|
Biological inspiration has made a significant contribution to robotics, due to the many similarities between the tasks solved by animals and those we would like our robots to achieve. Insects have attracted particular interest for a number of reasons. Their specialised sensors, for example to detect visual motion, sound direction or wind flow, can be copied in hardware. Their robust behaviours, such as homing, flight stabilisation and six-legged walking, provide useful capabilities to imitate. Their small brains suggest efficient processing algorithms for sensorimotor control. At the same time, robot implementations have led to better understanding of biology. As a technology for modelling hypothesised mechanisms of animal behaviour, robots offers several advantages. They enforce consideration of the complete loop from sensors to actuators, and the closing of that loop through the environment. They also allow models to be tested in real experiments in real environments. Insect sensorimotor systems are simple enough to make such complete models, at the neural level, plausible to implement and evaluate. I will describe several examples in which such implementations have made direct contributions to biology. Although relatively simple, insects are not just reactive systems. In fact, they are capable of much more complex behaviours - including learning, integration of multimodal cues, real-world navigation and flexible behavioural choice - than any existing autonomous robots. I will review some of these capabilities, and what is known of the neural circuits that underly them, and outline our recent work towards designing an insect brain control architecture for robotics.
Cellulose, in the form of biomass, is the ultimate renewable resource. Its conversion to starch would provide a hugely abundant source of material which could be used for the manufacture of biofuels or other biological products, as an animal feed supplement to release grain for human food use, or even as the basis of a food for human use. Given the present food and energy shortages, the advantages of such a process are clear. With this in mind, Edinburgh iGEM 2008 have devised systems for E. coli to degrade cellulose into glucose, to upregulate glycogen and terpenoid production, and to convert glycogen into starch. We have also designed software capable of generating a model in SBML format from a list of genes and promoters entered by the user. This is supported by a background database allowing users to build models based on published data."
The performance of machine translation systems varies greatly depending on the source and target languages involved. Determining the contribution of different characteristics of language pairs on system performance is key to knowing what aspects of machine translation to improve and which are irrelevant. This paper investigates the effect of different explanatory variables on the performance of a phrase-based system for 110 European language pairs. We show that three factors are strong predictors of performance in isolation: the amount of reordering, the morphological complexity of the target language and the historical relatedness of the two languages. Together, these factors contribute 75% to the variability of the performance of the system.
Big hopes rest on learning from demonstration, or imitation learning, to provide a flexible way of programming robots which would even enable a layman to teach a robot a new task. The main idea of this approach to robot programming is to use learning algorithms in order to detect and apply those aspects of demonstrated movements that are relevant to a new task, but obviously this is much harder than it sounds. In my talk I will present learning from demonstration, it's motivation and particularly it's open problems. I will also briefly discuss our contribution to the field and as a consequence will raise the question whether just observing movements is enough to learn a task.
"For ever and ever", "immer und ewig", "for ever and a day" - idiom, as well as Buzz Lightyear, uses phrases that, one might say, describe infinity and then some more - but this is mere rhetoric. In the late 19th century, Georg Cantor discovered that mathematically, it makes perfect sense to speak of infinities beyond infinity, and indeed it is forced on us by the assumptions underlying everyday mathematical practice. The Cantorian revolution was a shock to the foundations of mathematics unprecedented in history, and was resisted accordingly, but within a couple of decades widely accepted. Famously, David Hilbert, himself the greatest mathematician of his time, referred to "the paradise that Cantor has created for us" ["... dem Paradies, das Cantor uns geschaffen, ..."]. Nowadays, the infinities beyond infinity are an essential part of the everyday toolkit of many computer scientists at the LFCS end of the subject. This talk gives a gentle, mostly intuitive, introduction to the smallest of Cantor's infinities (putting one toe over the border of Paradise), and hints at how and why infinities are useful in computer science. For those who were at Stan Wainer's talk at the Bundy Symposium, we'll also see an example of what he meant by "fast-growing functions". No knowledge will be assumed beyond the ability to read algebraic notation (in the school sense) and maybe a bit of first-order logic. (For those who know, and wonder what I'm going to do: just the countable ordinals up to epsilon_0 (and maybe the Veblen hierarchy); ordinals as a measure of induction height and theory strength, illustrated by the Ackermann function for PRA and (my favourite) Goodstein function for PA.)
I will introduce the EU-funded JAST project ("Joint Action Science and Technology"). I will then give a detailed description (including videos!) of the human-robot dialogue system being built as part of the JAST project. The robot is able to work together with a human user to build construction toys on a joint work surface, with the two participants coordinating their actions through speech and gesture.
We propose a psycholinguistically motivated version of TAG which is
designed to model key properties of human sentence processing, viz.,
incrementality, connectedness, and prediction. We use findings from
human experiments to motivate an incremental grammar formalism that
makes it possible to build fully connected structures on a
word-by-word basis. A key idea of the approach is to explicitly model
the prediction of upcoming material and the subsequent verification
and integration processes. We also propose a linking theory that
links the predictions of our formalism to experimental data such as
reading times, and illustrate how it can capture psycholinguistic
results on the processing of "either... or" structures and relative
clauses.
Joint work with Frank Keller. Talk is a dry run for an
oral presentation at the TAG+9
workshop 2008.
For much of the twentieth century, the widespread view was that language evolution cannot be studied, owing to an absence of data. Nevertheless, recent decades have seen a revival of interest in the question of how human beings came to possess language, and how language came to have the shape it does. Such research has broken out of its traditional straitjacket by relying on data not only from fields like archaeology and comparative biology, but also from pioneering simulations of linguistic behaviour and language emergence. Computational modellers have had artificial agents interact in artificial environments, while robotic engineers have placed artificial agents in real-world environments. Recently, a third approach has been pioneered: the placing of real human agents in artificial environments. Much of this work has been conducted at the Language Evolution and Computation Research Unit, here at the University of Edinburgh. This talk will describe three such studies, and thus demonstrate that language evolution can be studied in the laboratory.
YouTute is a vicarious learning web-interface for students. They can use YouTute for example to improve their learning and/or recap their studies. As the name might suggest, this web-interface consists of video recordings taken of tutorials. Additionally to these recordings it contains recordings captured continuously from a "Smartboard", and learning material like tutorial scripts or lecture slides.
All data for this project was collected in Informatics at the University of Edinburgh, here especially from tutorials in Informatics 2B - Algorithms, Data Structures, Learning. Four of nine tutorials have been video recorded and were using a Smartboard during term. In the end three of these four tutorials were included into the database of the interface. This collected video data gives students the opportunity to watch and listen to how their peers might approach a certain tutorial question in class, possibly in a way different to their own. Hence the term Vicarious Learning since we think that watching others can help the onlooker to get a better understanding for solving a task/problem. In this spirit YouTute gives students of one tutorial the chance to look at another tutorial with the same topic but different tutor and other students discussing a subject. Thus onlooking students get a chance to broaden their knowledge and maybe get another perspective - one they maybe never thought about before or maybe weren't able to verbalise - on a problem.
YouTute is supposed to be designed in a usability friendly and ergonomic way listing tutorials, "tutes" and tutorial scripts in a drop down menu on one side, showing video data and control on the other side of the screen. Since the web-interface is designed to become hopefully a part of the students' community it allows all users to edit videos themselves and recommend these edited parts (here called tutes) to their peers.
Legal information retrieval currently uses boolean and conceptual search, relying on input from legal experts to perform adequately. An automated means of searching for analogous precedents to a present case that do not share keywords or hail from the exact practice area will help paralegals conduct comprehensive search critical to effective litigation. Starting from the assumption that legal cases can boil down to 'who did it', this talk will begin with a survey of causal theories from law and philosophy. Insights will be applied to propose a novel representation of court opinions as the substrate for search: events and causal structure. Exploring this idea further, the second part of the talk will introduce and compare language resources that facilitate the extraction of events and enable training of statistical systems for the key subtask of semantic role labelling (SRL). FrameNet, PropBank and VerbNet will be discussed as individual corpora and a unified language resource. Finally, an example will be worked through simulating computational analysis of opinions from two cases of the Supreme Court of Canada involving murder and a defence of not guilty by reason of insanity.
Given a formal mathematical theory T such as the theory of natural
numbers, the theory of abelian groups, etc., the ``decision problem for
T'' asks for the existence of an algorithm A s.t. given any conjecture C
in the language of T, A will decide (read: compute in a finite number of
steps) whether or not C is true. When no such algorithm can exist for a
theory T, we say that T is undecidable.
