School of Computing
Research Postgraduate Seminar Series
All seminars will take place in L221 at
4pm.
Abstracts:
Automatic
Extraction of Large-Scale, Multilingual Lexical Resources
Comprehensive lexicons are crucial for wide-coverage parsers and
machine translation engines using modern syntactic theories such as
Lexical-Functional Grammar (LFG). One
part of the lexicon is the description of the subcategorisation
requirements for all predicates -
that is, the arguments that the predicate must have in order to form a
grammatical construction.
Manually constructing a lexicon is time-consuming, error-prone and
expensive and has to be done anew for
every language. My research focuses on automatically and efficiently
building large-scale,
high-quality lexical resources for multiple languages. I will present a
methodology for extracting
subcategorisation information from the Penn-II and Penn-III Treebanks
which have been automatically annotated with LFG
f-structures. My approach allows us to control the level of detail in
the frames: for example
whether particles and prepositions are specified. It differentiates
between active and passive frames and
fully reflects long distance dependencies in the source data
structures. The extracted frames can be filtered to
optimise coverage or quality. In contrast to many other approaches,
frames are learned from the data
rather than predefined. I have carried out a large-scale evaluation of
the entire extracted lexicon against
the COMLEX resource. To my knowledge, this is the largest evaluation of
subcategorisation frames for English. The extraction technique for
English has been successfully migrated to Spanish, German and Chinese
treebanks
despite typological differences and variations in treebank encoding.
Creation and
evaluation of a plurilingual dictionary tool
In my research I am developing plurilingual language learning
software for French, Italian and Spanish. The plurilingual learning
approach exploits learners' knowledge of similar
languages and aims to teach the languages contrastively.
One problem of existing plurilingual learning materials is the limited
number of available texts to work with. Therefore I decided to develop
a plurilingual dictionary tool which
allows the learner to enter any text in one of these languages.
Currently the tool is a stand-alone dictionary
but later it will be integrated into different language learning
activities.
In my talk I will first give an overview of the features of the
dictionary tool. Then I will describe the technical side of development
focusing on how data storage and processing is done
with XML, Flash and PHP. Finally I will present some evaluation results.
If you want to test the dictionary tool beforehand just point your
browser to
http://www.computing.dcu.ie/~tkoller/phd/dict/dict.htm
Modelling
the
Liveweight
of Irish Dairy Cows
Many researchers have
modelled
cows’ liveweight from calving to maturity using growth curves, while
others
have modelled liveweight using body measurements. However, the
objective of
this study is to derive an equation which is biologically interpretable
that
will model the liveweight of Irish dairy cows over a lactation period. The dataset
consists of liveweight recordings on a weekly or monthly
basis throughout 12,496 lactations of spring calving cows, from 83
herds. As
the data used in this study is time series data it was decided to look
at
some
time series techniques initially. Splines were also examined to
indicate how
many dimensions were necessary to fit the dataset involved in this
study. As the liveweight curve is similar
to an inverted milk yield curve the models that were used to predict
milk
yield, were also tested. Ultimately liveweight changes between two
calvings
were modelled as a function of age, lactation and pregnancy. As
multicollinearity
was a severe problem with this function, the variance inflation factor
was
examined to find out which variables contributed to it and principal
component
analysis was carried out on the variables responsible for severe
multicollinearity. The new model has a 
of 0.76, the effect
of multicollinearity is weak and the residuals are normal,
homoskedastic and
independent. This new model therefore provides an acceptable level of
accuracy
in representing the shape of the liveweight curve for Irish dairy cows.
A new identity based two party key agreement
This is in the area of identity based cryptography. Those of you who
were at my transfer talk may remember that identity based cryptography
is a relatively new branch of cryptography in which one may transform a
public identifier into a PKI style public key using a well known
algorithm. There are currently two different public key generation
algorithms for
transforming an identifier into a public key. The first, that was
developed by Boneh and Franklin, and a new varient that was developed
by Sakai
and Kasahara. We will look at the differences between these algorithms
and we use this second as the basis for an authenticated key agreement.
A two party key agreement is simply a series of steps needed to be
performed by two entities so that they can agree a shared secret over
an insecure channel (such as the internet). This secret is usually used
to as the
key for a symmetric encryption algorithm and so may be quite short (160
bits). The aim is that any other entity that has the ability to see all
of the
messages that pass between the two entities can still not calculate the
shared secret.
