Logic, Logicism

and the Semantics

of Word Formation
 The Plan
• I’ll begin by trying to motivate a research program.
• Then I’ll formulate a proposal.
• Then I’ll illustrate it with some case studies.
 Some Linguistic Background
• The semantics of the lexicon is a weak point in Montague’s
framework. Montague was content with accounting for
logical words, especially quantifiers.
• David Dowty in his 1979 book attempted to account for
systematic relations among lexical items using a somewhat
conservative extension of Montague’s framework, with
limited success.
• Using events, Terence Parsons achieved better success in his
1990 book, confining himself mainly to tense and aspect.
But Parsons’ account of events is skeletal, making the
theory less explanatory than one would like.
• In his 1989 Tianjin lectures, Emmon Bach coined the term
“natural language metaphysics” and sketched some
examples.
 Philosophical Background
• Carnap’s ideas about how to deploy “meaning postulates”
were ahead of his time—and ahead of the logical resources
available to him.
• In Testability and Meaning (1936) Carnap abandons the
idea of defining “dispositional” predicates like soluble, but
instead deploys (rather simplistic) postulates about the
phenomenon of dissolving.
• Carnap called these “bilateral reduction sentences.”
• Later (1952) he included such postulates, as well as e.g.
transitivity postulates for comparatives, under the heading
of “meaning postulates.”
• They do, of course, have to do with meaning. But they are
also microtheories of the common sense world.
 Common Sense Logicism
• In work going back to 1969, John McCarthy (the computer
scientist, not the linguist) recommended understanding
common sense as the key to success in Artificial
Intelligence, • • •
• and logical formalization as the method for obtaining this
understanding.
• McCarthy’s program attacted many talented people, and
led to specific formalizations of several commonsense
domains, • • •
• including reasoning about actions and change, • • •

• and to logical innovations, including McCarthy’s Situation

Calculus and other formalisms for planning and temporal
reasoning, • • •
• and nonmonotonic logics.
 Coverage
• Formalizing commonsense reasoning overlaps with another
area of AI: commonsense physics.
• Topics that have been addressed (from Ernest Davis’
Representations of Commonsense Reasoning, 1990)
include:

1. Measurements
2. Reasoning about time
3. Reasoning about space
4. Reasoning about solids and liquids
5. Reasoning about minds
6. Planning and goal formation
 The Methods Are Different from Those
of Philosophical Analysis
• Philosophical analysis is concerned with definitions, not
postulates.
• Philosophical analyses will only sustain a moderate degree
of complexity until they become unmanageable. Therefore
they are not well suited to complex domains.
• Philosophical analyses are not exception-tolerant, which
makes them brittle.
• Computer scientists are familiar with techniques for
managing large and complicated formalizations.
• Software engineering techniques provide a model for how to
do this that is part of the common knowledge of the
discipline.
• These techniques can be extended to axiomatizations, and
include computerized techniques for managing, testing, and
debugging axiomatizations (e.g., theorem-proving,
simulation, actual robot-based implementations).
• These have been tested by the common sense knowledge
community on several benchmark challenges, e.g. the
“egg-cracking problem.”
 Nonmonotonic Logic
• Commonsense generalizations can involve exceptions.
• It is much easier to maintain an axiomatization by adding
axioms rather than by withdrawing and qualifying old
axioms.
• This makes exception-tolerant logics very useful in
axiomatizing comonsense domains.
• Important for our purposes: Some versions of NM logic
support explicit statements of commonsense normalities.
 McCarthy’s Circumscription
• McCarthy’s proposal for nonmonotonic logic was
Circumscription.
• It is logically conservative—it is based on classical
logic—and uses higher-order logic.
• Exception tolerant generalizations are formulated using
abnormality predicates: for instance, ‘A car will start if you
turn the key in the ignition’ would be formalized as:
[Car(x) ∧ Key(x, y) ∧ TurnKeyInIgnition(x, y, t)]
∧ ¬Ab1(x, y, t) →
Start](x, t)

