A Tune-Based Account of Turkish Information Structure
Umut Özge, January 2003
Languages differ in the means they avail themselves of for the
structural realization of information structure, where available options
are word order, prosody and morphology. Turkish has long been
characterized as predominantly using word order and its variation in
realizing information structure, where certain positions in a sentence are
associated with certain pragmatic functions related to information
structure. Prosody has been proposed to play only a secondary role
interacting with word order. Contrary to this widely established view, the
thesis argues that the notion that sentential positions have pragmatic
functions and word order variation is a syntactic means to realize these
functions can be abandoned, without any loss of explanatory power, in
favor of a tune-based perspective where prosody is the sole structural
determinant of information structure. In this setting word order
variations are argued to be prosodically motivated, in that Turkish
phonology imposes some precedence constraints on intonational contours.
Word order variation then turns out to be just a consequence as opposed to
being a determinant in attaining the right information structure required
by thediscourse context. To substantiate these claims a tune-based
account, based on Steedman's account of English information structure, is
proposed for the structural realization of information structure in
Turkish, whereby information structural units are directly associated with
prosodic phrases intonationally marked in certain ways. Validity of the
account is tried to be established by intonational analysis of recorded
speech data. As for the explanatory value, the information structure
phenomena that has received positional explanation in the relevant
literature, are tried to be captured only in prosodic terms, without
committing to positions, syntactic strategies and such.
Integrating Landmarks into a Computational Model of
Early Cognitive Mapping
Erkan Gurgen, January 2002
This work aims to expand
the computational theory of early cognitive mapping proposed by Yeap
(1988) and Yeap and Jefferies (1999). Yeap and Jefferies developed a
computational theory based on human cognitive mapping studies. Their
theory starts from a 2 1/2-D sketch that is the output of Marr's vision
theory (Marr, 1982). They attempt to answer the question of what is
computed first from 2 1/2-D sketch. They argue that the boundary of the
local space which one is surrounded by is computed first. Yeap and
Jefferies' theory and implementations are limited with early cognitive
mapping excluding landmark knowledge. This work aims to improve the
theory by integrating landmark information into early cognitive mapping.
Landmarks are used for encoding spatial relations within local spaces. A
local matching method by using landmarks is proposed for recognition of
revisited local spaces. Acquisition of more detailed metric information
within local spaces by using landmarks and improved ability in the
recognition of revisited local spaces are also among the
main
improvements obtained by integrating landmarks into Yeap and
Jefferies' cognitive maps.
Keywords: Cognitive Maps, Cognitive
Mapping, Landmarks, Place Recognition
Cognitive and Computational Aspects of Gender
Estimation From Faces
M. Koray Balci, January 2002
The aim of this work is to propose a computationally
feasible and cognitively plausible model for face processing and develop
a system for gender estimation from face images. For this purpose, we
propose a general face processing model that encapsulates all face
specific tasks. The model is inspired by the findings from cognitive
studies. Next, we implement the core of the whole model which uses
Principal Component Analysis (PCA) procedure and develop a classifier
for gender estimation. As classifier, we implement a Multi Layer
Perceptron (MLP). The MLP is further pruned for observing the minimal
input set necessary for the task. By our pruning approach we end up with
a robust and efficient classifier. We confirm the importance of
higher-eigenvalued eigenvectors and also show that only a small subset
of them are sufficient for gender estimation. We test our approach in
two different face databases, one of which is the largest face database
up to date. Until this study, PCA approach has not been tested on a
large database such as this one.
Keywords: Face Processing, Gender Estimation, Principal
Component Analysis (PCA), Multi Layer Perceptrons
A Minimalistic Approach To
Russian-English-Turkish
Multilingualism
Oya Ozagac, April 2002
The empirical question which
is the focus of present research is: How may the lexicons from different
languages interact in the course of one syntactical derivation,
resulting in code switching phenomena? We develop the following
hypothesis concerning code switching: The units of intrasentential code
switching are either heads or functional maximal projections. To get
support for this hypothesis, intrasentential code switching instances
from Russian-English-Turkish and Dutch-Turkish spoken data are analyzed
within the minimalist framework. In the data analysed, it has been
observed that the data gathered support this hypothesis and that the
Minimalist Program has an explanatory force for bilingual language
processing.
