PhD/Postdoc
project descriptions for the research program
Between
Logic and Common Sense: The Formal Semantics of Words
supported
by an NWO Vici grant
February
2010
Principal investigator: Yoad
Winter
Expected starting date for all projects: 1 October 2010 (negotiable)
Project 1: PhD/Postdoc project in
Formal Semantics – Logic and the semantics of concepts
Remark: All cross-references
below refer to the general program: http://www.phil.uu.nl/~yoad/vici/ViciProgram-Winter.pdf
The project
will model the effects of lexical meanings on logical meanings, employing
theories of concepts in cognitive psychology, especially Prototype Theory. An
important test case is described below, following work done in Kerem, Friedmann
and Winter (2009). Many relational expressions are assumed to have
representations as in (11), with defeasible and indefeasible restrictions on
their meaning:
(17) Functionality: pinch, give birth to, kick
Transitivity: equal,contain,
consecutive
Symmetry: overlap, similar, parallel
Asymmetry: stand on, follow
The Maximal
Typicality Hypothesis (section 4.2) assumes that such restrictions interface
logical modules of entailment. Project 1 will test and generalize this
hypothesis to other domains of formal semantics besides reciprocity and
relational meanings.
Research
goal:
Use and
generalize the Maximal Typicality Hypothesis for specifying the logical
structure in the common sense meaning of relational and other expressions and
its interactions with logical meanings.
Methodology:
The project will involve
three tasks:
1. Using the
correspondence principle (section 4.3) for elaborating the logical structure of
feature-based representations of relational (11) and other expressions.
2. Explaining
the interactions of these representations with the logical expressions like
reciprocals (each other), using the Maximal Typicality Hypothesis and
previous work on reciprocals (Dalrymple et al. 1998, Winter 2001b, Sabato and
Winter 2005, Kerem, Friedmann and Winter 2009).
3. Using the
plug-in principle on pragmatic effects with reciprocals in the formal
pragmatics of reciprocals and other logical expressions (Roberts 1987,
Schwarzschild 1996).
Contribution
to the program:
By modeling the interactions
between lexical semantics and logical semantics, project 1 will examine the
Maximal Typicality Hypothesis of Content Sensitive Formal Semantics in an
important test case.
Project 2:
PhD/Postdoc project in Psycholinguistics – Typicality effects with logical
expressions
Remark: All cross-references
below refer to the general program: http://www.phil.uu.nl/~yoad/vici/ViciProgram-Winter.pdf
According to
the main hypothesis (section 3) and its elaboration in the Maximal Typicality
Hypothesis (section 4.2), entailment involves semantic manipulation of
feature-based representations of common sense meanings. An expected result is
that there should be a clear reflection of typicality effects in the logical interpretation
of sentences. Project 2 will experimentally investigate this expectation for
typicality of relational expressions appearing within reciprocal sentences,
thereby informing the theory developed in project 1. The project will experimentally
address three types of data:
-
Typicality effects with relational
expressions and their account using feature-based representations.
-
The influence of such effects on
semantic judgments about sentences with reciprocal expressions: each
other, mutually, one another.
-
The expected parallelism (cf.
Maximal Typicality Hypothesis) between this influence and the influence of
typicality on the interpretation of modification constructions: red
car, red hair, striped apple.
The project
will aim to reproduce and systematize results by Kerem, Friedmann and Winter
(2009), and to check experimentally further implications of the correspondence
between lexical meaning and logical meaning.
Research
Goal:
Characterizing
experimentally the effects of typicality on the logical interpretation of sentences;
comparing such effects reciprocals to parallel effects with modification
constructions.
Methodology:
A
battery of tests will be developed, following Kerem, Friedmann and Winter
(2009), for characterizing typicality effects with various expressions (e.g.
transitive verbs). The experiments will be run together with a research
assistant. For testing typicality, the project will use standard methods of prototype theory (Smith 1990,
Connolly et al. 2007). For instance, in
(11) we hypothesize that typical situations with the concept pinch
involve only one patient per agent simultaneously. To check this hypothesis,
Figures like 4a and 4b will be presented separately to subjects, who will be
asked to classify them as appropriate or inappropriate pinching activities.
If our hypothesis is correct, we expect subjects to respond positively to both
figures, but react more quickly to Figure 4a than to Figure 4b.
Figure
4: typical
pinching vs. atypical pinching
In order to check the hypothesis that
typicality affects reciprocal sentences, subjects will be shown sentences with
a reciprocal verb phrase (e.g. pinch each other) and will be
asked to evaluate them in different situations like Figures 4c and 4d. If our
hypothesis about the effects of typicality on reciprocal sentences is correct,
two phenomena are expected. First, speakers are expected to classify Figure 4c
as a “pinch each other” activity, and do that at least as quickly as with
Figure 4d. Second, this tolerance of “missing” pinching relations in Figure 4c
is expected to crucially depend on the typicality test for the verb pinch.