Wind Energy - An Overview
Click here
for details.
Boosting parser
performance in specific domains
Parsing is a well researched area in natural language processing. To
date many high performance probabilistic parsers have been developed.
These parsers are trained on tree bank resources and suffer a drop in
performance
when confronted with text types different to the training data. Through
a number of experiments I have shown that the parsers suffer an
even greater drop
in performance when tested on particular sentence types, in this case
questions. I will present some of my work on improving parser
performance on direct
questions, and an overview of a method to semi automatically create
training resources to boost parser performance in particular areas
using direct questions
as an example.
Space-requirement
Implications for Heterogeneous Traffic
Manoeuvres
A
prototype micro-simulation model is presented for
heterogeneous motorised traffic in an urban context. The
heterogeneous mix traffic consists of short
(single-unit length) vehicles and long (double-unit length)
vehicles in
the first instance. Vehicle manoeuvrability at an urban single-lane
un-signalised intersection is considered for different traffic
distribution and
a range of arrival rates, using minimum acceptable-space rule. For long
vehicles, this implies occupation of
multiple road cells.
The impact on
throughput (the number of vehicles, which navigate through the
intersection in
a given time) and capacity, (the number of vehicles passing from an
entrance
road on to the intersection per unit time) for the geometry are
considered. The impact of overall throughput and capacity in a
TWSC
(Two-way Stop Control) intersection is found to depend on arrival rate
of long
vehicles as well as the arrival rates of major-roads. The occupation,
by a
single vehicle, of more than one cell, not only affects these
quantities directly, but also controls off-arterial flow.
Modern Foreign Languages and ICT in the
Primary School Environment
My current research involves the development of an Intelligent
Computer-Assisted Language Learning tool to help children in the
primary school environment learn German. This presentation sketches the
framework around my
research, looking at Modern Foreign Languages (MFL) in primary schools
in Europe and Ireland, as well as the Irish Department of Education and
Science’s
Information and Communications Technology (ICT) drive.
After a tremulous start, MFLs have taken root in primary school
education in Europe. Comparatively speaking, Ireland’s current Modern
Languages in Primary Schools Initiative is in its infancy, and
consideration must be given
to the lessons that can be learned from our European counterparts.
The integration of ICT into all sectors of education has been a
Department of Education and Science priority since 1997. Current ICT
policies complement the basic methodologies within existing language
curricula theory, making a
successful melding of ICT with languages possible.
Proposed Estimation Method for Financial
Auditing
Auditors often sample account
balances to
determine the accuracy of the published financial statements. The auditor typically tests the hypothesis
that the total error amount in the population exceeds a certain
critical level
called the materiality amount. The test
is usually conducted by constructing an upper confidence bound for the
total
error amount. If this upper bound equals
or exceeds the materiality amount, further investigation of the
accounts is
warranted. Early researchers first
attempted to construct an upper confidence bound for the total error
amount in
an auditing population using classical procedures.
These methods proved unsuccessful, however, because of
the
skewed nature of the population and the rare incidence of occurrence of
the
study variable; the confidence intervals for the means and totals were
found to
be unreliable, i.e. the coverages were below nominal levels.
The
Stringer
(1963) method is a non-classical heuristic procedure, based on the
Poisson
distribution, for computing an upper bound on the total error in a
population. All simulations done by
various researchers show the Stringer bound to give coverage above
nominal
levels. Furthermore all known
simulations studies find the bound to be too conservative with its
value being
far in excess of the true population error amount.
Despite its conservativeness, the Stringer bound is perhaps the
best known in auditing and has been used extensively in the literature
as a
standard of comparison for other methods.
In this talk I will show how a modification to
the Stringer method results in a considerably reduction in its
conservatism.
Computer Poker and
Opponent Modelling
Poker is
the classic game of asymmetric information where the players each have
different parts of the relevant information. This adds an extra
dimension to
the game not present in games of perfect information; it is important
to use
previous decisions by opponents to deduce the hidden information.
Finding
classic equilibria solutions (Minimax/Nash) for the popular variants is
a
non-trivial problem. Of more practical interest is modelling the play
of
typical opponents and exploiting this model to increase the overall
expected
return. The techniques of re-inforcement learning would seem to be
natural here
as there exists a clear reward signal (i.e. chips won/lost). Good
feature
selection will be vital, as it so often is.