• Here, Ab1 is an abnormality predicate. The entire

generalization is a normality or normality condition.
• Models for circumscription simultaneously minimize
abnormalities: that is, a theory T implies a formula φ not if
every model whatsoever that satisfies T satisfies φ, • • •
• but if every abnormality-minimal model that satisfies T
satisfies φ. (This should remind you of how the semantics
for conditional logics works, only here we are talking about
preferred models, not preferred worlds.)
• Abnormality minimization can be explicitly defined using
higher-order quantification over the abnormality predicates.
• You can refine nonmontonic axiomatizations by adding
qualifications to the abnormality theory.
• For instance, if you discover that a dead battery is an
exception to the car-starting regularity, you do not have to
withdraw the axiom.
• You can add an axiom:
DeadBattery(x, t) → Ab1(x, y, t)

• The inventory of abnormality predicates is open-ended.

• So, if you like, you can add abnormality predicates to the
exceptions, and qualify them later as well.
 Summarize the Motivation
• We should expect commonsense knowledge to be built into
many of the basic processes of word formation, • • •
• because, for instance, Aktionsarten, mass-count distinction,
and temporal and causal locutions should reflect and reveal
that knowledge.
• So the semantics of word-formation should appeal to
formalizations of commonsense theories of the physical,
cognitive, and social world.
• Work AI on commonsense knowledge and reasoning shows
us how to do this.
• The task, of course, is open-ended, but work in this field to
date has been encouraging.
Part II: The Proposal
• To make Montague’s framework capable of dealing with
lexical semantics, make two additions:

1. Add a domain of eventualities to the theory.

(Eventualities can be treated as individuals, so this is not
a modification of the logic, but an ontological extension.)
2. Make the logic nonmonotonic, using circumscription.
(This is a very conservative extension, consisting merely
of the addition of an open-ended special class of
abnormality predicates, and of a restriction of models to
those that are abnormality-minimized.)
Part III: Case Studies
Case 1: the -able Suffix
According to Laurie Bauer English Word-Formation,

there are, however, a number of processes which are

usually considered to be derivational and which do
display semantic regularity: consider, for example, the
formation of English adjectives in -able from transitive
verbs. This is probably a case where there are no gaps
in the derivational paradigm (any transitive verb can act
as the base) and the adjectives are semantically regular,
meaning ‘capable of being Ved’ (where V is the verb in
the base).
 It is Not as Simple as This
• In fact, the -able suffix illustrates a number of challenging
characteristics.

1. There is variation in the meanings it assumes, but the

variations are closely related shades of meaning. It is
hard to tell whether listing senses, to look for a single
underlying meaning allowing for different uses, to make
the meaning context-dependent.
2. The meanings themselves are difficult to formalize.
3. These meanings seem to invoke commonsense real world
knowledge rather than linguistic knowledge.
4. While it is possible to formalize separate instances in a
class of semantically similar group, it can be difficult to
formulate a general template for the entire class.
5. There are exceptional patterns.
• Typically, the -able construction works with verbs, and
these are transitive and broadly telic. (There are some
exceptions—e.g. Consider knowledgeable and palatable.
• Often, these verbs have three characteristics.

(i) They correspond to procedures that are in the normal

repertoire of actions of human agents.
(ii) Often there are normal or standard ways of initiating
these actions, and there is an end state associated with
the performance of the actions.
(iii) Often, this end state can be evaluated in terms of its
success in aiding the execution of a plan.

• In the simplest and easiest to analyze cases, all of these

features are present.
 Patterns
• From examining a corpus of -able words, I’ve identified
three major semantic patterns:

1. Conditional ‘if-will’ (this can be more or less occasional

or generic). The generic usage is what philosophers call
the “dispositional” case. (soluble is an example.)
2. Unharmful outcome. (edible is an example.)
3. Worthiness or obligatoriness. (honorable and (payable
are examples.)

