On the Relationship between a Computational Natural Logic and

Natural Language

Troels Andreasen

, Henrik Bulskov

, Jørgen Fischer Nilsson

and Per Anker Jensen

Computer Science, Roskilde University, Roskilde, Denmark

Mathematics and Computer Science, Technical University of Denmark, Lyngby, Denmark

International Business Communication, Copenhagen Business School, Frederiksberg, Denmark

Keywords:

Computational Semantics of Natural Languages, Models of Computation for Natural Language Processing,

Bio-Information and Natural Language.

Abstract:

This paper makes a case for adopting appropriate forms of natural logic as target language for computa-

tional reasoning with descriptive natural language. Natural logics are stylized fragments of natural language

where reasoning can be conducted directly by natural reasoning rules reﬂecting intuitive reasoning in natural

language. The approach taken in this paper is to extend natural logic stepwise with a view to covering succes-

sively larger parts of natural language. We envisage applications for computational querying and reasoning, in

particular within the life-sciences.

For better or for worse, most of the reasoning that is done in the

world is done in natural language.

G. Lakoff: Linguistics and Natural Logic, Synthese, 22, 1970.

1 INTRODUCTION

Traditionally, computational reasoning with informa-

tion given in natural language is carried out by con-

ducting a translation of sentences into ﬁrst order

predicate logic, see e.g. (Fuchs et al., 2008; Kuhn

et al., 2006), or derivatives thereof such as a descrip-

tion logic dialect, as in (de Azevedo et al., 2014;

Thorne et al., 2014). There are also natural logic

approaches which extend syllogistic proof systems

(Pratt-Hartmann and Moss, 2009) or which call on

forms of logical type theory, thereby taking advan-

tage of an assumed compositional semantics for natu-

ral language drawing on higher order denotations (Fy-

odorov et al., 2003).

The project described here relies on appropri-

ate forms of natural logic decomposed into graph-

structured knowledge bases. Natural logics are tiny,

stylized fragments of natural language in which the

deductive logical reasoning can be carried out directly

by simple, intuitive rules, that is, without taking resort

to predicate-logical reasoning systems such as reso-

lution. Natural logic originates from the traditional

Aristotelian categorial syllogistic logic (van Benthem,

1986; Kl

ıma, 2010; Nilsson, 2015), which became

further developed and reﬁned in late medieval times.

However, in the course of the late 19th century devel-

opment, forms of logic coming close to natural lan-

guage were largely deemed obsolete by Frege’s intro-

duction of the more general, mathematically inclined

quantiﬁer-based predicate logic. As is well-known,

the latter subsequently prevailed throughout the 20th

century.

This paper pursues the idea of choosing natural

logic as a target language for dealing computation-

ally with appropriately constrained and regimented,

yet rich, forms of afﬁrmative sentences in natural lan-

guage. The purported methodological advantage of

the present approach lies in the proximity of natu-

ral logic to natural language, very much in contrast

to predicate logic. Indeed then, the translation of

the considered natural language fragments becomes

a partial recasting of the considered sentences into an

even smaller language fragment, namely natural logic

sentences. Obviously, the chosen natural logic then

determines and conﬁnes the semantic range for partial

coverage of the considered sentences. A related use of

natural logic principles for partial computational un-

derstanding of natural language is found in (MacCart-

ney and Manning, 2009). The language translational

relationships are planned to be initially implemented

using the well-known, rather unsophisticated deﬁnite

Andreasen, T., Bulskov, H., Nilsson, J. and Jensen, P.

On the Relationship between a Computational Natural Logic and Natural Language.

DOI: 10.5220/0005848103350342

In Proceedings of the 8th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2016) - Volume 1, pages 335-342

ISBN: 978-989-758-172-4

335

clauses (clause grammars).

The paper is structured as follows: In section 2

we describe the semantic basis of the considered di-

alect of natural logic presented in section 3. In section

4 we discuss various mainly conservative extensions

of the natural logic in order to accommodate expres-

sion forms common in natural language. Section 5

discusses various sentence cases and the problems in-

volved in approaching free natural language formula-

tions. Finally, in section 6 a summary concludes the

paper.

2 SEMANTIC MOTIVATION

Our semantic framework comprises a selection of

stated classes of entities together with binary relation-

ships between the classes akin to the popular entity-

relationship models. First of all, there is a funda-

mental isa subclass relationship known from formal

ontologies. In addition, class-class relationships may

be be introduced according to needs, as discussed in

(Smith et al., 2005; Schulz and Hahn, 2004; Bittner

and Donnelly, 2007; Yu, 2006).