Many important theories in mathematics are undecidable. For instance,
Goedel and Church proved the first-order theory of the natural numbers
to be undecidable by the method of arithmetical coding. Similarly,
Julia Robinson proved the first-order theory of the rational number
field to be undecidable by exhibiting a collection of quadratic forms
that could be used to define the natural numbers within the language of
the rational number field, thus causing the first-order theory of the
rational number field to inherit the undecidability of the first-order
theory of the natural numbers. More recently, Yury Matiyasevich proved
the existential theory of of the natural numbers to be undecidable (this
is the Davis-Putnam-Robinson-Matiyasevich negative solution to Hilbert's
Tenth Problem, which was open for 70 years!).
All of the above undecidability results have a very important property
in common: They are about *discrete* number theories. The state of
affairs for continuous domains is decidedly rosier. For instance, in
the 1930s, Tarski famously proved the first-order theory of real closed
fields, and hence the real numbers, to be decidable. Similar results
hold for other mathematically convenient continuous domains, such as
algebraically closed fields like the complex numbers.
While we know that the full first-order theories of the discrete number
systems mentioned above are undecidable, my work concerns the isolation
of decidable fragments of these theories. One famous such result by
Presburger and Skolem is the fact that the *linear* theory of the
natural numbers is decidable. My interest is thus in isolating
decidable *nonlinear* fragments of discrete number theories.
I have recently proved a handful of simple results on this topic,
leading to the isolation of some nontrivial nonlinear fragments of both
the first-order theories of the natural and rational numbers. My
techniques are model-theoretic and topological in nature. Specifically,
I work to find linguistic restrictions one can place upon the language
of a discrete number theory that entail the topological properties it
can express are similar to the topological properties of a decidable
continuous domain. When such a result works out, it allows one to
soundly relax these linguistically restricted conjectures over the
discrete domain in question to related conjectures over a decidable
continuous domain, and thus to then decide their truth or falsity
algorithmically.
In this talk, we will explore the following questions:
The talk will describe work in the Virtual University of Edinburgh - Vue - and what some Informatics projects are doing in the Vue regions in Second Life, which is one of the many virtual worlds now emerging as developments of the old fashioned 2D world of the WWW. AIAI2 has an office in Second Life. It is a base for a range of collaborative systems work with others academics and industry. Why? What is it used for? Where is this leading? Can it be beneficial to your research?
The Nintendo Wii, since its release in late 2006, has surpassed all sales expectations and has become the must-have gaming console of 2008. Much of this is due to the innovative wireless Wiimote, which allows for gestural control, giving the potential for the most intuitive and universal gaming environment ever made. The past few years have also seen the growth of music-based video games such as Guitar Hero and Rock Band, as well as a widening availability of consumer music production software. In this seminar, Yann Seznec will discuss how these things can be combined, and how creative gaming environments, particularly for the Wii, should be developed. Yann Seznec is an EPIS entrepreneur at the University of Edinburgh developing creative music software using the Wii remote as an interface. See his work at http://www.theamazingrolo.net/wii/
The talk addresses a problem in the modelling of reasoning with evidence in a legal setting - the "alternative perpetrator" , "incrimination" or "Perry Mason" defence. Comparing dialogue based and ATMS based approaches to modelling evidentiary reasoning,we find that the legal treatment of this type of defence does not sit easily with either approach. An extension to the ATMS based approach to reasoning with hypothetical evidence is introduced and tested against the notorious " "Central Park Jogger" rape case. In this case, a group of teenagers was initially convicted of a particularly brutal attack on a female jogger. recently, another suspect in the original investigation confessed to the crime which he claims he committed on his own. While he was convicted, the consequences for the original verdict remain contested. The paper explores the contribution a computational theory of reasoning with evidence in law can make to explain why this is the case.
Gadder (www.gadder.co.uk) has been involved with the development of an Alumni Research Manager ARM software. The company's mission is to help fundraisers be more successful through providing timely and relevant information on their most promising prospects. The software comprises a multi-agent system that acts as crawler to extract information from the web. Then unstructured data mining is performed on the data. This information is then presented to the researcher for further analysis.
This talk presents several research projects that combine conceptual frameworks from psycholinguistics with tools from computational linguistics, motivated by problems in tutorial dialogue research.
First, we measured cohesion in tutorial dialogue corpora by counting "cohesive ties," which occur when the same word (or word stem) occurs in neighbouring tutor and student utterances. We found that this lexical reiteration measure of dialogue cohesion was correlated with learning for only our low-knowledge students, and interpreted these results in light of insights from the text processing literature.
Recently, we have extended this work by using WordNet and a notion of semantic similarity to measure both lexical and semantic cohesive ties between tutor and student. This extension improved our ability to predict learning among our high knowledge students.
Finally, convergence is the phenomenon in which two dialogue partners come to exhibit similar features. Motivated by the development of corpus-based measures of syntactic convergence, we developed measures of acoustic/prosodic and lexical convergence, and again applied them to tutoring dialogues. Our results show that these kinds of convergence occur in tutoring dialogue, and that the amount of convergence between student and tutor also predicts learning.
This is joint research with Arthur Ward, University of Pittsburgh.
Ontology has recently shifted from a purely philosophical study of existence to include the current efforts to create machines which can understand the world around them.
The aims and intuitions of ontology research within Informatics vary widely. From the formal ontology of Barry Smith to the cognitivist semantics of Peter Gardenfors (amongst others).
My interest is in the ability of an AI agent to evolve their representation of the world through their interaction with it. This is based on the assumption that no innate ontology given to the agent will be enough to encompass all the concepts which it would be useful for them to know. In this scenario an agent must be able to learn new concepts, to adapt old ones and to understand how these concepts relate to sensory observations.
The aim of the talk is to introduce the problem, discuss why it is of interest to AI in general and look at some possible mathematical frameworks which could be used to tackle the problem.
Multi-touch interfaces are touch screens which allow several simultaneous points of interaction with a computer. Although the idea is not new, there is a strong current in this direction at the moment, spurred on by Jeff Han, Microsoft's "Surface" and the Apple iPhone. There are two areas in which multi-touch is useful: allowing a more direct, intuitive and expressive interaction with digital objects, and allowing several people to work in the same digital space. In this talk, I'm going to talk a bit about what is happening in this area at the moment - some of the key players and technologies. Then, I'll outline how the "Truth Table" (an in-house multi-touch table) works. Finally, I'll talk about the software running on the table, which is designed to support playful brainstorming using Web 2.0 content. This will take about 20 minutes, after which there will be chance to play with the table itself and ask questions.
Data-driven methods allow effective system development with a reduced amount of idiosyncratic engineering.
One case example is open-domain question answering (Q&A), an information retrieval task where a system must answer any kind of question by pinpointing answers in a large document collection. I present the design and implementation of the LSV Alyssa Q&A system, which was built as an experimental platform to find out what works and what doesn't in Q&A, and in order to develop new, competitive algorithms for the task.
Alyssa is designed for simple experimentation and with flexibility and extensibility in mind, and it supports multiple IR paradigms. In this presentation I will focus on cascaded language-model based retrieval (LMIR) followed by maximum entropy-based candidate answer ranking and weighted answer stream fusion.
For factoid questions, Alyssa ranked third in the world after LCC1 and LCC2 at the recent TREC 2007 evaluation carried out by US NIST. For the definition questions, Alyssa ranked second, and overall achieved rank 4. I conclude with suggestions for further improvements that are conjectured to yield significant improvements.
This is joint work with the LSV Q&A group at Saarland University.
Short presentations by PhD students:
With ageing, human voices undergo some alteration due to anatomical and physiological changes. As a result, the acoustic properties of normal adult voices and voices showing signs of ageing, tend to be considerably different. This results in degradation of the performance of speaker independent Automatic Speech Recognition(ASR) Systems on speech from aged voices. One way to improve the performance is to adapt the ASR using speech of aged voices.
I am presently looking at approaches to identify data showing signs of ageing from a large pool of speech corpus. I shall discuss in the talk, about the acoustic cues of ageing voices and my experimental setup to use them in data selection and some results for data selection using Maximum Likelihood Linear Regression (MLLR).
The last decade has seen an explosion in the amount of information about individual components of the cell, e.g. genes and proteins. Little, however, is known about how they interact to form the complex emerging behaviour of a cell. Such a systems level understanding is sought in the young field of systems biology, in which computer science plays an increasingly important role.