Automated Tutoring for a Database Skills Training
Environment
Universities are increasingly offering courses online. Feedback,
assessment and guidance are important features of this online
courseware. Together, in the absence of a human tutor, they aid the
student in the learning process.
I will present a programming training environment for a database
course. It aims to offer a substitute for classroom based learning by
providing synchronous automated feedback to the student, along with
guidance based on a personalised assessment. The automated tutoring
system should promote procedural knowledge acquisition and skills
training.
Analysis and Synthesis of Speech based on
Nonlinear Dimensionality Reduction
Many problems in pattern recognition begin with the dimensionality
reduction of raw high dimensional signals, for example speech waveforms
or images of faces. Classical forms of dimensionality reduction, such
as PCA and MDS, are limited by the assumption that the data lies in a
linear subspace. Recently a number of nonlinear dimensionality
reduction techniques have been proposed, including locally linear
embedding (LLE) and isometric feature mapping (ISOMAP), which overcome
this limitation.
My current research involves investigating the application of these
nonlinear dimensionality reduction techniques to speech. If the
acoustic variability of a speech data set can be described by a small
number of features then we can view the data as lying on a low
dimensional manifold in the high dimensional space of speech waveforms.
In this talk I will describe the LLE algorithm and how it can be
applied to discover low dimensional embeddings of speech data. I shall
also discuss our approach to learning a nonlinear mapping from the low
dimensional embedding space to the high dimensional ‘speech space’ with
a view to speech synthesis and speaker transformation.
Genomic Information Retrieval Using Links
Nowadays a significant amount of biologists' work is spent outside of
the wet labs searching genomic databases. Such searches can be tedious
and involve navigation through a network of heterogeneous data,
following links in a similar way to that in which we browse the
internet. The goal of this project is to use the links to enable search
capabilities over heterogeneous genomic databases, much as Google does
for the web. This approach will reduce the amount of navigation
necessary and deliver a response that integrates diverse genomic data.
The talk will describe the heterogeneous genomic database environment,
the different types of data involved and the different types of links.
Preliminary results on experiments using MeSH (Medical Subject
Headings) data will be presented, along with future research directions.
Terabyte Search: Large Scale Web Search
Experiments for TREC 2004
As the size of the web increases, the task of developing a
effective search engine to deal with these large amounts of documents
becomes a major task. The presence of a new Terabyte track in TREC 2004
is just one example of how important large-scale retrieval has become.
In the seminar I will talk about my work since the development of a
large-scale search engine. This will include a description of my
group's participation in TREC, and subsequent experiments and
improvements made to the engine. I will discuss my current research
into the modification of the search engines architecture as well as the
evaluation of alternative indexes.
Ontology-based Adaptive Content Navigation
Ontologies are a key component of the Semantic Web initiative. Ontology
frameworks represent a method for modelling knowledge. I aim to show
how ontologies can be used in automated course creation and how
adaptivity can be exploited by using an ontological knowledge base. I
also hope to demonstrate how a learner can gain knowledge by browsing
the ontology in a pedagogically sound manner, using learning resources
associated with concepts within the ontology.
Imitation Learning in Interactive Computer
Games
Despite a history of games-based research, academia has generally
regarded commercial games as a distraction from the serious business of
AI, rather than as an opportunity to leverage this existing domain to
the advancement of our knowledge. Similarly, the computer game industry
still relies on techniques that were developed several decades ago, and
has shown little interest in adopting more progressive academic
approaches. In recent times, however, these attitudes have changed, as
each side begins to recognise the potential offered by the other;
under- and post-graduate games development courses are increasingly
common, while the industry itself is turning slowly but surely towards
the more modern AI fields of machine learning and pattern recognition.
One area which has not yet received much attention is imitation
learning, a subdiscipline of pattern recognition which seeks to
expedite the learning process by exploiting data harvested from
demonstrations of a given task. While substantial work has been done in
developing imitation techniques for humanoid robots, there has been
comparatively little exploration of the challenges posed by interactive
computer games. Given that such games generally encode behaviours which
are far more complex and interesting than simple limb movement, that
they often provide inbuilt facilities for recording human play, that
the generation and collection of data is therefore far easier than in
robotics, and that many games have vast pre-existing libraries of these
recorded demonstrations, it is fair to say that computer games
represent an extremely fertile domain for imitation learning research.