• The idea would be to concoct a family of similar meaning

postulate templates, one for each semantic pattern.
 Formalise the Conditional Pattern
• I rely on a compositional theory of telic eventualities, an
idea I got from from Marc Moens and Mark Steedman,
though it isn’t unique to them.
• A typical telic eventuality consists of:

(1) An inception (an event, which often has an agent),

(2) A body (a process, which progresses, often by
measurable stages, towards
(3) A culminating state.

• There are causal connection between these parts.

• I’m skipping some details here, but • • •

• for any telic eventuality there will be default normalities

linking (1) the initiation with the body, and (2) the body
with a culminating state.
• These normalities may be more or less qualified, ad hoc and
depending on the specifics of the case.
• The normalities represent causal connections, but I’m not
presenting this as an analysis of causality.
 Formalization of Water-Soluble
∀x[Water-Soluble(x) ↔
∀e[ [Put-In(e) ∧ theme(e) = x ∧ Water(medium(e))
∧ Occur(e) ∧ ¬AbInitiate-Dissolve )(e)] →

∃e′[ Telic(e′ ) ∧ Dissolution(e′) ∧ Occur(e)′

∧ theme(e′ ) = x ∧ inception(e)′ = e

∧ [¬AbProcess-Dissolve (e′ ) →

∃e′′ [culmination(e′ ) = e′′ ∧ Occur(e′′ )

∧ (In-Solution(e′′ ) ∧ theme(e) = x

∧ medium(e′′ ) = medium(e)] ]

]
]
]
 Remarks on the Formalization
• The main connective is a biconditional, so contrary to what
Carnap thought was possible, this is a definition of a
dispositional property • • •
• but the definition involves normality conditions.
• Many features of the postulate would be inherited from
generalizations about the formalization of telic eventualities.
• Any one-off formalization such as this must be provisional.
Large scale axiomatic knowledge bases need to be developed
and tested before one can have much confidence about
particular examples.
• I haven’t tried to incorporate time into the formalization.
Doing so raises many difficulties.
 Normality Conditions
• Almost everywhere you look in lexical semantics, you find
normality conditions.

1. A wastebasket is a basket-sized container that is

normally used to contain waste.
2. A pencil sharpener is a device that is normally used to
sharpen pencils.
3. To oar a boat is to move the boat by means of oars, used
in the way oars are normally used to move boats.
4. To wipe a table dry is to dry the table by wiping it with
something, in the normal way that tables are wiped with
something in order to dry them.
• The pervasiveness of such examples is a powerful argument
for a nonmonotonic logical framework
Case 2: Causatives
• I agree with Dowty about the nature of the overall problem:
provide an account of the semantic derivation of, for
instance, causative open from stative open.
• But I believe that the idea of using a propositional Cause
relation (of type hhe, ti, hhe, ti, tii) to do this is poorly
motivated and unworkable.
• That idea runs into logical problems.
• Also—and I regard this as a central problem—you can
cause a window to be open without opening it. (You get
someone else to open the window.)
• Some people think of this as a matter of “immediate
causality.” But the notion of “immediate causality” is
suspicious, and anyway it’s clearly wrong for this purpose.
• I have telic eventualities in my ontology.
• These are ready-made causal mechanisms, with their own
causal dynamics.
• So I don’t need to get the causality of causatives from
elsewhere.
• From my point of view, the problem is this: if a telic
eventuality culminates in a state of type X, • • •
• then what is its agent?
• It’s a matter of agency, not causality per se.
 The Simple Case is Easy
• The telic eventuality consists of an inception, a body and a
culminating state.
• It’s a causally isolated microsystem.
• The inception has an agent.
• Then the agent of inception is the agent of the containing
telic eventuality.
 The Simple Picture
Door’s opened
state