Keywords: bilingualism, multilingualism, code switching,
minimalism, lexicon
Cognitive Aspects of Image Zooming
Gaye Oncul , July 2002
In this work, digital image zooming methods and image
distortion metrics are examined in terms of a quality criterion based on
image appearance. Neurophysiological and psychophysical foundations of
perception driven image distortion metrics are presented. Results of a
comparative study based on both perceptual distortion metrics and
conventional distortion metrics for standard zooming methods are
discussed. Open problems in image quality definition as well as image
enlargement methods are presented and improvements on magnified image
quality are suggested.
2LRL: A Two-Level Multi-Agent Reinforcement Learning
Method With Communication
Guray Erus, August 2002
Learning is a key element
of an "intelligent" computational system. In Multi-agent Systems (MASs),
learning involves acquisition of a cooperative behavior in order to
satisfy the joint goals. Reinforcement Learning (RL) is a promising
unsupervised machine learning technique inspired from the earlier
studies in animal learning. In this thesis, we propose the Two Level
Reinforcement Learning with Communication (2LRL) method, a new RL
technique to provide cooperative action selection in a multi-agent
environment. In 2LRL, the decision mechanism of the agents is divided
into two hierarchical levels, in which the agents learn to select their
target in the first level and to select the action directed to their
target in the second level. The agents communicate their perception to
their neighbors and use the communication information in their
decision-making. We applied 2LRL method in a hunter-prey environment and
observed a satisfactory cooperative behavior.
Keywords: Multi-agent
Learning, Reinforcement Learning, Multi-agent Cooperation,
Communication
A Markov Model for Chorale Harmonization in the Style
of J.S. Bach
Kaan Biyikoglu, September 2002
Developing
algorithms for modeling musical intuition is beneficial because such an
endeavor enables the testing of our hypothesis about musical intuition
on empirical grounds. The present study aims to conduct an empirical
investigation of the syntax of harmonic progressions found in the
chorales of Johann Sebastian Bach, and to present a model for chorale
harmonization. First, about 170 chorales of Bach are analyzed to reveal
their harmonic progression by using a methodology similar to the
time-span reduction as given by Lerdahl and Jackendoff (1983). By virtue
of the same analysis, the chorales are segmented according to their
harmonies, and these segments are classified according to their chordal
functions (i.e., major-triad, half-diminished-7th, dominant-7th, etc.).
The obtained harmonic progressions are used for the training of a Markov
model and this model is used to generate suitable harmonies for novel
melodies. After the harmony is calculated for a novel melody, a pattern
matching module is called for the generation of a four-part
harmonization by using the previously obtained segments from Bach
chorales.
VP Shells and Argument Structure in Turkish
Cem Keskin, September 2002
This thesis deals with Turkish predicates within the
framework of the Minimalist Program, which is the current phase of the
Generative Grammar tradition based primarily on Chomsky’s work (e.g.
Chomsky, 1957-1995). It aims to discover what kind of structures the
Larsonian vp shell approach yields in the analysis of Turkish
predicates. The vp shell structure was introduced by Larson (1988) for
the analysis of double object constructions. The proposal was then
adopted, modified and expanded by various authors (e.g., Chomsky, 1995;
Hale and Keyser, 1993; Radford, 1997). The derivation of a vp shell
structure is now standardly assumed to involve the adjunction of a verb
(V) to a light verb (v), thereby forming a complex verb of some sort in
the form V-v (see Chomsky, 1995 and Radford, 1997). This thesis claims
that the direct application of this derivation to Turkish involves a
Lexical Insertion Principle violation and proposes a variant for the vp
shell derivation that involves replacing the light verb with a voice
feature. Then, using this alternative derivation, it undertakes the
analysis of Turkish predicates, also demonstrating problematic areas and
offering solutions to these.