This talk will present an overview of imitation learning, including a
breakdown of in-game behaviour based closely on an existing and
widely-accepted psychological model. An API based around Quake 2, a
well-known commercial game of the so-called "first person shooter"
genre, will also be presented. This API, known as the Quake Agent
Simulation Environment, is currently in joint development along with
Blekinge Institute, Sweden and Kyushu University of Japan. QASE is
capable of parsing data recorded during human gameplay, of providing an
interface between MatLab and the Quake 2 game server, and of realising
artificial agents which reproduce observed human behaviours. Some
in-game examples of such agents will also be shown.
Recent breakthroughs
against SHA-1
Last week, the cryptographic "guru" Bruce Schneier announced
a new and effective series of attacks against the Secure Hash Algorithm
by a research team based in China and the US. In this talk, we give a
basic overview of the Secure Hash Algorithm, sketch the (as yet
limited) details of the attack, and speculate on the implications for
the computing industry at large.
Hybrid SMT: Robust Sub-Sentential Alignment
of
Phrase-Structure Trees
Data-Oriented Translation (DOT), based on Data-Oriented Parsing is a
language-independent machine translation engine which exploits parsed,
aligned bitexts to produce very high-quality translations. However,
data acquisition constitutes a serious bottleneck as DOT requires
parsed sentences aligned at both sentential and sub-structural levels.
Manual sub-structural alignment is a time consuming and error prone
task, requiring knowledge of both the source and target languages as
well as how they are related.
My research focuses on the automation of this aligment process which is
essential in order to carry out the large-scale translation experiments
necessary to assess the full potential of DOT.
Long-Term Memory in
the Irish Market (ISEQ): Additional Insight from Wavelet Transforms.
<>Researchers
have
used many different methods to detect the possibility of long-term
dependence in stock market returns and, generally, there is mixed
evidence for the presence of long memory in these data. Here, three
different tests, (namely Rescaled Range (R/S), its modified
form, and (GPH), in addition to a
new approach using the discrete wavelet transform, (DWT), have been applied
to the daily returns of five Irish Stock Exchange (ISEQ) indices. These
methods have also been applied to the volatility measures (namely
absolute and squared returns). The aim is to investigate the existence
of long-term memory properties. The indices are Overall,
Financial, General, Small Cap and ITEQ and the results of these
approaches show that there is no evidence of long-range dependence in
the returns themselves, but there is strong evidence for such
dependence in the squared and absolute returns. Moreover, the discrete
wavelet transform (DWT) has the additional
advantage of providing an in-depth view of the data sets and this gives
us a real indication of structure in long memory effects e.g. giving
clear picture of the movements in the series.
Monte Carlo and Cellular Automata methods for
Simulation of Drug Dissolution
The objective of this investigation is to use Direct Monte
Carlo
techniques in simulating drug delivery from compacts of complex
composition, taking into consideration the special features of the
dissolution in vitro environment. This research focuses on
simulating
a binary system, consisting of poorly-soluble drug, dispersed in a
matrix of highly-soluble acid excipient. At dissolution, the acid
excipient develops certain mechanisms, based on local pH modifications
of the medium, which strongly influence drug release. Our model
directly accounts for such effects as local interactions of the
dissolving components, development of wall-roughness at the
solid-liquid interface, moving concentration boundary layer and mass
transport by advection. Results qualitalively agree with experimental
data and have demonstrated that when modelling dissolution in vitro,
special attention must be paid to including the particular conditions
of the dissolution environment.
Computer-Assisted Language Learning
(CALL) for Dyslexic Learners
Dyslexia is a Specific
Learning
Difficulty (SLD). It is a deficit
in the processing of phonological information and manifests itself as a
difficulty in reading, writing and spelling. Approximately
10% of the population have dyslexia, with 4% of the population being
severely
dyslexic.
My presentation provides a background to
dyslexia and special education issues in Ireland and outlines some of
the
Information and Communication Technology (ICT) tools that are
beneficial to
dyslexic people.
I will then discuss my
research into the
aspects of CALL that cater to dyslexic needs, with reference to CALL
courseware
for dyslexic teenagers that I am developing.