Causation

e1 Causation
Mary’s action
e2
of unlatching Door’s opening
and pushing process
• In the diagram, we’ve added causal links to complex
eventualities, • • •
• borrowing ideas from causal modeling in statistics and
qualitative causal reasoning.
• Among other things, event complexes can be causal graphs.
 But the Story is More Complicated
• Telic complexes can involve other eventualities.
• Some of these can be contributing causes.
• And some of the contributing causes can have agents.
• Which agent is the agent of the telic eventuality?
• We need to appeal to a version of the distinction between
animate and inanimate agents here.
• I’ll invent my own terminology: “efficient” versus
“automatic” agents. The latter are self-initiating sources of
causality. The former merely transmit causality.
• In the following hypothetical diagram, automatic agents are
labeled with “A”, efficient agents with “E”.
E: efficient agent
A: automatic agent

a2 , E

a1 , A

Inception Body Culmination

a0 , E

a3 , A

a4 , A
 Illustrating the Diagram
• Mary is directing a blacksmithing operation.
• a0: An assistant, Frank, initiates the process by putting bar
stock on the anvil.
• a2: Mary turns on a power hammer, which a1 hammers the
stock.
• a4: A switch turns on a bellows a3 which maintains the heat
of the forge.
 Conditions for Assigning an Agent
• An efficient agent is precluded by any efficient agent
between it and the body.
• An automatic agent is precluded by any agent between it
and the body, and by any causal ancestor of the body with
a efficient agent.
• The agent of the inception is precluded by any efficient
agent elsewhere in the causal ancestry of the body.
• These conditions make Mary the agent in the example. She
forged the metal.
Progressive
 Inertial Histories and the Modal Approach
• Dowty uses branching time and “inertial histories” to
account for progressive.
• The idea of inertia is close to the idea that I’m
recommending, of normal causal development of processes.
• But Dowty’s approach is modal: what is modal is closed
under conjunctions, and this is problematic.
• The problem is that simultaneous causal processes can
conflict.

1. While he was composing a requiem mass, Mozart was

dying.
2. Two trains are speeding down a single-track line, heading
srraight for a collision. One is going west to Chicago, the
other is going east to Detroit.
3. Judy is boiling a kettle of water on her gas stove. At the
same time, unknown to her, workmen are shutting off
her gas.

• I conclude that we don’t want global inertia, but local

inertia, confined to coherent causal microsystems • • •
• Even if the evidence isn’t entirely unequivocal.
 Continuous and Discrete Processes
• It’s natural to think of some commonsense processes as
discrete, and some as continuous. So it’s important to
provide a theory of becoming that works for both cases.
• ‘She filled the tub with water’.
• ‘She filled the jar with marbles’.
 What’s in a State?
• I want to think of states as loci of information, represented
by variables.
• These variables can track matters of degree. If a door is
capable of being open or closed, a variable will track the
extent to which it is closed—perhaps the same variable,
inversely, tracks extent to which it is open.
• Processes also supply values to variables.
• If a process P typically culminates in a state S, the two will
share many variables—there will be a match between the
two eventuality types.
 Becoming as Approaching a Limit
• The idea is that a telic process, if it follows its inner inertia,
will approach its culmination, is very appealing.
• The standard epsilon-delta definition of approaching a limit
(from below) for a continuous function f tells us this:

f approaches a certain value y as its argument

approaches x if for every range of values close to the
target value y there is an argument interval including
x such that all arguments in this interval take values
in the target range.