I will give a brief overview of systems biology and show how a programming language we are developing -- the "language of biological systems" (LBS) -- can be used to model a tiny biological system. The aim is to use subsequent computer simulation and analysis to predict the behaviour of the system, perhaps under the influence of novel drug candidates.
This work is from a PhD project supervised by Gordon Plotkin.
Language models are probability distributions over a set of unilingual natural language text used in many natural language processing tasks such as statistical machine translation, information retrieval, and speech processing. Since more well-formed training data means a better model and the increased availability of text via the Internet, the size of language modelling n-gram data sets have grown exponentially the past few years. The latest data sets available can no longer fit on a single computer. A recent investigation reported first known use of a probabilistic data structure to create a randomised language model capable of storing probability information for massive n-gram sets in a fraction of the space normally needed. We report and compare the properties of lossy language models using two probabilistic data structures: the Bloom filter and lossy dictionary. The Bloom filter has exceptional space requirements and only one-sided, false positive error returns but it is computationally slow in scale which is a potential drawback for a structure being queried millions of times per sentence. Lossy dictionaries have low space requirements and are very fast but with two-sided error that returns both false positives and false negatives. We also investigate combining the properties of both the Bloom filter and lossy dictionary and find this can be done to create a fast lossy LM with low one-sided error.
The world is facing an acute and growing shortage of teachers. According to UNESCO, there are currently 60 million teachers but 50% more are needed. The current model of teacher training, school building, and international monitoring will not scale quickly enough to meet the burgeoning demand of students globally. To meet demand, governments and educational institutions are turning to self-service e-learning systems that use a combination of automated multiple choice, automated short answer, and, most controversially, automated essay scoring (AES) technologies. In our next IDT, we investigate how AES works, who's already using it, and what it means for education in the future.
The Paperless Office is a vision that has been discussed for the past two decades. Spurred by the availability and adoption of the Personal Computer, computers are now the predominate tool in the workplace. However, due to the affordances of paper, the paperless office is still a myth. Part of the reason is that one of the core affordances of paper is its usefulness in collaborative work.
In this talk, I will discuss the concept of the Paperless Office, followed by recent innovations at two levels - the interface level and the collaborative work level. Examples of innovation at the interface level includes BumpTop (adding physics to the Desktop metaphor), G.hos.t (remote desktop environments), multiple touchscreens, and mobile office suites. At the collaborative work level, I will mainly draw from web based technologies which facilitate distributed working practices including wikis and groupware. Examples of which include Huddle, Basecamp, MediaWiki. I finish the talk by discussing the implications these innovations will have for the paperless office.
Playing around with the Wiimote
It's recently been noted that "the movement-sensing functions of the Wiimote do not have any precedents to help designers understand which make most intuitive sense to users for different actions. For example, what is the best way to allow users to open a door?" (Weedon, 2007). In this informal session, we'll explore
some of the things you can do with a Wiimote and a Nintendo Wii, and related usability issues. Time permitting, we'll survey things you can do
with a Wiimote *without* a Wii. Audience participation compulsory.
In this talk, I will present an algorithm based on linguistic tools that we have developed for watermarking a natural language text. The algorithm expands on the morphosyntactic repertoire of a morphologically rich language like Turkish.
NLW is an emerging research area which aims to hide information in natural language texts. It has various goals ranging from content and authorship authentication to enriching the text with metadata and fingerprinting it for distribution. It basically shares the same goals as multimedia watermarking, but the two obviously employ very different techniques. In the initial stages of the technology, non-linguistic NLW techniques were adapted from multimedia watermarking. Such approaches, however, had limited scope and were not robust against text reformatting and transcription attacks. Attention was then diverted to implementing linguistic tools such as semantic and syntactic transformations, morphological and punctuation manipulations, lexical substitutions, translations and word level typographical alterations. Our contribution to the field is to develop a novel NLW scheme that is imperceptible, secure and based on morphosyntactic manipulations of texts. We also propose a method for making the system robust.
In this talk, I will present the GIVE challenge for natural language generation (NLG) systems. In this scenario, a user must solve a certain task in a 3D virtual environment by moving around, manipulating objects, etc. The computer has a plan that tells it how to perform this task, and the NLG system's job is to communicate this plan to the user, i.e. to generate natural-language instructions that help them solve their task. We plan to run GIVE as a shared task for the NLG community for the first time in 2008. I will present the challenge itself and discuss some problems that it poses for an NLG system and some first ideas for how to solve them.
This challenge is interesting for a variety of reasons. First of all, it provides a new evaluation methodology for NLG systems because it allows us to observe the interaction of the system with a remote user that only connects to it over the Internet; that is, we have access to cheap experimental subjects in the style of psycholinguistic "web experiments". Second, the generation problem now takes place in the context of a (simulated) physical environment, which changes the rules of communication significantly. Third, the instruction giving challenge is related to a number of real-world applications, including human-robot communication, but studies them in a sanitized, simplified environment. Fourth, the instructions must be generated under near real-time conditions, i.e. very efficiently. And finally, 3D games are fun, and I can perhaps demo the very rough prototype I have so far.
"Our Just in Time (JIT) procurement software works in Germany. Why doesn't it work in Brazil? It's the same technology - what's the difference?" (IT Manager - automotive supply chain)
This will be a Cook's tour of some of the more recurrent problems: solution scenarios encountered by one researcher over multiple projects, presented from a socio-technical systems perspective. The session will include examples from projects involving e-business portals, supply chains and HealthGrids, to highlight very typical socio-technical problems that designers and IT managers faced.
I hope to pass on some of my own fascination with the ways 'coupled' socio-technical systems can align social, technical and organisational architectures to advantage, as for example in Web 2.0 applications, and the ways they can be misaligned.
There will be examples of the effect of misalignments that went badly wrong (!) and an opportunity to break out into groups to guess / suggest approaches to some of the socio-technical problems faced by one large UK financial services company when designing and rolling out an Intranet.
During the research, IT project managers often highlighted the lack of practical support or training in dealing with socio-technical problems, and this appears to be increasingly the case as distributed systems support increasingly diverse and distributed user communities, in very dynamic and unpredictable environments.
Speaker:
Jenny is a psychologist currently working in the Social Informatics Cluster in the School of Informatics in Edinburgh University, researching collaboration and data integration in Grid based eHealth systems. Prior to this she was a Research Fellow in the Management of Technology at Edinburgh University Management School and a Visiting Research Fellow at Curtin University of Technology looking at the design and management of complex 'socio-technical' networks.
At present, most biomedical Information Retrieval and Extraction tools process abstracts rather than full-text articles. The increasing availability of full text will allow more knowledge to be extracted with greater reliability. To investigate the challenges of full-text processing, we manually annotated a corpus of cited articles from a Molecular Interaction Map (Kohn, 1999).
Our analysis demonstrates the necessity of full-text processing; identifies the article sections where interactions are most commonly stated; and quantifies both the amount of external knowledge required and the proportion of interactions requiring multiple or deeper inference steps. Further, it identifies a range of NLP tools required, including: identifying synonyms, and resolving coreference and negated expressions.
Using this corpus we have also investigated the challenges of sentence and passage retrieval. From the annotated facts we generate keywords for sentence retrieval, and analyse the impact of various query relaxation strategies on performance. We also investigate the use of hedging and commitment in the reporting of scientific results on retrieval.
The presentation will include the Research & Development (R&D) process
of the Dot Red Engine (©) (DRE) casual games platform within the
School of Informatics research environment. Additionally, more
intimate details for the design and implementation of the elaborate AI
algorithms used for Dot Red Games's innovative first game title, Crazy
Space, will be presented. The game creation process included taking in
full consideration HCI factors for designing a game to appeal in the
casual games market, a market mostly targeting a mature women's
audience with a preference for relaxing, non-violent, less complex,
but still very fun games.
See and try out the "Crazyspace" game at www.dotredgames.com!
The Penn Treebank does not annotate within base noun phrases (NPs), committing only to flat structures that ignore the complexity of English NPs. This means that tools trained on Treebank data cannot learn the correct internal structure of NPs. This paper details the process of adding gold-standard bracketing within each noun phrase in the Penn Treebank. We then examine the consistency and reliability of our annotations. Finally, we use this resource to determine NP structure using several statistical approaches, thus demonstrating the utility of the corpus. This adds detail to the Penn Treebank that is necessary for many NLP applications.
