Date: Wednesday 8th December 2010, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Pierre Andrews, University of Trento, Department of Information Engineering and Computer Science
Semantics for the Users?
One of the cornerstones of what we now call the 'Web 2.0' is
unconstrained user collaboration and creation of content. Some of the
first sites to allow such features were del.icio.us and Flickr where
users could share resources -- bookmarks and photos respectively -- and
freely annotate them. Both websites allowed the creation of
Folksonomies: social classification of resources created by the
community that have shown to be very important for organising the large
amount of content online, but also for, later on, studying the
collaborative creation of shared vocabularies and taxonomies.
However, in these folksonomies, tags are free-form terms with no
explicit semantic, therefore a number of issues arise from their use,
such as:
- the loss in precision due to the ambiguity of tags -- for example, the
tag ``java'' can refer to the ``Indonesian island'', the ``programming
language'', and a ``beverage''.
- the loss of recall due to the synonymy of terms -- for instance, if
you search for the tag ``travel'', you might be interested by the
results for the tag ``journey''.
These issues are also exacerbated by the use of different forms of the
same word as some users would, for example, use the tag ``running'',
others would use instead ``run'', ``runs'', ``torun'', etc.
The presentation discusses the cases of Flickr.com and del.icio.us and
look at the tagging habits in these folksonomies. The study of two large
datasets show a number of properties of the vocabulary
shared by the users of these folksonomies and identify important
features that have been overlooked by previous studies on
disambiguation and sense extraction from such folksonomies.
Date: Wednesday 1st December 2010, 11:15, CSE102 Computer Science, Univ. of York
Group meeting
Date: Wednesday 24th November 2010, 11:15, CSE202 Computer Science, Univ. of York
Speaker: Peter Hines, University of York, Department of Computer Science
Types in models of meaning (and elsewhere)
This talk begins by describing an ambitious project to produce models of meaning for natural languages that are both distributional and compositional. A general outline of the scheme - built using the tools of categorical logic - is described, and a toy example of how this might work is given. However, even this simple toy example uncovers numerous subtleties in such proposed models, relating to questions of both typing (in the technical sense), and whether deduction in models of logic is necessarily a process that loses information. Exploring these questions immediately leads to some very deep structures related to models of lambda calculus, and taking the process even further we see the unexpected emergence of classic structures from a different field entirely.
Date: Wednesday 17th November 2010, 11:15, CSE102 Computer Science, Univ. of York
No Meeting - Speaker cancelled due to illness
Speaker: Myroslava Dzikovska, University of Edinburgh, Human
Communication Research Centre, School of Informatics
Better tutoring with natural language dialogue
One-on-one tutoring is widely considered to be the most effective form
of instruction. One possible explanation for this is that dialogue with a
tutor keeps students engaged with the material and forces them to explain
their reasoning, resulting in better understanding. Intelligent tutoring
systems aim to deliver improved learning through interaction with a computer
rather than a human, by replicating effective tutoring strategies employed by
human tutors.
I will discuss how tutorial dialogue differs from task-oriented and
information-seeking dialogue, and the challenges it brings to natural
language processing. I will then describe the Beetle2 tutorial dialogue
system being developed at the University of Edinburgh as a testbed for
research in effective dialogue management and NLP techniques for tutoring.
First system evaluation was completed in 2009, and I will discuss its results
and future research directions.
Date: Wednesday 10th November 2010, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Daniel Kudenko, University of York, Department of Computer Science
An introduction to Game Theory
Any agent, whether alive or man-made, physical or virtual, interacts with other agents. This interaction can take the form of cooperation or competition, or both. In order to be able to implement effective agents for such a multi-agent environent, we need to understand these interactions. Game theory offers mathematical tools to represent and reason about interactions. In the seminar I will discuss the issues surrounding the representation of interactions, definitions of optimality, and algorithms that compute solutions to games. I will assume no prior knowledge in game theory.