It is a key feature of our approach that the given

named classes may be used to form subclasses ad li-

bitum by restriction with relationships to other classes

in the natural logic. As an example, given the classes

cell and hormone and the relation(transitive verb) pro-

duce, one may form the subclass

cell that produce hormone

This is a phrase in the applied natural logic forming a

new class which is a subclass of cell.

Unlike what is the case in predicate logic, in our

framework the entities belonging to the classes are not

dealt with explicitly. Individual entities may, how-

ever, if necessary, be dealt with as stipulated singleton

classes (having no subclasses in so far as the empty

class is left out in our setup).

This basis, while being application-neutral at the

outset, appears to be particularly useful for applica-

tions within the bio-sciences as discussed in (Smith

et al., 2005; Schulz and Hahn, 2004) as well as in

our (Andreasen et al., 2014a; Andreasen et al., 2014b;

Andreasen and Nilsson, 2014; Nilsson, 2015; An-

dreasen et al., 2015). General, natural language de-

scriptions in natural sciences abound with classes, let

us just mention the Linnean and chemical and medical

taxonomies. The class-relationship framework might

also ﬁnd more innovative use for ontology-structured

knowledge base concept organization and speciﬁca-

tion in semi-exact sciences, e.g. in linguistics.

3 CORE NATURAL LOGIC

Having introduced the semantic motivation, we turn

next to the logical sentences serving the mentioned

class-relationship setup. We consider a natural logic

which has the general form of expression shown in

(1):

Cterm

R Q

Cterm

(1)

where

• Q

are quantiﬁers (determiners) every/all, some/a,

• the grammatical subject term Cterm

and the

grammatical object term Cterm

are class expres-

sions, and

• R is a relation name.

In linguistic parlance, the Cterms are noun

phrases, and R is a transitive verb. In simpler cases

Cterms are just class names (nouns) C. Among the

quantiﬁer options here, we focus on the quantiﬁer

structure

every Cterm

R some Cterm

(2)

This form is pivotal in our treatment because it repre-

sents the default interpretation of sentences like be-

tacells produce insulin in ordinary descriptive lan-

guage. Further, it conforms with the functioning of

the class restrictions as to be described. Example:

every betacell produce some insulin. An explication

of this form in predicate logic with quantiﬁer struc-

ture ∀x∃y is straightforward, cf. our references above,

and therefore not repeated here. The inherent linguis-

tic structural ambiguity corresponding to the scope

choice ∀x∃y versus ∃y∀x is overcome by stipulating

the ∀x∃y reading, which is the useful one in practice.

See also (2) Accordingly, we have

∀x(betacell(x) → ∃y(produce(x,y) ∧ insulin(y))

Observe that we are not going to translate the sen-

tences into their predicate logical form with individ-

ual variables. Rather, it is a crucial feature of our ap-

proach that we decompose the sentences into simpler

constituents forming a graph without variables as ex-

plained in section 3.4.

Our natural logic notations rely on the convention

that if no quantiﬁers are mentioned explicitly, the in-

terpretation follows the scope pattern ∀x∃y. Thus,

the sentence betacell produce insulin is semantically

equivalent to our sample sentence above, every beta-

cell produce some insulin. Note further that we per-

sistently use uninﬂected forms of nouns and verbs

rather than morphologically correct forms in our nat-

ural logic expressions.

PUaNLP 2016 - Special Session on Partiality, Underspeciﬁcation, and Natural Language Processing

336

3.1 Subclass Through Copula Sentence

Within the above natural logic afﬁrmative sentence

template in (1), there is an extremely important sub-

class, namely the copula sentence form shown in (3)

Cterm

isa Cterm

(3)

known from categorial syllogistic logic. cf. (Nils-

son, 2013). Using the symbol C in class names as in

isa C

we say that the class C

specializes the class

, and, conversely, that C

generalizes the class C

As explained in (Nilsson, 2013) the copula form (3)

may actually be understood as a special case of the

above general form (2) with the relation being equal-

ity. For example, the sentence betacell isa cell is pred-

icate logically construed as

∀x(betacell(x) → ∃y(x = y ∧ cell(y))

giving in turn

∀x(betacell(x) → cell(x))

Note once again that we are not using these predicate

logical forms in our reasoning with natural logic.