In this talk I will discuss MAMA (Musical Acts - Musical Agents), an architecture for musical agents. I will discuss the anatomy of a musical agent, and how it relates to other agents. I will explore the issue of representing music such that distributed, logical agents can play a given "score". This will use the notion of Musical Acts - a layer of abstraction which allows an agent to reason about the actions of others. Finally, there will be a short demo (technology permitting) of a current case study using the system.
This talk is about experimental psycholinguistic work aimed at distinguishing between resource-limitation and resource-allocation theories of processing difficulty in human sentence comprehension. We focus on the effect of manipulating relations between verbs and their dependents on the processing difficulty of the verbs themselves. One of these experiments involves clause-final verbs in matrix clauses of German, and the other involves embedded verbs in relative clauses of Russian. Our experimental results give strong evidence for the existence of both resource-limitation and resource-allocation theories, and shed light on the precise conditions under which the predictions of each type of theory emerge.
This paper explores cognitive mechanisms that process models of complex systems - represented in their implicit form - in order to produce 'conceptual' redescriptions, which (we hypothesise) could reveal knowledge about these models that is not accessible on the implicit level. The aim of this exploration is to support new ways of conceptualising the phenomenon of emergence, the main characterising feature of complex systems in general. Here, we focus on exemplar Cellular Automata (CA) rules developed to perform the density (majority) classification task (as defined by Mitchell et al., 1993). Conceptual representations of the best known rules for this task will be presented and we will show how the resulting abstractions can be considered suitable for the formation of "Conceptual Spaces", wherein rules that perform similar computations are positioned in close proximity.
A short talk on the development of a robot arm that plays connect 4 against humans - which was seen recently at the science festival - followed by an opportunity for people to play against the arm. The project was specifically developed as a public demo and as such contains little ground breaking AI work. Despite this people are genuinely fascinated by a machine that plays in their world, rather than on a computer screen. The talk will motivate the creation of this demonstration, outline how it works, what need improving and give anecdotal evidence of how people reacted to it.
Abstract: Can artistic applications of computers combining interaction and visualisation be seen as useful or are they simply entertaining distractions? In my presentation I will illustrate how "Art as a Mode of Enquiry" represents a methodology by which new ideas are created and software and hardware systems can be challenged with the possibility of creating new paradigms of computational theory.
I will illustrate these ideas through examples of the interactive works:
Abstract: The problem of trust in distributed open systems may be defined as the problem of who to interact with, when to interact with, and how to interact with. Available research has mainly focused on analysing, improving, and developing strategies that tackle the 'who/when/how' question of trust. In this talk we focus not on the strategies, but on the specification and verification of such strategies. We present our specification model, which is based on the combination of interaction and deontic models, and our dynamic model checker, which may be invoked by agents at runtime when the conditions of verification are met.
Abstract: The browser, once a simple hypertext navigation tool, has evolved into a platform for delivering sophisticated applications over the web. Web programs fall into three categories: server-side programs with HTML form-based interaction, client-side programs that transform the DOM in response to user events, and "AJAX" programs that divide computation between communicating processes on the server and the client. Similar problems arise in each of these cases: the constraints of the architecture impose an "inside-out" program structure which makes it difficult to write large, coherent applications. The solutions are also similar: in each case continuations put the thread of control back in the hands of the application programmer.
Surprising as it may seem, most images of natural scenes exhibit certain statistical regularities. For example, for almost all natural images the two-dimensional Fourier transform yields a power spectrum which follows (approximately) an inverse-square law dependant on frequency. Biologist have in the past argued that certain aspects of mammalian visual processing have evolved to take advantage of such invariances. I believe that natural image statistics can provide useful prior knowledge in the design of robotic visual navigation systems. I will discuss some key image statistics and describe an attempt I've made to justify the usefulness of a particular visual navigation algorithm using a key natural image statistic.
Empirical studies suggest that speakers show a reliable increase in the use of particular syntactic structures after repeating or hearing them in an unrelated sentence (e.g., Bock 1986, Pickering and Branigan, 1998). In this talk, I report two psycholinguistic experiments to investigate the effect of syntactic priming in Japanese. Using a picture-description task, Japanese speakers read the `prime' sentence and describe the `target' pictures orally. Word order/voice and animacy in prime sentences were manipulated (experiment 1: SOV/OSV, animate-inanimate, inanimate-animate, experiment 2: active/passive, animate-inanimate, inanimate-animate). Both alternatives could be used to describe the unrelated target pictures. While I found that speakers were more likely to use `primed' word order and voice in the target picture description, there was a lack of conceptual (animacy) effect (e.g., animate object (agent) in passive to animate object (patient) in active), contrary to the finding by Bock, Loebell and Morey (1992). I suggest that these effects are instances of syntactic but not conceptual priming, and describe how these results are interpreted in the current model of language production (Bock and Levelt, 1994). In addition, I discuss how the current study can evaluate accounts of relation-changing operations (e.g., transformational grammar, Chomsky, 1981).
The behaviour of the virtual characters we encounter in films and games is for the most part created by hand. For every eyebrow raise, blink, or turn of the head the animator had to specify the trajectory, the timing, and numerous other parameters. Our goal was to specify some of these parameters automatically. Stochastic methods are employed that learn the distribution of movement from a certain speaker. At synthesis time, new motions are created by sampling from that distribution.
I will talk about our paper which was recently accepted at Cognitive Psychology. It reports two studies which showed evidence for syntactic priming in comprehension. Syntactic priming is a phenomenon that people tend to reuse the same syntactic structure across consecutive sentences and it has been demonstrated by many language production studies (e.g., Bock, 1986). It is often assumed that comprehension and production share similar mechanisms and thus that syntactic priming also occurs during comprehension (e.g., Pickering & Garrod, 2004). However, previous research investigating priming during comprehension has mainly focused on syntactic ambiguities that are very different from the meaning-equivalent structures used in production research. Using the visual-world technique, we found that priming during comprehension occurs in ditransitive sentences similar to those used in production research. Furthermore, we also found no evidence for priming when the verbs were different in prime and target. This finding highlights the difference between comprehension and production mechanisms and we argue that in contrast to production, the priming effect in comprehension is completely lexically dependent. Also, I discuss that our results are the first demonstration that during comprehension, people anticipate not only the type of upcoming arguments but also the order of upcoming arguments, which indicates that anticipations during sentence processing include a structural component.
I will discuss my research (in progress) on the acquisition of Italian grammatical gender in second language learners. For this study, I am going to compare two groups of advanced learners of Italian (a group with English as their first language and a group with Dutch as their first language) to a native speaker control group. Dutch and English learners have been chosen because Dutch and English differ in one crucial respect: English does not have grammatical gender, but Dutch does.
The aim of this research is to shed light on various issues: can gender be acquired by second language learners (i.e. do the experimental groups perform similarly to the control group)? Is there a difference in the ability of the two experimental groups to acquire gender? And if there is a difference either between the two experimental groups or between the experimental groups and the control group, is this difference at the level of processing or at the level of knowledge? I hope to answer these questions by asking the experimental groups and the control group to perform an online self-paced reading task and an offline six-point-scale acceptability judgement task.
E-science is hailed as a new way of doing science. Its aim is to create an infrastructure which enables collaborative and distributed work, coupled with large computational processing power and data storage facilities. E-Science brings together a number of challenges both social and technical, and most of the time, a combination of both. These include collaborative working, information governance, database linking, security, etc.
Producing knowledge from data is part and parcel of the scientific process. As a result, data production and data analysis are closely coupled. However, under the vision of e-science, it is envisaged that data can be produced and repurposed for a number of different ways. This changes the process in which knowledge is produced.
In this talk, I will be presenting on some initial findings from my PhD research about the role of data and its affects on the development of e-science systems. I will be presenting empirical work from a 10 month observational study. Then providing some discussion points on the building of e-science systems.
The field of second language acquisition has traditionally investigated properties of the developing grammar of the L2, while the study of how L2 learners process the L2 input has been relatively neglected. In this talk I will outline some of the key issues in second language sentence processing, and discuss how processing difficulties may lead to non-nativelike language behaviour. In particular I will examine how the visual-world paradigm can be successfully applied to address these issues.
In this paper, I outline some of the theoretical ideas that underpin my research into Animal Geographies and Nature Conservation. Nature conservation is seen as the process of managing and reversing (human) actions deemed to have degraded nature. Part of this process involves the definition and consequent legislation of what natures belong in certain places, and therefore, by extension, what natures fail to conform to these `rules'. It is common practice to attempt to eradicate or move the offending natures out of the `wrong' places. I will use examples from urban and wild natures in Scotland to illustrate the anomalies associated with such actions, and link these ideas into a wider debate about indigeneity.
