the future in their

hearts’ ⇒ ‘The soldiers want to keep open a road to

the future in their hearts’ (mean: 1, median: 1). The

lowest rated inferences are either trivial or too vague

to be useful, e.g., ‘The little problems help me to do

so’ ⇒ ‘I do so’ (mean: 2.75, median: 2.5). The low

correlation among judges can be attributed to differ-

ing interpretations as to how seriously sentence con-

tent and quality should be taken. But this is a minor

concern, given the generally high scores.

Some of the chained-forward inferences illustrate

the need to attend to the projection problem. For in-

stance, the inference ‘They refuse to mention that

they’re not there’ ⇒ ‘They don’t mention that they’re

not there’ is obtained by the negative implication of

‘refuse’. EPILOG then infers from this conclusion that

‘They’re not there’ by the presuppositional nature of

‘mention’. However, it is dubious if this latter infer-

ence projects from the initial embedding.

It is also interesting to note the frequency of the

words in the sampled sentences (Table 4); a vast ma-

jority are attitude verbs like ‘think’, illustrating our

tendency to express personal opinion—and thereby

the importance of extracting information from them.

5 CONCLUSIONS

We have taken a step toward combining “shallow” and

“deep” linguistic inference methodologies by equip-

ping a general reasoner with NLog-like inference ca-

pabilities. In addition to laying out some important im-

plementation issues and addressing relevant linguis-

tic phenomena, we have argued that our approach has

speciﬁc advantagesover ones based on less expressive

logics or on shallow, indexical NLog reasoning alone.

Though the work is far from complete (with regard

to automatic processing of NL sentences, efﬁcient in-

ference, and the handling of the projection problem),

our evaluation on the Brown corpus indicates that this

is a promising direction for further advancing lan-

guage understanding and, thereby, the acquisition of

inference-capable knowledge from language.

ACKNOWLEDGEMENTS

We thank Fabrizio Morbini for accommodating the

EPILOG facility required for the work in this paper.

This work was supported by NSF grants IIS-1016735

IIS-0916599, and ONR STTR subcontract N00014-

10-M-0297.

REFERENCES

Beaver, D. I. and Geurts, B. (2011). Presupposition. In

Zalta, E. N., editor, The Stanford Encyclopedia of Phi-

losophy. Summer 2011 edition.

Clausen, D. and Manning, C. (2009). Presupposed content

and entailments in natural language inference. In Proc.

of the ACL-IJCNLP Workshop on Applied Textual In-

ference.

Danescu-Niculescu-Mizil, C., Lee, L., and Ducott, R.

(2009). Without a ‘doubt’? unsupervised discovery

EPISODIC LOGIC: NATURAL LOGIC + REASONING

309

of downward-entailing operators. In Proc. of NAACL

HLT.

Hughes, G. E. and Cresswell, M. J. (1996). A New Introduc-

tion to Modal Logic. Routledge.

Karttunen, L. (1973). Presuppositions of compound sen-

tences. In Linguistic Inquiry, volume 4, pages 167–

193.

MacCartney, B. and Manning, C. D. (2008). Modeling se-

mantic containment and exclusion in natural language

inference. In Proc. of the 22nd International Confer-

ence on Computational Linguistics (COLING ’08).

Morbini, F. and Schubert, L. K. (2007). Towards real-

istic autocognitive inference. In Proc. of the AAAI

Spring Symp. on Logical Formalizations of Common-

sense Reasoning.

Morbini, F. and Schubert, L. K. (2008). Metareasoning

as an integral part of commonsense and autocognitive

reasoning. In Proc. of the AAAI Workshop on Metar-

easoning.

Nairn, R., Condoravdi, C., and Karttunen, L. (2006). Com-

puting relative polarity for textual inference. In Proc.

of Inference in Computational Semantics (ICoS-5).

Schubert, L. K. and Hwang, C. (2000). Episodic Logic

meets Little Red Riding Hood: A comprehensive, nat-

ural representation for language understanding. In

Iwanska, L. and Shapiro, S., editors, Natural Lan-

guage Processing and Knowledge Representation:

Language for Knowledge and Knowledge for Lan-

guage.

Schubert, L. K., Van Durme, B., and Bazrafshan, M. (2010).

Entailment inference in a natural logic–like general

reasoner. In Proc. of the AAAI 2010 Symp. on Com-

monsense Knowledge.

Strawson, P. F. (1952). Introduction to Logical Theory.

Methuen.

KEOD 2011 - International Conference on Knowledge Engineering and Ontology Development

310