The categorial syllogistic no C

isa C

is not made

available, since class disjointness is assumed initially

for pairs of classes by default in our setup, cf. (Nils-

son, 2015), which also discusses the relation to the

well-known square of opposition in traditional logic.

The upshot of our convention is that two classes are

disjoint unless one is stipulated as a subclass of the

other, or that they have a common subclass introduced

by the copula form. This default convention conforms

with use of classes in scientiﬁc practice as reﬂected in

formal ontologies. However, one may observe that the

convention deviates from the description logic prin-

ciple, which follows predicate logic with the open

world assumption.

3.2 Simple and Compound Class Terms

Compound terms Cterm in the natural logic take the

form of a class name C adorned with various forms of

restrictions giving rise to a virtually unlimited num-

ber subclasses of C. Linguistically, this “generativity”

is provided by constructions like restrictive relative

clauses and adnominal prepositional phrases (PPs).

Accordingly, in the present context we consider

Cterm in the form of a class name (noun) C option-

ally followed by

• a stylized relative clause: that R Cterm

or optionally by

• a PP in the logical form R

prep

Cterm, in turn

optionally followed by a relative clause.

The relation R

prep

is to be provided by the entry for

the pertinent preposition in the applied vocabulary.

Sample class terms illustrating these patterns:

cell

cell that produce hormone

cell in pancreas

cell in pancreas that produce hormone

Ontologically, these four classes form a (trans-

hierarchical) diamond by the isa subclass relation:

cell-in-pancreas-that-produce-hormone

cell-that-produce-hormone cell-in-pancreas

cell

The sample class term cell in pancreas that pro-

duce hormone is aligned as in cell (in pancreas) (that

produce hormone).

For the moment we disregard the more tricky re-

strictions provided by adjectives (even when assumed

to behave restrictively), noun-noun compounds

, and

genitives. This is because these constructs, unlike

the case of relative clauses and PPs, do not explic-

itly yield a speciﬁc relation R, cf. the discussions in

(Jensen and Nilsson, 2006; Vikner and Jensen, 2002).

The natural logic sub-language where Cterm is

simply a class name we call atomic natural logic.

As described below, in due course we shall also

admit conjoined constructions with the conjunction

and in class terms, and, further, compound relation

terms, Rterm for R, linguistically comprising selected

adverbs and adverbial PPs modifying verbs and verb

phrases. See also (Andreasen et al., 2014b) for vari-

ous extensions of the natural logic.

3.3 Reasoning with Natural Logic

The key inference rules for the considered natural

logic are the so-called monotonicity rules (van Ben-

them, 1986). They very intuitively admit specializa-

tion of the grammatical subject class and generaliza-

tion of the grammatical object class. Accordingly,

given A isa B and [every ] B R [some] C, one may de-

rive [every] A R [some] C,

A isa B [every] B R [some] C

[every] A R [some] C

and given [every] B R [some] C and C isa D, one may

Some class names (given terms) in an application may

consist of more than one word,but are still to be considered

as simple, ﬁxed terms.

On the Relationship between a Computational Natural Logic and Natural Language

337

insulin

hormone

cell-that-

produce-hormone

produce

cell

cell-that-

produce-insulin

produce

Figure 1: Inferred inclusion by subsumption.

derive [every] B R [some] D,

[every] B R [some] C C isa D

[every] B R [some] D

In particular, these rules provide transitivity of the isa

subclass relationship with R being isa.

In addition, we provide a subsumption inference

rule which makes the properties assigned to classes

act restrictively as detailed with an algorithm in (An-

dreasen et al., 2015). By way of example, the sub-

sumption rule ensures that

cell that produce insulin

is recognized as a specialization of

cell that produce hormone,

given that insulin isa hormone, cf. ﬁgure 1.

3.4 Atomic Natural Logic as Graph

In our framework and prototype system, the core

natural logic introduced above is decomposed into

atomic natural logic devoid of compound class names,

that is, Cterm is simply a class name C. This is ac-

complished by introduction of fresh, internal class

names such as cell-that-produce-hormone, which is

formally conceived of as a class name. In turn, this

is deﬁned by two atomic natural logic sentences

cell-that-produce-hormone isa cell

cell-that-produce-hormone produce hor mone.