The topic of this talk is closing-in phenomenon, which is a behaviour observed in graphic copying tasks, in which the copy is placed abnormally close to the model, or even directly on top of it. This behaviour is commonly observed in patients with dementia, and in early childhood (2-5 years). The purpose of our research is to evaluate competing hypotheses proposed to explain closing-in behaviour. In particular, we wish to test between visuospatial, working memory and attentional explanations of the phenomenon. We will report the results of initial experiments carried out with young children and patients with dementia and present future research plans.
Amodal completion is a realistic problem in 3D computer vision.
As the requirement for more realistic 3D environments is pushed forward by the computer {graphics|movie|simulation|games} industry, attention turns away from the creation of purely synthetic, artist derived, environments toward the use of real world 3D captures. However, common 3D acquisition techniques are realistically only 2.5D in nature - such that the backs and occluded portions of objects cannot be realised from a single viewpoint.
Amodal completion is required to produce realistic and plausible 3D environments from original 2.5D real-world captures. Here we detail a completion approach based on the non-parametric synthesis of localised surface relief over an underlying geometric surface model. Additionally we show how this technique can be extended to the completion of 3D colour data and the concept of 3D texture transfer.
Many theories of language evolution assume a selection pressure for the communication of propositional content. However, if the content of such utterances is of value then information sharing is altruistic, in that it provides a benefit to another at expense to oneself. Traditional evolutionary explanations of altruism are kin selection and reciprocal altruism. However, neither of these applies to the case of conversation and information sharing. Instead, I present a series of arguments that information sharing is in fact selfish behaviour.
The 'Health and Personality Processes: Links Explored' (HAPPLE) study is a longitudinal study of personality traits and health in British Internet users, funded by the ESRC. The study was granted ethical approval by the NHS Multi Region Ethics Committee for Scotland (A/B) and is hosted at www.happle.org.uk. Participants complete the 300-item version of the International Personality Item Pool NEO-FFI (Goldberg, 1999), the Health Behaviour Marker Scales (Vickers et al., 1990) and the SF-36 Health Outcome Survey (Ware, 1993). Data is transmitted using HTML and PHP to a MySQL database using 128bit encryption technology. At least three months after registering, participants return to 'log in' and complete later waves. This paper describes the study design and selected findings from the first wave of the study. Particular reference is made to methodological, logistical and ethical challenges which researchers may face when conducting Internet mediated research (IMR). For example, how to obtain informed consent electronically and the non-equivalence of web-based psychological measures to traditional paper-and-pencil scales. Future plans for analyses of the HAPPLE dataset are outlined.
The phrase-based, joint probability model provides a strong probabilistic framework within which to learn both word and phrase translations simultaneously. However, the model's usefulness is limited by its computational complexity. This project proposes novel hybrid models which use the highly optimized standard phrase-based models to restrict and guide the search space of the joint probability model. We show that this reduces the complexity of the model and allows it to scale up to training corpora that are four times larger, while at the same time improving the quality of the final translations.
We also present several other methods for reducing the training time and improving the quality of the hybrid models.
Spoken language is usually understood and analysed without making reference to written language. However, written language has a pervasive influence on both our conceptualisation and our analysis of spoken language. For most practical purposes there may not be a pressing need to recognize the influence that written language has on spoken language, but there is one area where it is crucial to come to terms with the relationships between spoken and written language - that is, when observed literacy impairments (specifically, in developmental dyslexia) are attributed to impaired phonological representations. In this talk I will provide a critique of the conclusions I drew from an earlier project on the phonological deficit in dyslexia, and show how the role that literacy acquisition plays in shaping individuals' representation of language still needs to be adequately recognized before we can draw firm conclusions about the sources or causes of developmental dyslexia.
Meiotic mapping methods allow the identification of genes influencing a disease when the biological basis of the disease is unknown. These approaches have been very successful in mapping genes for certain kinds of disease, and provide an outstanding opportunity for students without a scientific background to make substantive contributions to human epidemiological research. In this talk, I will provide an overview of meiotic mapping methods and discuss the recent successful mapping of a gene implicated in Age Related Macular Degeneration. I will also discuss how some modeling techniques taught within the School of Informatics may help overcome limitations of existing mapping approaches.
This talk is about learning competent behavior in complex environments with very little prior knowledge. The objective of an agent is to maximize its expected cumulative reward over time. In order to produce competent behavior quickly, it is necessary for the agent to build a model of the world and use this model to plan its behavior. I will discuss ways of adapting models and plans in response to the experience accumulated by the agent.
People who use English as a foreign language often struggle with the handling of the modal auxiliary verbs. This is mainly because of the various applications with each modal auxiliary verb. There are many similar, overlapped meanings between the modals. However, at the same time, it seems that there exists some difference in the detailed significations between them.
Modal auxiliary verbs are applied to several politeness expressions. In this context, it looks as if the modal is a word specifically for politeness use (e.g. a term of respect). People whose mother tongue has terms of respect tend to easily regard it so. On the other hand we encounter other applications which do not show particular politeness or respect in other expressions with the same modal.
I investigate the treatment of modal auxiliary verbs going back to the period of Late Modern English and thereafter to find the relationship between politeness and the modal auxiliary verb. What we can see is the situation where the use of the terms has developed in a conflict between fixed prescriptions and fluid social factors. We are also able to notice that politeness in English has a wide capacity which includes the concept: for example, `respect' and `honour' just as parts of it.
1:10pm, Thursday 18th August, HCRC Seminar Room, 4 Buccleuch Place.
We describe the initial design, training and evaluation of a prototype system enabling the automatic and flexible expansion of an abbreviated, typed text input, into a reconstructed sentence. The system?s target user group is cognitively unimpaired users with motor disabilities, for whom typing can be slow and tiring. It is intended that, by reducing the number of keystrokes required to generate a sentence, without imposing a rigid correspondence between abbreviation and full word, their typed communication might be made more comfortable and expansive. The system employs several techniques and statistics, including vowel deletion, phonetic replacement, and word truncation, extracted from two studies of the methods used by people in abbreviation. Encouraging initial results and evaluation are discussed, along with planned futur e work.
As this is a dry run for a conference next week, and the slot then is only 20 minutes, this talk will only be about 15 minutes, followed by any questions.
Friday 12th August, 4-5pm, Lecture room A, Music, Alison House, 12 Nicolson Square (followed by wine reception).
Considering the differences and similarities in the perception and processing of language and music, we designed an fMRI-experiment to compare the organisation and underlying neural networks of working memory for tonal (sine wave tones) and verbal (syllables) stimuli. In addition, we were interested in the question if musical expertise or the possession of absolute pitch has an influence on the cognitive processes and the neural organisation of working memory for tonal stimuli. Seventeen nonmusicians, sixteen musicians and eight musicians with absolute pitch performed a tonal and a verbal working memory task. Two types of tonal sequences were designed: "in-key" sequences (tones are in key, and three of them belong to one triad) and "non-key" sequences (consists of neither triad nor key). We also investigated the influence of verbal interference on tonal working memory processes in absolute pitch musicians. Results indicate that, the neural networks underlying working memory for tonal and verbal stimuli mainly overlap. However, there are areas that seem to be engaged specifically and/or are more involved in verbal or tonal working memory processes (e.g. verbal rehearsal engaged more Broca's area, whereas during tonal rehearsal the angular gyrus showed stronger activation). The significantly higher performance of musicians compared to nonmusicians, and the different activation patterns between both groups, especially the finding that musicians show stronger activation of the left inferior parietal lobe (an area known to be engaged in the storage during working memory processes), might reflect different strategies and/or use-dependent plasticity triggered by musical expertise concerning working memory processes for tonal stimuli.
Thursday 16th June, 4-5pm, room S1, Psychology, 7 George Square.
Viewers are often surprisingly insensitive to an obvious visual change in a scene when that change takes place during a saccadic eye movement. This phenomenon and related results have led some investigators to propose that visual scene representations do not exist, a strong claim that has profound implications for theories of cognitive architecture and the nature of mind. In this talk, I will present research from my laboratory exploring scene representation using a variety of change detection and memory probe techniques tied to eye movements. New results examining incidental visual encoding will also be discussed. The results demonstrate that high delity and long lasting visual memories can be (and typically are) produced for attended and xated scene regions as a natural consequence of scene perception. We conclude that insensitivity to visual change during scene perception does not require any special explanation. Change insensitivity is a consequence of predictably fallible memory encoding and retrieval processes, and not the result of an architectural inability to create visual representation.