Date: Wednesday 3rd November 2010, 14:00, CSE082 Computer Science, Univ. of York
Speaker: James Cussens, University of York, Department of Computer Science
Integer linear programming for machine learning
Most machine learning is optimisation of one sort of another. For example, in the Bayesian approach the goal is either to find the most probable statistical model (conditional on data and domain knowledge) or to find the best approximation to a probability distribution over all possible models. I have recently started to use integer linear programming to solve machine learning optimisation problems and like any new convert am keen to evangelise. In a linear program (LP) the goal is to maximise a linear function of some real-valued variables subject to linear constraints on those variables. In integer linear programming (ILP), some, perhaps all, of these variables must take integer values which makes it substantially harder than LP. This seminar will be largely at a tutorial level. The basics of LP and ILP will be presented and then an application of ILP to the problem of learning Bayesian networks will be given. Experiments with the SCIP (http://scip.zib.de/) ILP solver will be presented.
Date: Wednesday 3rd November 2010, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Alan Baddeley, University of York, Department of Psychology
Working memory
Cognitive psychology has relied heavily on ideas from computer science and artificial intelligence for its development of theory. This is the case, even when theories are driven principally by an experimental bottom-up, rather than a theory-driven top-down approach. This will be illustrated in the case of working memory, a theory that was originally driven by simple computer-based analogies, and that has continued to be productive for over thirty years. Implications for the relationship between cognitive psychology and computer science will be briefly discussed.
Date: Wednesday 27th October 2010, 14:00, CSE082 Computer Science, Univ. of York
Speaker: Zoubin Ghahramani, University of Cambridge, Department of Engineering
Nonparametric Bayesian Modelling
Because uncertainty, data, and inference play a fundamental role in the design of systems that learn, probabilistic modelling has become one of the cornerstones of the field of machine learning. Once a probabilistic model is defined, Bayesian statistics (which used to be called "inverse probability") can be used to make inferences and predictions from the model. Bayesian methods also elucidate how probabilities can be used to coherently represent degrees of belief in a rational artificial agent. Bayesian methods work best when they are applied to models that are flexible enough to capture the complexity of real-world data. Recent work on non-parametric Bayesian machine learning provides this flexibility. I will touch upon key developments in the field, including Gaussian processes, Dirichlet processes, and the Indian buffet process (IBP). Focusing on the IBP, I will describe how this can be used in a number of applications such as collaborative filtering, bioinformatics, cognitive modelling, independent components analysis, time series modelling, and causal discovery. Finally, I will outline the main challenges in the field: how to develop new models, new fast inference algorithms, and compelling applications.
Date: Wednesday 27th October 2010, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Mark Stevenson, University of Sheffield, Department of Computer Science
Word Sense Disambiguation in the Biomedical Domain
Like text in other domains, biomedical documents contain a range of terms with more than one possible meaning. These ambiguities form a significant obstacle to the automatic processing of these texts. Previous approaches to resolving this problem have made use of a variety of knowledge sources including the context in which the ambiguous term is used and domain-specific resources (such as UMLS). We compare a range of knowledge sources which have been previously used and introduce a novel one: MeSH terms. The best performance is obtained using linguistic features in combination with MeSH terms. Performance exceeds previously reported results on a standard test set. Our approach is supervised and therefore relies on annotated training examples. A novel approach to automatically acquiring additional training data, based on the relevance feedback technique from Information Retrieval, is presented. Applying this method to generate additional training examples is shown to lead to a further increase in performance.
Date: Wednesday 20th October 2010, 11:15, CSE102 Computer Science, Univ. of York
Speaker: Mark Bartlett, University of York, Department of Computer Science
Bayesian Network Learning for Instruction Cache Analysis
In Real-Time Systems, the temporal properties of a program are as important as the functional properties; for example, there would be no point in a car braking system that took a minute to activate. One important issue in analysing the time that such programs take to execute is to determine which instructions are in the cache when needed, as main memory access times can be orders of magnitude greater. At present, this cache analysis is performed using detailed mathematical models of the program and the code under examination, but this requires perfect knowledge of the system, intelligent human insight and is frequently overly pessimistic. I introduce instead a novel method in which a cache use model is automatically learned from observations of the programming running. This takes the form of a Bayesian network, and I demonstrate how this can be used to obtain accurate estimates of the cache behaviour.