The knowledge base of the decomposed sentences

may be viewed as one single labeled graph whose

nodes are uniquely labeled with classes. Except for

those relationships that follow from transitivity, we

make sure that all valid isa relationships between

nodes are materialized in the graph by the subsump-

tion rule, so that for instance

cell-that-produce-insulin isa

cell-that-produce-hormone

is recorded. As such, the graph appears as an

extended formal ontology with the isa relationship

forming the skeleton, as it were.

In our system, besides deductive querying the

graph is used for pathﬁnding between classes (An-

dreasen et al., 2015). Actually, in our system the

atomic natural logic graph is embedded in function-

free logical clauses (e.g. DATALOG). This embedding

approach means that natural logic sentences become

encoded as variable-free logical terms. The logical

variables in the clauses then range over class and re-

lationship terms enabling formulation of the inference

rules and hence reasoning and deductive querying.

The clausal embedding also facilitates formulation of

domain speciﬁc inference rules such as transitivity of

causation (Andreasen et al., 2014b).

4 EXTENDING CORE NATURAL

LOGIC

We now turn to a partial treatment of natural language

sentences using natural logic as a vantage point. In

a conventional approach, one may proceed by devis-

ing a (partial) translation from the considered natural

language text sentence by sentence into natural logic.

Here, we choose to proceed by introducing a series of

conservative extensions to core natural logic. Thus,

these extensions do not increase the semantic range

of the core natural logic as stipulated above. How-

ever, the extensions provide paraphrases commonly

encountered in natural language. These extensions,

when taking jointly, form an extended natural logic

coming closer to free natural language formulations,

remaining, however, within the conﬁnes of the seman-

tics of core natural logic. The extensions may be used

in turn in developing a partial translator from natural

language into core natural logic.

4.1 Extension with Conjunctions

Let us ﬁrst consider conjunctions of Cterms including

the linguistic conjunction and assuming distributive

(in contrast to collective) readings. Conjunctions in

the grammatical object

Cterm

R Cterm

and Cterm

as in

pancreas contain betacell and alphacell

straightforwardly give rise to the decomposition:

Cterm

R Cterm

Cterm

R Cterm

Conjunctions in the grammatical subject

Cterm

and Cterm

R Cterm

as in

betacell and alphacell is-contained-in pancreas

are conventionally interpreted as disjunction rather

than overlap of the two classes and therefore decom-

posed into

Cterm

R Cterm

Cterm

R Cterm

PUaNLP 2016 - Special Session on Partiality, Underspeciﬁcation, and Natural Language Processing

338

These conventions are justiﬁed by the underlying

predicate logical explication of core natural logic.

The linguistic disjunction or in the linguistic sub-

ject seems irrelevant from the point of view of the

considered domains. It may be considered a case

where predicate logic covers more than needed.

More interesting are disjunctions in the grammat-

ical object, viz.

Cterm

R Cterm

or Cterm

which cannot simply be decomposed into two core

natural logic sentences. One approach is to appeal

to a common general term Cterm

sup

for Cterm

and

Cterm

if one is available in the KB. More precisely,

one seeks a Cterm

sup

, such that

Cterm

isa Cterm

sup

Cterm

isa Cterm

sup

and such that for all different Cterm

having these

properties Cterm

sup

isa Cterm

. This supremum re-

quirement seems reasonable in cases where the con-

sidered disjunction is pragmatically relevant at all.

Notice that all of the above reductions of conjunc-

tions endorse the desired commutativity and associa-

tivity properties.

It goes without saying that the presence of con-

junctions together with relative clauses and preposi-

tions gives rise to structural ambiguities. Introduction

of appropriate default readings and/or addition of aux-

iliary parentheses are the simplest ways to eliminate

these.

Collective, i.e. non-distributive, readings such as

co-presence of A and B cause C call for separate

treatment, which goes beyond the scope of the present

approach.

4.2 Extension with Appositions and

Parenthetical Relative Clauses

Let us consider natural logic sentences

Cterm

R Cterm

extended with appositions bounded by commas

Cterm

, [a|an] Cterm

appo

, R Cterm

This is paraphrased into the pair

Cterm

R Cterm

Cterm

isa Cterm

appo

Analogously for the grammatical object, Cterm

, as

betacell produce insulin, a peptide hormone

In this extended natural logic, the pronoun ’that’

is formally set off for restrictive relative clauses as

accounted for above.

By contrast, consider the case of parenthetical rel-

ative clauses for which in the formal natural logic we

use ’which’ together with commas

Cterm

, which R

par

Cterm

par

, R Cterm

as in

insulin, which isa peptide hormone, ...