Thus, tolerance to “missing” pinching relations will be compared to other cases
such as know each other (2). It is expected that under time pressure,
tolerance to exception in the latter case is significantly lower.[1]
In order to test the parallelism between different typicality
effects on sentential meaning, the results in both tests will be compared to
the results by Smith et al. (1988) about modification constructions. The
project will examine if the similarity-based model of Smith et al. can be
extended to fit the results with reciprocal expressions. If needed, further
experiments with modification constructions will be designed and run in order
to also evaluate the extended model with modification constructions.
Case study:
A pilot version of the
typicality test was run with 53 native Hebrew speakers (Kerem, Friedmann and
Winter 2009). The results showed clear typicality effects supporting our
hypothesis about many transitive action verbs like pinch, in contrast to
other verbs.
Contribution
to the program:
By characterizing typicality
effects with relational expressions and semantic judgments on reciprocal
sentences, project 2 will provide experimental data informing the part of
Content Sensitive Formal Semantics developed in project 1. By comparing these
results to other experimental data on modification constructions, the project
will contribute to the development of the general framework.
Project 3: PhD/Postdoc project in
Formal Semantics – Quantification and spatial expressions
Remark: All cross-references
below refer to the general program: http://www.phil.uu.nl/~yoad/vici/ViciProgram-Winter.pdf
A central aim of this program is to lay the foundations for a general
theory of the lexical semantic parameters that affect entailment phenomena.
Project 3 will concentrate on spatial/mereological relational
expressions as in (18), whose effects on entailment go beyond the realm of
typicality-based accounts.
(18) Locative prepositions: inside, outside, behind
Directional prepositions: around, to, from
Dimensional adjectives: big, far, narrow, fast
Geometrical relations: cross,
overlap, contain, touch, be disjoint from,
be
adjacent to, be parallel to
Mereological relations: constitute, comprise, member of
Group-referring nouns: group,
couple, forest
It was only rarely observed that the common sense effects of spatial/mereological
expressions on entailments are as far-reaching as those of other
well-appreciated phenomena like generic sentences (Carlson and Pelletier 1995),
tense and aspect phenomena (Dowty 1979, Hay et al. 1999, Winter 2006),
plurality (Link 1983, Scha 1981, Winter 2001a) and intensionality (Zimmermann
2006).
Examples (3a-c) above illustrated this
claim for the spatial words far and near and plural
definite descriptions. The same is true for indefinite descriptions.
Suppose that you are in
(19) a. We are far from a gas station. b. We are near a gas station.
The two sentences have different implications for the fate of your
journey. While (19b) means that some gas station is nearby, (19a) does not
simply mean that some gas station is far away. Rather, (19a) conveys the
stronger, unfavorable proposition that all gas stations are far from
your location. Thus, in (19b) the indefinite a gas station is
existential as expected by traditional treatments, but replacing the relational
expression in (19b) by its antonym far from leads to a universal
statement (19a). The lexical properties of the main relation in (19) interact
with quantificational devices to express a statement about a spatial
concept: the location of gas stations.
Quantificational variability with plural definites is not
restricted to locative prepositions, as the following examples with spatial
transitive verbs illustrate:
(20) a. Zeeburg contains
the industrial zones.
b.
Zeeburg overlaps the industrial zones.
(21) a. Mary circled
the lakes.
b. Mary reached
the lakes.
(20a)
entails that Zeeburg contains all the industrial zones under discussion.
However, for (20b) to be true it is sufficient that one industrial zone
overlaps Zeeburg. Similarly, in (21a) the relation circle requires that
the path described by Mary's movement surrounds all the lakes, but (21b) can
hold if Mary only reached one lake, adjacent to others in a big lake district. Similar effects on logical semantics appear
with the other spatial/mereological expressions in (18). The project
will adopt the following hypothesis:
Spatial Quantification Hypothesis:
Quantificational variability appears
in entailments with spatial/mereological
relations and in/definite descriptions, because descriptions can function as direct arguments of those
relations.
Descriptions
are assumed to be arguments of predicates, as in previous analyses of semantic
incorporation (Zimmermann 1993, Farkas and de Swart 2003, McNally and Van
Geenhoven 2005). The Spatial Quantification Hypothesis is an innovative
contribution to these accounts in the realm between logic and common sense
meaning. In (3a), the description the lakes, denoting a location,
is a direct argument of the spatial relation far from. Universal
quantification over lakes follows from the definition of this spatial concept
– being far from a location means being far from all its
sub-locations. In (3b), existential quantification over lakes follows from the
different definition of the spatial concept near – being near a location
only requires being near one of its sub-locations. Thus, the in/validity
of the entailments in (3c) follows from the Spatial-Quantification Hypothesis,
and similarly for examples (19)-(21).