Poitin: Distilling Theorems from Programs
In this research work, the unfold/fold based transformation technique
distillation, which is an extension of the supercompilation technique
is used to prove inductive theorems. Generalisation is performed for
the termination of both the supercompilation and distillation
techniques, but less generalisation is performed for distillation, thus
allowing more theorems to be proved. This more powerful distillation
technique can prove theorems fully automatically which would otherwise
require intermediate lemmas, and can therefore prove a vast range of
theorems which cause problems for existing theorem provers.
Web Service Technologies and Real World Applications
Web Services are a collection of technologies that realise
heterogeneous interoperability by utilising an open and standardised
set of protocols for data exchange. Although there is much talk about
Web Services and their associated technologies there are few examples
of Web Services in action in industry. In this talk I will attempt to
address this issue by providing some interesting examples of real world
applications of Web Services. I will also motivate the need for more
complex Web Services infrastructures to support enterprise
applications.
Some useful background resources related to the talk are available at...
http://en.wikipedia.org/wiki/Web_services
http://en.wikipedia.org/wiki/Service-oriented_architecture
http://en.wikipedia.org/wiki/BPEL
TransBooster: piece-wise machine translation
While state-of-the-art machine translation systems seem to deal well
with short sentences and phrases, introducing longer sentences often
leads to errors. The purpose of the current project (TransBooster) is
to break down sentences and then recompose their translations in such a
way that critical syntactic context is not omitted, but the chunks that
are submitted for translation are of manageable length for the MT
system. Issues involve finding the best substitution schema for
replacing complex syntactic elements with dummy variables, singling out
the syntactic elements that may be omitted from translation and
identifying which parts of the MT output correspond to which chunks of
the input. We have been evaluating TransBooster in conjunction with the
LogoMedia MT engine.
Project Management
Failure - Does a practical solution exist ?
Industry statistics suggest that there is a high incidence of project
management failure globally with an associated cost of failure that is
equally high but largely unrecognised or ignored by organisations. This
study investigates if it is possible to define a solution which can
reduce the incidence of project management failure for an organisation
and is also relatively effortless in its application.The solution
proposed consists of two components -
- The project management methodology type
- The training approach
Both of these components have been implemented in a large
multi-national test organisation and their influence on the outcome of
a test group of projects will be measured over a 3-9 month period. It
is hoped that the outcome will be positive for all or some of these
test group of projects and that some correlation will be established
between the outcome and the project management methodology and training
approach applied.
Transbooster: boosting
the performance of existing MT by complex sentence reduction.
This presentation is an extension of last week's seminar on
Transbooster. The goal of Transbooster is to improve the quality of
current Machine Translation output, not by proposing redevelopment from
scratch, but by building on the strengths of existing MT engines while
trying to correct their common shortcomings. The motivation for the
project lies in the fact that many existing wide-coverage translation
systems handle simple or short sentences better than linguistically
complex ones. The input to the algorithm that we have developed to
reduce the complexity for the client Machine Translation system, will
eventually have to be produced by a parser if previously unseen data
are to be processed. In this presentation, I will give an overview of
parsing techniques that could be used and sum up the
strenghts/weaknesses of our approach.
Automatic Summarisation
of Digitial Visoe based on Analysis of Musical Score
There are thousands of movies available to the viewing public, but the
question is how do we choose among them when deciding what to watch? Do
we yield to the marketing strategies of the studios, listen to the
critics’ reviews, or follow viewer recommendations? In our opinion, it
would be better to be able to access a summary of a film based what
happens in it and how it’s intended to
make you feel, and then we can make up our own minds. To that end,
we're
working on an approach to automatic summarisation of movies, based
primarily
on an analysis of its musical score as an indicator of content -
automatically
analysing movie content, with a view to providing content description.
Level-based Indexing
for Optimising XML Queries
Many of the problems with native XML databases relate to query
performance and subsequently, it can be difficult to convince
traditional database users of the benefits of using semi- or
unstructured databases. In particular, the ongoing development of the
XQuery language requires that performance related issues are resolved.