Retaining logical equivalence, this can be para-

phrased into the joint pair

Cterm

R Cterm

Cterm

par

Cterm

par

and similarly and recursively for Cterm

and

Cterm

par

4.3 Beyond Core Natural Logic

From the point of view of application functionalities,

verbs should be allowed extensions with adverbial

PPs yielding restricted relations, Rterm, for plain R,

say, as in

A produce in pancreas B

with variants A produce B in pancreas and in pan-

creas A produce B, with obvious additional structural

ambiguity problems.

On our agenda for non-conservative extensions of

core natural logic, let us mention passive voice verb

forms, nominalisation, and plural formation. As far

as negation is concerned we rely throughout on the

closed world assumption in the query answering.

Of course there are numerous genuine (that is,

non-conservative) language extensions which go be-

yond the semantic range of core natural logic, even

within the given scope of monadic and dyadic re-

lations in afﬁrmative sentences. Thus admission of

anaphora as in the infamous donkey sentences breaks

the boundaries as mentioned in (Kl

ıma, 2010). An ex-

ample of this is seen in every cell that has a nucleus

is-controlled-by it. The point is that in the applied nat-

ural logic, the subject noun term and the object noun

term are independent, connected solely by the relator

verb and unconnected by anaphora.

5 COMPUTING NATURAL

LOGIC FROM SENTENCES

Methodologically, rather than the usual forward or

bottom-up translation following the phrase structure,

we devise a top down processing governed by the nat-

ural logic. In this process we try to cover as much as

possible of the considered sentence in a (partial) “best

ﬁt” process.

A prototype is under development and is currently

functioning in a preliminary version. In the present

approach natural language input is processed sentence

by sentence. Thus sentential context is not exploited.

Each input sentence is preprocessed for markup, and

the result is further processed by a parser that also

On the Relationship between a Computational Natural Logic and Natural Language

339

insulin

peptide_hormone-producedby-

betacell-in-pancreas

producedby

peptide_hormone

pancreas

betacell

Figure 2: Graph representation of the sentence insulin is a peptide hormone produced by betacells in the pancreas.

functions as a natural logic generator. During the pre-

processing the sentence is tokenised into a list of lists,

where each word from the sentence is represented

by a list of possible lexemes specifying base form

of word and word category (part of speech) for each

word. A lexeme is included as possible if it matches

the lemma of the input form of the word. Finally, the

preprocessing applies a domain speciﬁc vocabulary to

identify multiword expressions in the input sentence.

To ensure that multiwords are treated as inseparable

units they are replaced by unique symbols. A prepro-

cessing of the sentence: Insulin is a peptide hormone

produced by betacells in the pancreas returns the fol-

lowing tagged and lemmatised word list, where the

word sequences ’is a’, ’peptide hormone’ and ’pro-

duced by’ are replaced by symbols:

{{insulin/NN}, {isa/VB}, {peptide hormone/NN},

{produced by/VB}, {betacell/NN}, {in/JJ, in/NN,

in/RB, in/IN}, {pancreas/NN}}

Each possible lexeme for a word (and multiword) W

is included as an element L

i j

if W

has lemma L

i j

for category C

i j

. The categories are denoted using

Penn Treebank POS-tags. VB, NN, RB, JJ and IN cor-

respond to verb, noun, adverb, adjective and preposi-

tion, respectively.

Our approach to recognizing and deriving natu-

ral logic expressions from natural language texts can

be considered a knowledge extraction task where the

goal is to extract expressions that cover as much as

possible of the meaning content from the source text.

The search is guided by a natural logic grammar,

and the aim is to create propositions that comprise

well-formed natural logic expressions. The “best ﬁt”

approach is basically a guiding principle aiming for

largest possible coverage of the input text. Thus, if

an expression, that covers the full input sentence can

be derived, it would be considered the “best”, and if

not, the aim is a partial coverage where larger means

“better”.

In the prototype, we apply a simple approach to

deriving a partial coverage. A “best ﬁt” is provided

by iterating through a series of “sub-sentences” of the

input sentence of repeatedly smaller size until one is

found from which a proposition can be derived. A

sub-sentence arises from removing zero, one or more

words from the input sentence. This approach has ob-

vious drawbacks, most importantly, it’s quite inefﬁ-

cient – especially due to load from forcing the parser

to repeat identical subtasks over and over again. How-

ever, it has one important advantage as a prototype

approach – it allows a clear separation of parsing and

selection of partial expressions.