• That is, if we want the values of f to be this close to y, we

can always do that by choosing arguments for f this close
to f .
 Proposed Definition for Successful Becoming
At time t0 and history h, e is a process culminating in a state s
in which a has feature F .
Definition: Become(e, s, h, t0, F, a).
Let e be a process, s a state, a an individual, F a feature, h a
history, and t0 a time. Then Become(e, s, h, t0, F, a) iff:
(1) THEME(e) = a, Holds(e, t0, h), and Supports(s, F, a).
(2) The following conditions hold:
(i) There is a least upper bound t1 for
{t | Holds(e, h′, t)}.
(ii) For all states s1 such that Comparable(s, s1, F, a),
there is a time t2 ≤ t3 such that for all t′ ∈ [t2, t3]:
NoLess(s′ , s1 , F, a),
where Exhibits(e, s′, h, t′).
The idea is that if you choose any possible state of F-edness for
a, you can find a time in the process where F (a) is persistently
at least as true as is in this state.
We then say that becoming is a matter of successful becoming
in all inertial histories.
At time t0 and history h, e is a process tending toward a
culminating state in which a has feature F . Let e be a process,
a an individual, F a feature, h a history, and t0 a time.
Definition: Become(e, s, h, t0, F, a) iff:
for all appropriately nonabnormal histories h′
that are nonabnormal with respect to e, there
is a state s such that Become(e, s, h′, t0, F, a).
 There are Two Cases
• In the case of a continuous process—for instance, a process
in which Judy goes home—
• this is equivalent to the standard limit definition.
• It says that in each inertial history Judy arrives home, and
her distance from home approaches 0 as a limit as time
approaches the time of her arrival.
• In the discrete case—for instance, a process of filling a shelf
with books—the condition simply means that the process
has the right type and that in any inertial history the last
state of the process is one in which the shelf is full.
• This may seem too simple a condition, but as far as I can
tell it is hard to find convincing counterexamples.
 An Exceptional Verb
• In lexical semantics, one comes to expect elements, one
finds exceptional semantic subpatterns.
• I have only found one such pattern for progressive: this is
‘winning’, as applied, for instance, to games.
• To say that a player is winning a game doen’t mean that in
all normal continuations the player will win.
• It seems simply to mean that the player has a higher score.
• There may be other exceptional cases, but so far I haven’t
found any.
 Conclusion
• I’ve convinced myself that this is a promising and
worthwhile research program.
• Hopefully, the case studies helped to convey why I think
that.
• The trouble is that pursuing this project properly is just as
much a matter of computer science as of linguistics or logic,
• • •

• and interdisciplinary projects can be difficult to sustain.

• Also, I don’t think the prospects for funding this research
are very encouraging, at least in the US.
• As far as I can see, this is not a project with commercial
applications.
Bibliography

[Bach, 1989] Emmon Bach. Informal Lectures on Formal Semantics. State University of
New York Press, Albany, NY, 1989.
[Davis, 1991] Ernest Davis. Representations of Common Sense Knowledge. Morgan
Kaufmann, San Francisco, 1991.
[Dowty, 1979] David R. Dowty. Word Meaning in Montague Grammar. D. Reidel
Publishing Co., Dordrecht, Holland, 1979.
[Group, 2000] Formal Reasoning Group. Web pages for the formal reasoning group.
http://www-formal.stanford.edu/, 2000.
[Lifschitz, 1988] Vladimir Lifschitz. Circumscriptive theories: A logic-based framework for
knowledge representation. Journal of Philosophical Logic, 17(3):391–441, 1988.
[Morgenstern, 2001] Leora Morgenstern. Mid-sized axiomatizations of common-sense
problems: A case study in egg-cracking. Studia Logica, 67(3):333–384, 2001.
[Parsons, 1990] Terence Parsons. Events in the Semantics of English: a Study in
Subatomic Semantics. The MIT Press, Cambridge, Massachusetts, 1990.
[Shanahan, 2001] Murray Shanahan. An attempt to formalize a nontrivial benchmark
problem in common sense reasoning. Artificial Intelligence, 153(5–6):141–165, 2001.
[Thomason, 1999] Richmond H. Thomason. Formalizing the semantics of derived words.
Unpublished manuscript, available at
http://www.eecs.umich.edu/˜rthomaso/documents.html., 1999.
[Thomason, 2004] Richmond H. Thomason. Non-monotonic formalisms for natural
language semantics. http://web.eecs.umich.edu/ rthomaso/documents/nls/nms.pdf,
2004.
[Thomason, 2005] Richmond H. Thomason. Formalizing the semantics of derived words.
http://web.eecs.umich.edu/ rthomaso/documents/nls/nmslite.pdf, 2005.
[Thomason, 2015] Richmond H. Thomason. Formal semantics for causal constructions. In
Bridget Copley and Fabienne Martin, editors, Causation in Grammatical Structures,
pages 58–75. Oxford University Press, Oxford, 2015.
http://web.eecs.umich.edu/ rthomaso/documents/nls/cause.pdf.