Research
goal:
Characterize
the lexical meanings of spatial and mereological expressions and the ways they
affect entailment modules using the Spatial Quantification Hypothesis.
Methodology:
This PhD/Postdoc project will
involve theoretical work, and use questionnaires and linguistic judgments of
native speakers to support its claims. The theory developed using the Spatial
Quantification Hypothesis will extend the theory of project 1 for spatial
relations.
Contribution
to the program:
While it has
been previously proposed (Gärdenfors 2000) that models of typicality are
related to spatial models, this possibility has not been empirically tested in
formal semantics. Project 3 will thus make an important contribution to
characterizing effects of common sense meanings on entailments.
Project 4: PhD project in Computational Linguistics – Formal
semantic annotation of textual entailments
Remark: All cross-references
below refer to the general program: http://www.phil.uu.nl/~yoad/vici/ViciProgram-Winter.pdf
The RTE corpus (Dagan et al. 2006), which is the only currently available
resource of textual entailments, marks candidate entailments as valid/invalid.
However, the RTE corpus makes no indication of the linguistic and informational
processes that underlie entailment. This situation is comparable to the
situation of statistical parsing prior to the introduction of syntactically
annotated corpora like the Penn Treebank (Marcus et al. 1993). In the lack of
theoretical conceptions and empirical documentation of the processes that
support entailments, current works that attempt to automatically recognize
entailment have mostly had to rely on heuristic principles. Some recent works
(Bos and Markert 2005, MacCartney and Manning 2007) stress the importance of lexical-logical principles (Van Benthem 1987, Sànchez-Valencia 1991, Fyodorov, Winter and Francez
2003) for models of natural language entailments. However, semantic theory
itself currently lacks the core principles and typologies that would comprehensively
support this challenge for natural language processing. This project will use
the new framework of Content Sensitive Formal Semantics, together with previous
results, for analyzing and annotating the entailments of the RTE.
Research
Goal:
Develop a
version of semantic theory that is suitable for annotating a corpus of textual
entailments.
Methodology:
Project
4 will both add data to be modeled by the theory developed in projects 1-3 and
be used for testing its applicability for the central challenge of entailment
recognition. Entailments from the RTE corpus will be annotated using
sound semantic relations with a rigorous interpretation in lexical and formal
semantics. The annotation scheme and its theoretical foundations will be
developed by a PhD student and the principal investigator. The actual
annotation will be performed by two assistants, supervised by the principal
investigator and supported by the PhD student.
Pilot
studies:
As
an initial examination of the feasibility of entailment annotation, the principal
investigator examined a pilot of twenty valid entailments, picked at random
from the RTE corpus. These entailments were annotated according to a
provisional scheme that includes eight semantic categories. Six of the
categories that were identified are simple relations like (16.1-4), and two
categories are of a higher-level nature, similar to (16.5). Of the 87 relations
identified in the 20 entailments, about 90% (78 relations) were of the six
simple categories. This suggests the feasibility of an annotation scheme using
these eight categories, as well as the feasibility of their automatic
acquisition and utilization for improving the recognition of entailments. In a
subsequent study (Van Strien 2009), an MA student of the principal investigator,
developed an XML scheme for annotating common relations of restrictive
modification and apposition in the RTE corpus. Two linguist
annotators used this scheme for annotating 200 sentences from the RTE corpus,
uncovering these two relations as operational in 55% of the valid entailments
and achieving 90% cross-annotator agreement results.
Project 5: PhD/Postdoc project in Computational Linguistics – Automatic
acquisition of textual entailments using semantic corpora
Remark: All cross-references
below refer to the general program: http://www.phil.uu.nl/~yoad/vici/ViciProgram-Winter.pdf
As
experience in natural language processing has shown, general machine learning
algorithms are useful for acquiring effective linguistic generalizations from
annotated corpora about language use (Daelemans and van den Bosch 2005, Lappin
and Shieber 2007). In this project, generalizations about acquisition of
entailment from linguistic resources will be obtained from the annotated corpus
(project 4) by general machine learning techniques.
Research
goal:
Use
annotated entailments for developing machine learning algorithms that recognize
unseen entailments on the basis of lexical/syntactic tools and
resources.
Methodology:
The PhD/postdoc researcher
will develop a machine learning model that will use the annotated corpus of
project 4. The corpus will be used as a training set for the acquisition and
utilization of tools and resources (e.g. WordNet, Penn Treebank, OpenCyc) for
parsing entailments. The PhD/postdoc researcher will be a specialist in machine
learning techniques in corpus-based natural language processing, whose
expertise will complement those of the principal investigator.
Synchronization
between projects 4 and 5:
Projects 4 and 5 are closely
connected. The semantic annotation scheme developed in project 4 will inform
the computational model of project 5, and using the semantic annotations on the
resources will affect the definition of the annotation scheme in project 4.