Presently, there still lacks an index structure providing efficient
support for both navigational and structural queries and the
traditional data-centric and content queries. This paper presents an
extended index structure based on the preorder traversal rank and the
level (or depth) rank of each node in a document tree. The extended
index fully supports the navigation of all XPath axes while efficiently
supporting data-centric queries. The ability to start path traversals
from arbitrary nodes in a document tree also enables the extended index
to support the evaluation of path traversals embedded in XQuery
expressions. Furthermore, an encoding technique for this extended index
structure is presented, where properties of a level ranking may be
exploited to provide efficient and optimised path traversals and in
certain cases, optimal solutions to path traversals.
Consistency management mechanisms in Internet content
delivery
Facilitating the exchange of data between organisations located at
geographically distributed sites was the main motivation for the
development of the Internet. Today, as a consequence of the popularity
of the world-wide-web, and as the E-business revolution gathers
momentum, new methods for the distribution of content from origin
servers to end-users are required. In the coming years, data will be
highly replicated, and processed deep inside the network, introducing a
whole host of consistency management issues. In this presentation, a
variety of content delivery architectures, from the traditional
client-server model, to the futuristic on-demand/edge computing
platform will be illustrated. The issue of consistency management will
be explored, followed by a description of current techniques. Details
of my own work thus far in the area will be given, together with an
outline of future research direction.
A Metadata Repository for Facilitating
Process Composition
Service oriented architectures provide a modern paradigm for web
services allowing seamless interoperation among network applications
and supporting a flexible approach to building large complex
information systems. A number of industrial standards have emerged to
exploit this paradigm with the development of the J2EE and .NET
infrastructure platforms, communication protocol SOAP, description
language WSDL and orchestration languages BPEL, XLANG and WSCI. At the
same time the Semantic Web enables automated use of ontologies to
describe web services in a machine interpretable language. In previous
work, we presented a Peer-to-Peer infrastructure for large scale data
integration. In this presentation a service infrastructure to provide
an e-business layer exploiting current web service technologies will be
displayed. In this context, we present a distributed service
repository over a super-peer network facilitating process composition.
To provide tangible reliability for services and processes, our
framework is introduced to support the e-business layer.
Finding
New News: Novelty Detection in Broadcast
News
The automatic
detection of novelty, or newness, as part of an information retrieval
system
would greatly improve a searcher's experience by presenting
"documents" in order of how much extra information they add to what
is already known instead of how similar they are to a user's query.
This would
be particularly useful in applications such as searching broadcast
news. We
present a novelty detection system evaluated on the AQUAINT text
collection as
part of our TREC 2004 Novelty Track experiments. We also discuss how we
are
extending the text-only approach to novelty detection to also include
input
from video analysis.
Side Channel Analysis of Pairing-Based
Cryptography
Pairing-based cryptography (PBC) is currently one of the most popular
topics in cryptography. Pairings such as the Weil, Tate or Eta, have
the essential property of bilinearity, which is the core operation in
the construction of a number of ECC based protocols. With the
development of efficient algorithms to compute Pairings, makes them and
their respective protocols, perfect candidates for usage on smartcard
systems.
Implementation of these primitives on smartcards cannot pass without
the assessment of vulnerability of Pairings to SCA, since it is one of
the most potent forms of cryptanalysis of smartcard systems.
In this talk an overview of Pairing-based cryptography will be
presented and a first look at the application of SCA to PBC will be
discussed.
Probabilistic detection of ungrammatical
sentences
In theory, parsing directly yields a grammaticality judgement. A
sentence is grammatical, if and only if, it can be parsed with a
grammar of the language in question. Unfortunately, writing grammars
for natural languages is very difficult and nobody has actually
succeeded in providing a complete grammar yet. Hand-written grammars
usually achieve deficient coverage, i.e. a parser will reject sentences
that are judged grammatical by most native speakers. Data-driven
methods, on the other hand, often generalise too much and produce
grammars that render nearly any sequence of words grammatical. The
latter
grammars are still useful in applications that assume valid input and
just need an analysis of the input. Probabilistic parsers use
statistical information attached to the grammar in order to select a
plausible parse
among the set of possible parses. Over-generalisation and probabilistic
selection together result in high robustness to errors and broad
coverage of language. These are very desirable properties in many
applications. We propose to exploit the output of existing
probabilistic parsers to judge grammaticality. I will explain why a
simple threshold method cannot work and outline our idea to tackle this
problem. Prelimniary experiments show that the main prerequisite of our
approach might actually hold.
Distillation: Higher Order Transformation
for
Higher Order Programs
15 May 2005