Date: Wednesday 23rd June 2010, 11:15, CS202J Computer Science, Univ. of York
Speaker: Ali Karami, University of York, Department of Computer Science
The Dynamic Temporal Knapsack Problem and Its Solutions
The Dynamic Temporal Knapsack Problem (DTKP) is a new combinatorial optimisation problem. The temporal knapsack problem was first defined by Bartlett et al in 2005 and a set of solvers were introduced. A work by Van Hentenryck and Bent in 2006 introduced a multi-scenario approach based on a set of anticipatory algorithms for solving some combinatorial optimisation problems such as online reservation problems. The DTKP can be considered as a variation of the online reservation problem. This suggests the use of anticipatory algorithms in the DTKP solvers. In This talk we introduce the dynamic temporal knapsack problem and a set of solvers based on multi-scenario approach. Also, the effect of anticipation of the future on the quality of the DTKP solutions is investigated.
Speaker: Paul Giannoros, University of York, Department of Computer Science
SAT Encodings of Grammars
Recent work in the field has already produced two similar encodings of context-free grammars based on the same parsing algorithm. This project shows that a more direct approach to the problem is possible, presenting new SAT encodings inspired by the parses produced by regular and context-free grammars. The encodings presented are evaluated for efficiency and benchmarked against the two prior context-free grammar encodings, with competitive results for the regular grammar encoding. In addition, comparisons of the two prior context-free grammar encodings introduced by Axelsson et al. and Quimper and Walsh are compared and contrasted, and criteria under which they are equivalent are presented.
Date: Wednesday 16th June 2010, 14:00, CS103 Computer Science, Univ. of York
Speaker: Peter Cowling, University of Bradford, School of Informatics
Date: Wednesday 16th June 2010, 11:15, CS202J Computer Science, Univ. of York
Speaker: Edward Hartwell Goose, University of York, Department of Computer Science
Unsupervised Keyword Identification Using Topic Models
Keyword identification is the process of choosing specific words from a corpus which identify the content or the subject of an article. Vast quantities of information are now published online and tagging these articles with keywords has the potential to allow users to easily discover information that interests them. The Discovery Challenge is a competition designed to evaluate methods for creating these tags for the website BibSonomy.org. Another method of organising information is the use of topic models. A topic model organises information into distinct topics which, for example, highlight similarities between two different newspaper articles. The aim of this project is to develop and evaluate a method that utilises topic models to create keywords for the Discovery Challenge as well as a student media website in York, namely, The Yorker. The two evaluations show potential in the use of topic models for keyword identification, but more work needs to be done. The work for The Yorker involved a human evaluation which could be expanded while the work for BibSonomy revealed problems with noise and languages.
Date: Wednesday 9th June 2010, 14:00, CS103 Computer Science, Univ. of York
Speaker: Ross King, University of Wales, Computational Biology Group
Date: Wednesday 2nd June 2010, 11:15, CS202J Computer Science, Univ. of York
Speaker: Simon Hickinbotham, University of York, Department of Computer Science, Non-Standard Computation Group
A WORLD WITHOUT REASONING: FROM ARTIFICIAL CHEMISTRIES TO BEHAVIOUR-BASED ROBOTICS
In the 1980s, Rodney Brooks proposed a computational model in which a combination of perception and action systems combine to give the illusion of cognition. While we do not pretend that this idea can be universally applied, the approch has given rise to a variety of successful robot architectures. A good example of this is the subsumption architecture in which competing behavioural subunits have the opportunity to sieze control of a system should the stuation demand it. A major problem with these architectures is that they are not directly amenable to optimisation, because the connections between processing units tend to be `brittle'. Work on the Plazzmid project here at York has attempted to address these issues, and to make versions of these systems available for evolutionary control. This can be achieved by implementing an artificial chemistry to run the control architecture. The chemistry works on two levels. The first level defines a processing network in which each node is represented by an artificial 'molecule' and each edge shows a reaction between molecules. We derive a level-two representation of the network in which each molecular species is defined by a sequence of symbols that species both the binding affinity between molecules, and the product of the reaction between them.