As far as parsing is concerned, and as already

mentioned, our approach is a top down processing

governed by the natural logic. The word list given as

input is ambiguous due to the multiple categories as-

signed to each word. Thus, the parser should be able

to recognize an input proposition if such exists for at

least one combination of possible lexemes of the in-

put words. Therefore, in addition to processing the

grammar (given below), the parser must ensure that

all combinations are tried before failing the recogni-

tion of a proposition.

Core natural logic, as described in section 3, with-

out the extensions sketched in section 4, can be spec-

iﬁed by the following grammar.

Prop ::= Cterm R Cterm

Cterm ::= NN [RelClauseterm | Prepterm]

RelClauseterm ::= [that|which|who] R Cterm

R ::= VB

Prep

::= IN

Prepterm ::= R

Prep

Cterm

Notice that terminals are speciﬁed as either speciﬁc

words or word categories using Penn Treebank-tags.

Example 1: The preprocessing result of the exam-

ple sentence insulin is a peptide hormone produced

PUaNLP 2016 - Special Session on Partiality, Underspeciﬁcation, and Natural Language Processing

340

cell-that-

produce-insulin

cell insulin

produce

pancreatic-gland

located:in

Figure 3: Graph representation of the sentence cells that produce insulin are located in the pancreatic gland

by betacells in pancreas is shown above. In this case,

the full sentence can be recognized as a proposition

with the given grammar and the following arcs can be

derived from the parse tree:

insulin isa

peptide hormone-producedby-betacell-in-pancreas

producedby betacell-in-pancreas

peptide hormone-producedby-betacell-in-pancreas

isa peptide hormone

betacell-in-pancreas in pancreas

betacell-in-pancreas isa betacell

The corresponding graph is shown in graph form in

ﬁgure 2.

Example 2: Preprocessing the sentence cells that pro-

duce insulin are located in the pancreatic gland leads

to:

{{cell/NN}, {that/WDT}, {produce/VB, pro-

duce/NN}, {insulin/NN}, {locatedin/VB},

{pancreatic gland/NN}}

where ’is located in’ and ’pancreatic gland’ are re-

placed by symbols. As it appears, again the full sen-

tence can be recognized as a proposition. From the

corresponding parse tree the following 3 arcs can be

derived:

cell-produce-insulin locatedin pancreatic gland

cell-produce-insulin produce insulin

cell-produce-insulin isa cell

The result is shown in graph form in ﬁgure 3.

Example 3:The grammar above does not cover con-

junctions. Thus pancreas contains betacells and al-

phacells will not be recognized as a proposition.

However, the best-ﬁt principle will iterate through

partial-cover sub-sentences (where one or more words

are left out). Among these are pancreas contain beta-

cell and pancreas contain alphacell, which will both

be recognized. Thus, in this case continued best-

match iteration will lead to a result that complies with

the conjunction.

6 SUMMARY AND CONCLUSION

We have described how to use dedicated forms of nat-

ural logic for partial computational comprehension of

natural language with a particular view to descriptive

scientiﬁc corpora within the life sciences. The applied

natural logic, called core natural logic, has a well-

deﬁned semantic foundation in predicate logic, and

we have devised an appropriate inference engine for

query answering functionalities. Core natural logic

is extended in the paper with common paraphrase

schemes.

Our approach features – as explained and illus-

trated with the ﬁgures – a shared graph represen-

tation of all sentences. In this representation con-

cepts are (at least ideally) uniquely represented as

nodes. The directed arcs represent atomic natural

logic sentences, and appropriate labelling conventions

ensure that the individual natural logic sentences can

be reconstructed from their atomic components in the

graph, modulo paraphrasing. The graph representa-

tion eases pathﬁnding between concepts and serves

deductive querying.

Finally, we illustrate processing of sentences from

the text using core natural logic. As a next step, we

are going to try our prototype on selected life science

corpora.

ACKNOWLEDGEMENT

The authors would like to thank the reviewers for their

careful reading of the draft with suggestions for im-

provements and clariﬁcations, and for provision of

further comprehensive references for our continuing

work.

REFERENCES

Andreasen, T., Bulskov, H., Fischer Nilsson, J., and Jensen,

P. A. (2014a). Computing pathways in bio-models de-

rived from bio-science text sources. In Proceedings of

On the Relationship between a Computational Natural Logic and Natural Language

341

the IWBBIO International Work-Conference on Bioin-

formatics and Biomedical Engineering, Granada,

April, pages 217–226.