Due to a severe attentional impairment, patients with 'unilateral neglect syndrome' ignore half of their visual world. They are often not aware of objects or persons on their (usually) left side. However, it is becoming increasingly clear that neglect patients can still process visual information in their 'neglected' visual field implicitly. Through an experiment which capitalises on this ability of neglect patients to implicitly process visual information, I will try to address the question of whether visual perception and representation are overlapping or independent cognitive processes.
This talk is a practice for a seminar I will give at the Stroke Research Group at the Western General Hospital next week. The talk is thus originally meant for an audience of clinicians, rather than academics.
How well can monolingual English speakers translate Arabic text? Better than you'd expect. I will present some work into building computer aides that allow non-Arabic speakers to translate Arabic texts into English. I participated in a US Government sponsored competition for Arabic-English machine translation. My entry investigated the efficacy of manually correcting the output of a fully-automatic machine translation system, and the efficacy of allowing people to choose among possible English translations of Arabic phrases and construct their own translations. I will describe the methodology that I used, and give a high level overview of statistical machine translation. The talk will be brief (about 20-25 minutes in length).
This talk will be a practice for the NIST Machine Translation Evaluation workshop.
Psychology seminar, S1, Psychology, 7 George Square.
Professor Nelson Cowan (University of Missouri-Columbia): Time Limits and Item Limits in Working Memory
There are two ways in which working memory might be distinguished from the rest of memory: in time limits of the active representations it contains, and in item capacity limits. I will examine evidence that processes that have been assumed to be time-limited are often not limited in that way after all. Storage mechanisms with an apparent time limit may serve to back up other storage mechanisms with an item limit. Evidence on these points seems convergent across three diverse experimental topics that will be described: (1) two-tone comparisons, (2) delayed recall in amnesic patients, and (3) immediate recall for lists of words and associated word pairs. Underlying brain mechanisms for time- and item-limited working memory are proposed.
Tanzania is one of the few African countries which has a long history of a non-European language, namely Swahili, as a nation-wide medium of instruction. At the same time, more than 120 minority languages are spoken in Tanzania. When it comes to linguistic description of these languages, European languages have so far been used predominantly, thus making the results of such descriptions inaccessible to the majority of the language populations concerned.
In this talk, I will introduce a project of developing linguistic descriptions of Tanzanian Bantu languages in Swahili, which has been initiated by SIL International. This includes the development of a template covering the main grammatical features of Bantu languages, the teaching of this template to and data elicitation by mother tongue speakers of ten Bantu languages in an April 2004 workshop, and the write-up phase of the data in two sessions (November 2004 and January 2005). In particular, I will discuss problems of transfering linguistic terminology into Swahili, of introducing the template into the workshop setting, and of integrating the elicited data into a unified grammar publication. While it is too early to report any effects of these descriptions, as they have not been published and disseminated yet, I will give preliminary conclusions from feedback of workshop participants.
Psycholinguistics Conference, S1, Psychology, 7 George Square.
For details see the Psycholinguistics Postgraduate Conference website
Mood strongly influences the behaviour of animals and people. The control of mood is complicated. It has to be capable of reacting quickly to emergencies, adapting to long term changes, and synchronising itself appropriately to the various natural cycles of environmental changes such as day and night, tides, and the seasons. An engineer who was trying to design a mood controller capable of those kinds of things would naturally think of the classic simple PID control system of "spring and damper", because it can to be tuned to exhibit many of the required properties. Several would be needed to be combined to cover the various speeds and frequencies of response required.
These systems involve typical compromises and trade offs, such as being able to get faster response to change if you're prepared to tolerate some overshoot in response. These systems break down and malfunction when broken in various typical ways. It turns out that if we hypothesise a mood control system based a set of hierarchical mood controllers with different natural periods, then such a system has typical modes of breakdown which are reminiscent of the features of manic depressive illness. It also suggests some explanations of otherwise puzzling or paradoxical behaviour in manic depressive illness, reactions to certain drugs, etc..
This talk will explain the general features of manic depressive illness, the behaviour of PID controllers, and how the latter allows us to explain some features of the former. A preliminary review of some of the research literature from psychiatry and chronobiology which supports this hypothesis will be presented.
The very general idea on which this hypothesis is based in not new. It wouldn't surprise me if Kraepelin has a paragraph or two somewhere describing the basic idea rather well. The problem is that psychiatrists and psychologists rarely know much about control theory, and that mechanical and electroinc engineers, who do know about control theory, rarely know much about psychology and psychiatry,
The status of this research is that it seems like an interesting interdisciplinary idea which I hope to develop into a journal paper.
The main purpose of the present talk is providing an extrapositional account to cleft sentences in the Combinatory Categorial Grammar tradition (Steedman 2000) - more precisely in the Unificational Combinatory Categorial Grammar framework (Traat & Bos 2004). To date the few attempts on clefts in frameworks related to CCG always followed the expletive tradition.
The two main approaches to the structure of clefts are known as the extraposition approach and the expletive approach. The biggest difference between the two approaches is that the expletive approach considers the cleft pronoun to be just a dummy pronoun with no semantics, but for the extrapositional approach it is a normal referential pronoun. According to the expletive approach the sentences 'Sam wants Fido' and 'It is Sam that wants Fido' are semantically identical. We believe that these sentences differ in meaning: they are not interchangeably usable in a context. Therefore we take the position of the extrapositional approach. For us, while the above sentences have the same truth conditions, they have distinct semantics.
Clefts differ from their corresponding non-cleft counterparts in some important respects: they are presupposition-inducing syntactic structures (Delin 1990). If the above two sentences are negated, the presupposition 'Somebody wants Fido' is not cancelled in the cleft sentence.
In the expletive CCG analysis the copula has a special category in cleft sentences that differs from its category in the 'normal' copular sentences. In extrapositional approach the copula in clefts is considered just a regular copula, playing its usual linking role. Therefore, we cannot directly use the CCG syntactic categories used in the expletive account.
Having made the decision that the copula in clefts is a regular copula, we have to give the cleft pronoun somewhat special syntactic treatment. This is well justified, since semantically it seems to participate in two slots in the cleft sentence:
| Somebody (=it) wants Fido. | wants(x,Fido) |
| The person (=it) is Sam. | x=Sam |
The semantics we use for the cleft pronoun is straightforward: it only introduces a new discourse referent in the domain of the DRS. We show how due to the choice of its UCCG syntactic category the semantic value gets introduced in the two appropriate slots in the DRS.
To state that unmasking the true origin of cancer remains the Geneticists' quest for the Holy Grail is certainly not an unreasonable statement. Conventional intervention merely delays the inevitable prognosis of the major cancers. Could we be targeting the wrong cells?
Recent empirical evidence may serve as a compass at an old theoretical crossroads. We are leaving behind the scenario in which all cells within a tumour are deemed characteristically the same and arriving at one that reveals that only a few select bad seeds may initiate a tumour. These bad seeds are cancer stem cells.
Traditional treatments have relied on the ability to target and indiscriminately kill cycling/proliferating cells within the tumour. According to the cancer stem cell model, this method of intervention would fail to eradicate the minority population of tumour-initiating cells which, like their normal stem cell counterparts, may remain in a quiescent state removed from the cycle of cell division.
The elucidation of the cancer stem cell is testament to a common flaw in molecular biology and genetics research - the confusion between cause and consequence. In this respect we have been missing the cancer culprit due to a misrepresentation of cause and symptom. Cancer therapies of the future may finally serve to target its cause.
Most approaches to language understanding can be described as either "knowledge-lean" statistical models or "knowledge-rich" logic-based models. This division is unfortunate, as it leads to a severe limitation: despite obvious benefits, it is difficult to use statistics and knowledge bases alongside one another. In this talk, I will discuss the "knowledge-greedy" approach to language understanding, which overcomes this limitation and allows the combination of statistics and logic. My goal is to review work on knowledge-greedy reasoning, analysing the benefits and difficulties in using this work for language understanding.
What Is EI anyway? Trait and Ability theories
Testing EI
EI Findings
Issues with EI: Nomological Net, Predictions, Experimental Evidence
Shameless Plug: My Research
References
Matthews, G., Zeidner, M., & Roberts, R.D. (2002). Emotional Intelligence: Science and Myth. Cambridge, MA. MIT Press.