Date: Wednesday 26th May 2010, 11:15, CS202J Computer Science, Univ. of York
Speaker: Katja Markert, University of Leeds, Department of Computer Science
Automatic Modelling of Lexical Sentiment
The field of sentiment and opinion mining has grown immensely over the last decade, with myriads of applications such as review summarisation, detecting bias in newspaper articles or predicting gobal trends. Most work starts with automatic annotation of word lists for sentiment as the basis. In this talk we discuss the following questions: What are the limitations of current work on lexical sentiment? How can we automatically create large-scale reliable ressources for sentiment mining? What are the limits of concentrating on lexically expressed sentiment, for example in bias detection?
Date: Wednesday 19th May 2010, 11:15, CS103 Computer Science, Univ. of York
Speaker: David Weir, University of Sussex, Department of Informatics
Exploiting the Distributional Hypothesis in Natural Language Engineering
The distributional hypothesis asserts that words that occur in similar contexts tend to have similar meanings. A growing body of research has been concerned with exploiting the connection between language use and meaning, and much of this work has involved measuring the distributional similarity of words based on the extent that they share similar contexts. In this talk I consider a variety of approaches that have been developed to measure distributional similarity, and look more generally at the potential of this approach by considering the following questions. What aspects of the semantics of a word can be characterised using distributional similarity? Is it feasible to use distributional similiarity to compare the meaning, not just of individual words, but of phrases? What kinds of semantic relationships can be captured with distributional semantics, and to what extent can distributional similarity be used to determine ontological relationships such as hyponomy between word senses? Given the availability of large collections of different kinds of text, how can we exploit the fact that the distributional semantics can be automatically derived from any sufficiently large corpus of text?
Date: Wednesday 5th May 2010, 14:00, CS103 Computer Science, Univ. of York
Speaker: Peter Dayan, University College London, Gatsby Computational Neuroscience Unit
Neural Reinforcement Learning: The Good (and maybe the Bad and the Ugly)
Reinforcement learning has become a wide conduit that links ideas and results in computer science, statistics, control theory and economics to psychological data on animal and human decision-making, and the neural basis of choice. I will describe the resulting synthesis, show how and why refined theoretical distinctions such as that between model-free and model-based reinforcement learning have a direct resonance in empirical data, and sketch out areas in which the conduit is beginning to leak.
Date: Wednesday 28th April 2010, 11:15, CS202J Computer Science, Univ. of York
Speaker: Diana McCarthy, University of Sussex, Department of Informatics
Graded Word Meaning Annotations
For over a decade, researchers in word sense disambiguation have been wrestling with how best to represent word senses. Traditionally, a manually produced sense inventory is selected and then a word in context is labelled with the identifier of the best fitting sense description from this inventory. This talk will contrast such traditional gold standard data with datasets that we have created using graded sense assignment, sentence similarity and synonyms in context (lexical substitutes). In this talk I will describe our motivation for moving from the traditional winner-takes-all annotation, the production of the new datasets, and some findings on the relationships between these graded annotations and traditional datasets. I will also give a brief summary of some preliminary experiments evaluating various computational models on some of these datasets.
Date: Wednesday 17th March 2010, 11:15, CS202J Computer Science, Univ. of York
Speaker: Richard Wilson, University of York, Department of Computer Science, Computer Vision and Pattern Recognition Group
Learning with difficult data for Pattern Recognition
Statistical Machine Learning is the fundamental tool which has allowed rapid progress in the fields of computer vision and pattern recognition over the last decade. Large amounts of data and the rise of computing power have made it possible to learn sophisticated models for vision, speech recognition, biometrics and many other problems. However, some types of data are difficult to learn about. In this talk, I will highlight some examples of difficult data we have encountered in the CVPR group, and what we did with them. I will discuss the problem of learning on non-Euclidean manifolds and the exponential map. I will then talk about the problem of learning fields of directional data, and finally an application to discovering object shape.
Date: Wednesday 10th March 2010, 11:15, CS202J Computer Science, Univ. of York
Group meeting
Date: Wednesday 3th March 2010, 11:15, CS202J Computer Science, Univ. of York
Speaker: George Tsoulas, University of York, Department of Language and Linguistic Science
Types of Semantic Composition
In this talk I present some data supporting the semantic operation Restrict,proposed by Chung and Ladusaw (2004). The basic idea is that alongside the standard semantic operations or functional application, predate modification, predicate abstraction (as well as existential closure) an operation which composes a verb with a (generally) property denoting expression without saturating the former is needed. I first review some of the evidence Chung and Ladusaw have adduced oin favour of this proposal (from Maori) and then consider constructions from Japanese and Korean which present a nominative marked noun phrase (subject) followed by a seemingly fully saturated sentence. I argue that these constructions are also best treated in terms of the operation Resrtict (predicate).