Keywords: vp shells, argument structure, light verbs,
voice
Chord Priming Beyond Association
Nart Bedin Atalay, September 2002
Perception of harmony is one of the main topics for the
study of music cognition. One aspect of this study is related with the
perception of chordal progressions. Chord progressions in Western tonal
music are regular and they are organized by rules. The current
explanation of the psychology of harmony emphasizes the knowledge about
hieararchical organization between elements. This knowledge is encoded
as a neural network. By this way, rules are not encoded explicitly but
they emerge as a result of micro-computations within the network.
According to this account, perception is shaped by spreading activation
in the network. In this study, the question of whether associationist
description of harmonic knowledge is sufficient to explain the
perception of harmony is investigated. The main investigation is
revolved around the question of whether the knowledge of neapolitan
sixth progression has been acquired by individuals. The neapolitan is
important because it shows regularities which cannot be captured by
hieararchical organization between elements. If the knowledge of
neapolitan progression has been acquired by individuals then
explanations based on associationist paradigm for the psychology of
harmony should be revised. Results of psychological experiments showed
that subjects were aware of neapolitan progressions in sequences. The
current theory of harmony should be revised.
Word Order Variations in Turkish: Evidence From Binding
and Scope
Ceyhan Temurcu, November 2001
This study aims to contribute to the efforts of
explaining word order variations observed in Turkish, by granting a role
to the hierarchical structure for certain word order alternations, and
by admitting discourse-driven 'extra-syntactic' mechanisms for others.
It will propose a tentative extension to the framework of the Minimalist
Program (as it appears in Chomsky, 1995), by adopting its basic tenets
and introducing the information structure (IS) as a representation that
feeds the phonological component. IS differences will be maintained to
be responsible from what is frequently referred to in the literature as
'optional' or 'stylistic' movements. Extraposition and contrastive focus
fronting will be argued to be IS-driven movements in Turkish. It will be
maintained that the SOV-OSV alternation in the surface forms with
pre-verbal focus signals a difference in the hierarchical structure,
whereas other word orders in transitives can best be viewed as sharing
the phrase structures, hence being discursive variants of either the SOV
or the OSV construction. OSV will be analyzed as involving a grammatical
movement of the object to an IP- external Spec position. In addition,
OSV will be proposed to represent the inverse voice in Turkish, The
phrase structure in Turkish will be analyzed from a Minimalist
perspective, and structural analyses will be given for what is referred
to as 'incorporation' of objects and subjects in Turkish.
Keywords: Turkish, Word Order, Phrase Structure,
Information Structure, Phonological Component, Extraposition,
Contrastive Fronting, Topic-driven Movement, Case Checking, Object
Shift, Incorporation, VP-shells, Scrambling, Inverse Voice, Basic Word
Order.
Semantic Plausibility and Category Effects on Unbounded
Dependency Processing in Turkish Relative Clauses
Sukru Baris Demiral, September 2001
Unbounded dependencies represent special types of
constructions to understand initial preferences of human language
parser. How the language parser relates unbounded items is the main
concern of both psycholinguists and computational linguists. In order to
understand syntactic and semantic constraints in initial parsing
preferences, varieties of sentence types for different languages have
been constructed. Turkish, with its characteristic syntactic and
morpho-syntactic structure, enables us to observe the behavior of the
parser on different categories and constituents. This thesis aims to
find out syntactic and semantic constraints used by the parser to form
unbounded dependencies as it proceeds on Turkish relative clauses. In
order to observe initial parsing preferences of the parser, a self-paced
reading experiment was conducted. In this experiment, some parts of the
sentences, which were difficult to parse, caused long reading times when
they were compared with the reading times of the base condition. It has
been observed that syntactic constraints, lexical category information
and semantic plausibility affected parsing. The verb of the relative
clause assigned its thematic expectancy on the following words
immediately. Our findings indicate that verb-specific information is
processed in a very fast manner such that thematic features generated by
the verb direct the parsing process. The results of the experiment are
in accord with the constraint-based models of parsing.