Petrides, K.V., & Furnham, A. (2001). Trait emotional intelligence: Psychometric investigation with reference to established trait taxonomies. European Journal of Personality, 15, 425-448
Van Rooy, D.L., & Viswesvaran, C. (2004) Emotional intelligence: A meta-analytic investigation of predictive validity and nomological net. Journal of Vocational Behavior, 65, 71-95
Dyslexia is a term used to refer to individuals with an average to high intelligence rating, but a specific impairment in their reading ability. Despite the fact that this disorder affects a significant part of the population, its exact cause remains the subject of heated debate. In this talk I will outline two of the main theories of dyslexia; the phonological and magnocellular hypotheses, and report on a study I conducted for my Masters thesis in the summer. Using the picture-word interference task, the aim of the study was to examine whether dyslexics experience semantic as well as phonological interference as a way of investigating the existence of a deficit other than phonological problems in a sample of the dyslexic population. The results suggest that dyslexics experience increased difficulty relative to controls when trying to process two consecutively presented stimuli. These findings are in line with a domain general magnocellular deficit.
There are two contrasting ways of incorporating information into a database. The first is manual curation which relies on human expertise. Its advantage is its high quality, its drawback is lack of scalability to high volumes of incoming data. The alternative is automated information extraction. It can deal with large volumes of data, but the accuracy of the data obtained is uncertain. The Nuclear Protein Database (NPD) is a hand-curated database which contains information on vertebrate proteins located in the cell nucleus.
I will first talk about the interface I developed last year for my Master by Research, which is now in use. I worked on automatically classifying articles from the biological literature into two classes: those which should be curated into the NPD and those which should not.
I will then talk about my ideas since the start of my PhD. I am now working on information extraction and named entity recognition on full text papers.
Evidence for Theory of Mind-- the capacity to perceive mental states of self and others (Baron-Cohen, Leslie & Frith, 1985)-- has been primarily sought in children's use of mental verbs and their performance on a linguistically and cognitively rather complex false-belief task. As a consequence, mental states, such as beliefs, are said to be achieved only around the age of three or four (Baron-Cohen, Leslie & Frith, 1985; Wellman & Gelman, 1992; Gopnik & Meltzoff, 1994; Leslie, 1994). Arguably, however, there are earlier signs of Theory of Mind in children's more ordinary linguistic accomplishments, such as verb argument realization in everyday conversation. In this talk, I will review the results of a study on spontaneous speech data from four monolingual children acquiring Inuktitut (2;0-3;6). I will show that a choice of argument form to represent a referent not only reflects the child's knowledge of syntax and evaluation of the discourse-pragmatic context, but more importantly, it also requires and is determined by the child's ability to assess the knowledge state of the interlocutor. As such, this knowledge can be viewed as a substantial milestone toward understanding the mental world of another person, constituting one of the precursors of Theory of Mind.
I will be talking about our on-going work on recognising textual entailment (TE). The task is to decide, given two text fragments, whether the meaning of one is entailed (can be inferred) from another. For example, the first sentence of this abstract entails "Our work on TE is far from finished". Entailment is akin (but more general) to paraphrasing and is believed to be a common core to many semantic-oriented NLP tasks, including automatic Question Answering. As a result of a detailed analysis of a small corpus of entailment pairs, we propose a two-stage method for recognizing entailment: first, identify similar entities in the text fragments, and second, check whether the syntactic structures enbedding the matching entities are "similar". We use corpus- and Web-based methods for both subtasks.
Time permitting, I will also briefly present our ideas for a completely different project: mining, organizing and answering FAQs from the Web. This fascinating task requires techniques from data mining, NLP and semi-structured data retrieval, and can potentially help you with questions like "How do I feed a baby hedgehog?"
My paper has two main themes. The first is methodological: I examine how we might study consciousness from a first-person and second-person point of view. I will explain how reflection on our own conscious experiences and those of others can contribute to a scientific investigation of the conscious mind.
My second aim will be to briefly describe some of the results this approach to studying consciousness delivers. Here I shall draw on the work of the late Francisco Varela and more recent work by Alva Noe and Kevin O'Regan. I will show how in both cases this research dovetails with what reflection tells us about our own experiences and the experiences of others.
- One way of selecting one's acts is through answering the following two questions: 1) Which human being(s) shall I treat as an end? 2) What is it to treat a human being as an end?
- Assumptions. This talk assumes that one is indeed selecting one's acts through answering questions (1) & (2), and that one is entitled to put to one side any other possible ways of selecting one's acts. This talk also takes it to be the case that the answer to question (2) is, roughly, that to treat a human being as an end is to allow one's acts to be guided by their bearing upon him (if any). Then the question will arise, 'What bearing shall I seek to have upon him?' One possibility would be to ensure that one does not inflict harm on him, and, in addition, to try to promote his happiness & virtue. However there is not time to discuss question (2) today.
- This talk aims to answer question (1). The arguments in the literature which attempt to answer this question are not very satisfactory. This talk puts forward a better way of going about answering this question, which, if one follows the argument, will lead one to always treat every human being as an end. The argument goes as follows.
- Firstly, either one treats at least one human being as an end, or one does not select any acts. This is because, without treating at least one human being as an end, one would have no answer to question (1). Given that one is selecting one's acts by answering questions (1) and (2), it follows that, without an answer to question (1), one does not select any mental or physical acts at all. Thus if one rejects not selecting any acts, then one treats at least one human being as an end.
- Secondly, when trying to answer the question 'Which human being(s) shall I treat as an end?' one has a choice of two paths, (A) or (B):
If one takes Path (A) then one sorts human beings into two or more categories, for example the category of those whom one always treats as an end, and the category of those whom one does not always treat as an end.
If one takes Path (B) then one does not sort human beings into two or more categories, but places each in the same category; in other words, treats each and every human being similarly with respect to the question as to whether or not to treat him/her as an end.
With respect to Path (A), it can be shown that there are no non-arbitrarily selected criteria which can govern one's act of sorting human beings into such categories. Therefore one's sorting of human beings into such categories can only be done arbitrarily. Clearly, if one decides arbitrarily who to treat as an end then one acts arbitrarily (given that who one treats as an end determines how one acts). Thus if one rejects acting arbitrarily then one rejects sorting human beings arbitrarily into two or more categories, and instead takes Path (B).
- Thirdly, we saw earlier that if one rejects not selecting any acts, then one treats at least one human being as an end, and we have just seen that if one rejects acting arbitrarily then one treats every human being similarly. Conjoining these two points, leads to the conclusion that if one rejects not selecting any acts, and if one rejects acting arbitrarily, then one treats every human being as an end.
- Fourthly, the arguments for rejecting acting arbitrarily will be outlined.
- Finally, objections which might be raised on behalf of the egoist and the dice-man, will be rebutted.
Primitivism is an emerging account for secondary qualities (such as colour) as primitive intrinsic properties of objects. According to it, secondary qualities are not phenomenal properties of our visual field, but are constitutively connected to them (McGinn, McDowell, Campbell). Further, secondary qualities are instantiated in the physical world. Primitivism's appeal consists in its apparent phenomenal adequacy.
However, Chalmers [2004] argues that secondary properties as primitivism conceives them cannot possibly be instantiated in the world. For, given an agent with inverted spectrum, an object would have to be, say, as red and as green appears to us, at the same time. But no object, says Chalmers, can be both ways.
In this paper I respond to Chalmers, offering support for the primitivist view. My starting point is Aristotle's perceptual theory which gives an account of secondary qualities that respects the primitivist intuitions. I show how this account can overcome Chalmers' difficulty.
One of the key challenges facing artificial life researchers, as well as biologists, is to explain the origin of living organisms from a non-living environment. Furthermore, in order to build artificial evolutionary systems, we would like to know how to produce highly evolvable systems, in which agents can control and exploit their environment in unlimited and increasingly complex ways (i.e. we would like them to be capable of open-ended evolution). In this talk, I will give a brief critique of some previous work in this area, and introduce a new perspective to building artificial open-ended evolutionary systems. This involves removing the representational distinction between organism and environment and treating them as a single dynamical system. I will introduce a simulation platform, called EvoCA, to demonstrate these ideas. Using the results of experiments with EvoCA, I will discuss the implications of this modelling approach on topics such as the origin of symbols and semantics (as entailed in the genotype-phenotype relationship) in a system governed by physical laws, and the evolution of new sensory and effector capabilities. I will conclude by mentioning some of the main research questions that arise from this approach.