Date: Wednesday 24th February 2010, 11:15, CS202J Computer Science, Univ. of York
Speaker: Sandy Louchart, Heriot-Watt University, School of Mathematics and Computer Science
Multi-Level Drama Management - AI requirements
The Interactive Storytelling community is a melting pot of disciplines, arts, sciences, people, backgrounds, visions and objectives. Thus many aspects of the research are still relatively prone to interpretation: What is a story? When is a story interactive? Are video games interactive stories? Most importantly, as a user What I am supposed to do? Shall I create/author stories? Shall I aim to control stories? Or shall I just enjoy them as interactive storytelling is primordially entertainment! In his talk, Dr Louchart aims to outline in a simple fashion the complexities and paradigms of research in Interactive Storytelling and describe the different levels of narrative representations through the concept of Distributed Drama Management.
Date: Wednesday 10th February 2010, 11:15, CS202J Computer Science, Univ. of York
Speaker: Maria Arinbjarnar, University of York, Department of Computer Science, Artificial Intelligence Group
Autonomous agents and Bayesian networks in Directed Emergent Drama (DED)
The Directed Emergent Drama architecture facilitates emergence in a virtual reality through user and agents' interactions. The emergence is directed in order to ensure that the drama emerges into an expected drama genre, e.g. love or mystery drama. The agents are autonomous to facilitate distributed computation and to allow for a more fluid response to user actions. The autonomous agents have traits, moods and emotions in order to increase believability of the agents as having distinct personalities, so that the majority of users will find the moods and traits of the characters plausible. Bayesian networks (BNs) are particularly suitable to implement autonomous agents' decision mechanism because they support transient emotions and decision making and accommodate for conflicts between the agents goals. In their basic form, BN reasoning algorithms do not scale well, since the cost of updating values in a BN grows exponentially with its size. We use relevance reasoning to reduce the complexity of computing inference by extracting a subset of variables from a BN that is relevant to a localised inference based on target and input variables.
Date: Wednesday 3rd February 2010, 11:15, CS202J Computer Science, Univ. of York
Group meeting
Date: Wednesday 20th January 2010, 11:15, CS202J Computer Science, Univ. of York
Speaker: Adel Aloraini, University of York, Department of Computer Science, Artificial Intelligence Group
Extending the Graphical Representation of KEGG Pathways for a Better Understanding of Prostate Cancer Using Machine Learning.
The problem of variables selection with large number of candidates have been addressed in the field of biology recently. This is due to the importance of data collection technologies. One example is microarray data analysis techniques where the number of genes (predictors) to be examined exceeds the number of samples(observations).In such model selection problem ,the hope is to achieve: an accurate prediction, and interpretable models. In this talk extending the graphical representation of KEGG pathways of prostate cancer will be used as a case study to show how variables Selection methods can be used to learn an extended Graphs for KEGG pathways.
Date: Wednesday 6th January 2010, 11:15, CS103 Computer Science, Univ. of York
Speaker: Matthew Butler, University of York, Department of Computer Science, Artificial Intelligence Group
Financial Forecasting with Artificial Intelligence
An overview of Artificial Intelligence techniques and their application in forecasting financial assets.
Date: Wednesday 6th January 2010, 11:15, CS103 Computer Science, Univ. of York
Speaker: Suraj Pandey, University of York, Department of Computer Science, Artificial Intelligence Group
Sentiment Analysis
Sentiment analysis focuses on identifying positive and negative opinions, emotions and evaluations expressed in natural language. In its most basic form Sentiment Analysis is classifying sentences or text into either a positive sentiment bearing or negative sentiment bearing one.This talk will mainly focus on, how SA evolved ,different approach on how it is accomplished and what are the problems that remains outstanding in SA systems.