Keywords: Unbounded Dependency, Parsing Decisions,
Syntax, Category Information, Semantic Plausibility, Turkish Relative
Clauses, Self-Paced Reading Experiment, Thematic Expectancy,
Constraint-Based Accounts
Refining The Representational Basis of The
Construction-Integration Model of Text Comprehension with Syntactic
Cues
Evren Kapusuz, September 2001
In this thesis, within the framework provided by the
Construction-Integration (CI) model of comprehension, we investigate the
topicality of referents in discourse in terms of the activations of the
memory elements corresponding to those referents. We discovered the
inherent deficiencies of the CI model in displaying the linguistic
observations about how different syntactic cues relate to sentence-level
topic. This thesis proposes enchantments to the CI model so that it
incorporates and makes use of the two source of grammatical information
prevalent for topicality in sentence level: the grammatical role of the
participants and the causal organization of sentences. Computational
tests are carried out to compare the original and modified CI models.
The comparisons have shown that the modified CI model is more successful
at identifying the topic of single sentences. We have also implemented
two aspects of discourse-level topic, i.e. topic shifts and referential
continuity, in the modified CI model.
Keywords: Cognitive Modelling, Construction-Integration
Model, Sentence-Level Topic, Topic Continuity, Centering Theory
Developing A Feature-Based Edge Detector Using
Artificial Neural Networks
Güven Burç Arpat, March 2001
In vision science, edges are considered amongst the
most important features of low-level image processing. They can be used
for a variety of subsequent image processing steps, such as object
recognition and motion analysis. Yet, the concept of edge detection is
not limited to optical images only. Non-optical images that are mostly
used in measurement sciences also require use of edge detection
techniques to be processed/interpreted with high precision and/or speed.
Yet, the edges in non-optical images are hard to define due to their
non-luminance/intensity based nature.
The classical way to obtain
edges in vision science is to process the image using image derivative
operators like in the famous Canny method. Unfortunately, almost all
popular edge detection algorithms make use of variations of this
approach and are developed such that they initially can only be used on
optical images/intensity maps. To overcome this difficulty of
applicability to non-optical images and to be able to automatically
calibrate the edge detector to the image type/edge type of interest, a
new technique making use of artificial neural networks is proposed. The
proposed network uses analogous structures that can be found in human
vision system (analogs to retina, simple cells and hypercolumns) and
utilizes a concept called "training images" to be able to adjust itself
to any image/edge type of interest with an initial help from an
instructor, i.e. using a supervised learning scheme.
To demonstrate
the capabilities of the proposed method, the study also includes samples
of successful applications to luminance edge detection (a standard
optical/intensity image application) with a comparison to Canny edge
operator and rock-type interpretation from seismic measurements (an
earth sciences / measurement sciences application as the non-optical
example).
Keywords: Edge Detection, Image Processing, Human Eye,
Artificial Neural Networks (ANNs), Non-optical Images.
A Computational Analysis of Information Structure in
Turkish
Filiz Yilmaz Bican, January 2001
The purpose of this study is to investigate the
information structure in Turkish both linguistically and
computationally. The thesis includes an empirical study where properties
of information structure units in Turkish, namely topic and focus, are
obtained statistically from a collection of naturally-occurring data.
These results are reflected to an implementation which tests the
analyticity of the results on a control data. The implementation is a
simulation of the knowledge-stores of speakers and a possible part of
future Natural Language Processing applications, as well as a
demonstration of the computability of information structure units.
Keywords: Information Structure, Topic, Focus,
Turkish
Machine Learning and Language Acquisition: A Model of
Child's Learning of Turkish Morphophonology
Yasemin Altun, July 1999
Every normal child who is exposed to linguistic input
in an interactional environment acquires the complex structure of the
language at a very early age, in a very short time and without any
explicit training. This fascinating character of language acquisition
lead many researchers to study on the aspects of language acquisition.
The present study is on the morphological analysis of Turkish and on
learning the morphophonology of Turkish by using the non-monotonic
setting of Inductive Logic Programming, ILP. A model for the acquisition
of morphophonology by children is proposed and is implemented in a
computational system. Assuming that grammar is composed of the lexicon,
the phonological representation and the semantic representation which
are connected to the central structure, i.e. surface grammar, we aim at
generating a part of the central structure, namely the morphophonology,
when phonological, semantic and lexical knowledge are
given.