Andreasen, T., Bulskov, H., Nilsson, J. F., and Jensen, P. A.

(2014b). A system for computing conceptual path-

ways in bio-medical text models. In Foundations of

Intelligent Systems, pages 264–273. Springer.

Andreasen, T., Bulskov, H., Nilsson, J. F., and Jensen, P. A.

(2015). A system for conceptual pathway ﬁnding and

deductive querying. In Flexible Query Answering Sys-

tems 2015, pages 461–472. Springer.

Andreasen, T. and Nilsson, J. F. (2014). A case for em-

bedded natural logic for ontological knowledge bases.

In 6th International Conference on Knowledge Engi-

neering and Ontology Development.

Bittner, T. and Donnelly, M. (2007). Logical properties of

foundational relations in bio-ontologies. Artiﬁcial In-

telligence in Medicine, 39(3):197–216.

de Azevedo, R. R., Freitas, F., Rocha, R., de Menezes, J.

A. A., and Pereira, L. F. A. (2014). Generating de-

scription logic ALC from text in natural language. In

Foundations of Intelligent Systems, pages 305–314.

Springer.

Fuchs, N. E., Kaljurand, K., and Kuhn, T. (2008). Attempto

controlled english for knowledge representation. In

Reasoning Web, pages 104–124. Springer.

Fyodorov, Y., Winter, Y., and Fyodorov, N. (2003). Order-

based inference in natural logic. Log. J. IGPL,

11(4):385–417. Inference in computational seman-

tics: the Dagstuhl Workshop 2000.

Jensen, P. A. and Nilsson, J. F. (2006). Ontology-based

semantics for prepositions. In Saint-Dizier, P., editor,

Syntax and semantics of prepositions, pages 217–233.

Springer Science & Business Media.

ıma, G. (2010). Natural logic, medieval logic and formal

semantics. Logic, Language, Mathematics, 2.

Kuhn, T., Royer, L., Fuchs, N. E., and Schroeder, M.

(2006). Improving text mining with controlled nat-

ural language: A case study for protein interactions.

In Data Integration in the Life Sciences, pages 66–81.

Springer.

MacCartney, B. and Manning, C. D. (2009). An extended

model of natural logic. In Proceedings of the eighth

international conference on computational semantics,

pages 140–156. Association for Computational Lin-

guistics.

Nilsson, J. F. (2013). Diagrammatic reasoning with classes

and relationships. In Visual Reasoning with Diagrams,

pages 83–100. Springer.

Nilsson, J. F. (2015). In pursuit of natural logics for

ontology-structured knowledge bases. In The Sev-

enth International Conference on Advanced Cognitive

Technologies and Applications.

Pratt-Hartmann, I. and Moss, L. S. (2009). Logics for

the relational syllogistic. Review of Symbolic Logic,

2(4):647–683.

Schulz, S. and Hahn, U. (2004). Ontological foundations of

biological continuants. In Formal ontology in infor-

mation systems. Proceedings of the 3rd international

conference-FOIS, pages 4–6.

Smith, B., Ceusters, W., Klagges, B., K

ohler, J., Kumar, A.,

Lomax, J., Mungall, C., Neuhaus, F., Rector, A. L.,

and Rosse, C. (2005). Relations in biomedical ontolo-

gies. Genome biology, 6(5):R46.

Thorne, C., Bernardi, R., and Calvanese, D. (2014). De-

signing efﬁcient controlled languages for ontologies.

In Bunt, H., Bos, J., and Pulman, S., editors, Com-

puting Meaning, Volume 4, volume 47 of Text, Speech

and Language Technology, pages 149–173. Springer.

van Benthem, J. (1986). Essays in Logical Semantics, Vol-

ume 29 of Studies in Linguistics and Philosophy. D.

Reidel, Dordrecht, Holland.

Vikner, C. and Jensen, P. A. (2002). A semantic analysis of

the english genitive. interaction of lexical and formal

semantics. Studia Linguistica, 56(2):191–226.

Yu, A. C. (2006). Methods in biomedical ontology. Journal

of Biomedical Informatics, 39(3):252 – 266. Biomed-

ical Ontologies.

PUaNLP 2016 - Special Session on Partiality, Underspeciﬁcation, and Natural Language Processing

342