Sudan, the biggest African country, is divided sharply into two distinct areas, both in geographical area and ethnic group, and cultural systems. It is noted that the Northern part of the Sudan is occupied by a hybrid Arab race that is united by their one language, one culture, and one religion and they look to the Arab world for their cultural and political inspiration. The people of Southern Sudan, on the other hand, belong to the African ethnic group of East Africa. They do not only differ from the hybrid Arab race in origin, arrangement and basic systems, but in all conceivable purposes. The two regions have been at war for around thirty years, but the month of May 2004 saw a landmark in the history of Sudan when the government and the southern opposition signed key peace protocols on the 26th in the Kenyan town of Naivasha. One of the peace protocols contains a significant section on language policy. It has become like a fact of history that politicians turned a deaf ear to the issue of linguistic diversity in Sudan; however, the language policy in the peace accord may function as a reference to the fact that the question of language is there at the heart of the struggle to achieve a peaceful co-existence.
In this talk I will attempt to critically analyse the language policy statements. The analysis will be carried out with reference to the existing language policy, with the aim at arriving to a better understanding of the relationship between the language policy and the social relations involved. The primary purpose is to uncover the implications of taking up one interpretation position of the language policy statements over another by critically engaging with questions of power, ideology and inequality. I will adopt a problematising practice that questions the components of the current reality that are taken for granted and touches on the official accounts on how it came to be the way it is.
In this talk we present a multi-layered Question Answering (Q.A.) architecture suitable for enhancing current Q.A. capabilities with the possibility of processing complex questions. That is, questions whose answer needs to be gathered from pieces of factual information scattered in different documents. Specifically, we have designed a layer oriented to process the different types of temporal questions. Complex temporal questions are first decomposed into simpler ones, according to the temporal relationships expressed in the original question. In the same way, the answers of each simple question are re-composed, fulfilling the temporal restrictions of the original complex question. Using this architecture, a Temporal Q.A. system has been developed. We focus on explaining the first part of the process: the decomposition of the complex questions. Furthermore, it has been evaluated with the TERQAS question corpus of 112 temporal questions. For the task of question splitting our system has performed, in terms of precision and recall, 85% and 71%, respectively.
This talk will be given at ACL 2004.
Although experiences of apparent communication between people who are physically separated (and are not using mobile phones etc!), are not uncommon, the "telepathic" nature of such experiences is often highly questionable. Most such experiences can be easily explained as the combined effect of coincidence and selective memory, i.e. we tend to remember impressive incidents when our internal experience (e.g. thinking about someone) coincided with an independent external event (e.g. them calling us on the phone), but not the unremarkable (and far more numerous) instances when our internal experience and external events did not share any meaningful association.
Although some reports of such experiences do seem suggestive of some form of genuine communication, the possibility of error, fraud or of an alternative less exotic route of communication being involved cannot be ruled out, and the only way to tackle the question methodically is through laboratory experiments.
A large variety of such experiments has been conducted over the past 50 years or so, but I will be focusing on studies using measures of psychophysiological activity (EEG and EDA), giving an overview of past findings and then present my own PhD research so far. This aimed to explore the possibility of psychophysiological interactions between (physically isolated) pairs of strangers and pairs of people in an empathic relationship, using an experimental protocol of photic stimulation and EEG measurements.
The vision behind Ken Yersel is a network of community groups around Scotland - and hopefully beyond - who use simple discussion games to consider socially important issues, and then share their results through a website. Our method is inspired by a Freirian based community education approach and the McCall method of Philosophical Inquiry, using argument visualisation research performed by the Archelogos project.
After a particular group discussion, an argument map of the relevant issues is created and published on the web. This can then be refined together with other groups to paint as complete a picture of all the important facets of the issue in question as is possible. Through organisations such as the Scottish Civic Forum or IDEAS the final discussion map would be presented to decision makers who would be asked to justify policy decisions with respect to the issues set out by the community groups.
The project is a collaboration between Archelogos at the Philosophy Department, the Edinburgh Adult Learning Project as well as other community education groups throughout Scotland.
This talk is concerned with the evolution of socio-technical systems. Although it is possible to identify an evolutionary space for socio-technical systems, the methodologies that address the evolution of socio-technical systems are still patchy. Moreover, it is still challenging to address multidisciplinarity in modelling methodologies. This talk addresses the heterogeneous modelling of the evolution of socio-technical systems. The analysis of a case study highlights how the combination of diverse models allows the characterisation of requirements evolution.
This is a practice for a conference talk to be presented at the Workshop on Interdisciplinary approaches to achieving and analysing system dependability, DSN 2004.
Types are often used to control and analyze computer programs, but useful types are both difficult to supply manually and difficult to generate automatically.
System E is a new type system which uses expansion variables (introduced recently by Kfoury and Wells) to support flexible automatic generation of type information.
Expansion variables make it easy to refine the analysis of program fragments on a demand-driven basis, by allowing insertion of typing rules uses in nested positions.
System E provides flexibility and expressiveness via intersection types and the ! operator. Intersection types provide polymorphic/polyvariant analysis, and the ! operator is used to distinguish between linear (used exactly once) and non-linear types.
In this talk I will present findings from two experiments conducted in Hebrew on the acquisition of relative clauses by children (mean age 4;5). Relative clauses have been extensively studied within the field of language acquisition both because of the apparent difficulty children have in comprehending them (Sheldon, 1974) and because of their relevance to theories of syntactic development and breakdown. Many studies have shown that children have special difficulty with object relative clauses (show me the doctor that the soldier is drawing). In addition, an intruiging discrepency betwen production and comprehension wherein production precedes comprehension is also reported. One prominant account for patterns of relative clause acquisition is that children lack movement abilities(the ability to connect a moved element with its trace)thus conceptualizing the difficulty as a developmental stage. However, a reanalysis of previous studies in addition to adult data showing that adults also have difficulty with relative clauses seem to weaken this claim. The presented experiments tested both production and comprehension of relative clauses, with and without resumptive pronouns (which are not supposed to involve movement in Hebrew). Overall, the results confirmed the need to look for explanatory factors other than movement to account for relative clause acquisition, and underline the importance of close analysis of error patterns. I shall discuss some possible interpretations, focusing on the possibility of developmental sub-stages in the acquisition of relative clauses and the potential role of processing limitations in addition to structural limitations in accounting for both the individual variance revealed by these studies and the lag between production and comprehension.
Generalized quantifiers like "all beer" or "most nuts" are semantically well understood yet we know little about their neural representation. Our model of quantifier processing includes a numerosity device, operations that combine number elements, and working memory. Semantic theory posits two types of quantifiers: First-order quantifiers identify a number state (e.g. "at least"), and higher-order quantifiers additionally require maintaining a number state actively in working memory for comparison with another state (e.g. "less than half"). We used fMRI to test the hypothesis that all quantifiers recruit inferior parietal cortex associated with numerosity, while only higher-order quantifiers recruit prefrontal cortex associated with executive resources like working memory. First-order and higher-order quantifiers both recruited right inferior parietal cortex, suggesting that a numerosity component contributes to quantifier comprehension. Moreover, only probes of higher-order quantifiers recruited right dorsolateral prefrontal cortex, suggesting involvement of executive resources. We find parallel findings from patients with neurodegenerative diseases: corticobasal degeneration (CBD) patients with parietal lobe disease are significantly impaired on all quantifiers and frontotemporal dementia (FTD) patients with frontal lobe disease are only impaired on higher-order quantifiers. Our findings are consistent with a large-scale neural network centered in frontal and parietal cortex that supports comprehension of generalized quantifiers.
Linguistics Postgraduate Conference, B9, TAAL, Adam Ferguson Building.
For details see the Linguistics Postgraduate Conference website
Recent scientific breakthroughs and conflicting ethical and religious beliefs have contributed to propelling stem cell research into the political crosshairs. The therapeutic potential of undifferentiated human cells has given rise to a deeper exploration into the mechanisms governing biological control during early human development.
In cell-based therapies, differentiated cells derived from human embryonic stem cells (hESCs) would be transplanted or injected into the patient where they would integrate into the target tissue and thereby restore organ function or prevent or slow further deterioration. Such cells include neural cells, cardiomyocytes and pancreatic islet beta cells, regeneration of which may serve to alleviate conditions such as Parkinson's, heart disease and diabetes respectfully.
Here I give a general overview on the basic biological principles surrounding the human embryonic stem cell and the potential derivation of any cell lineage of the human body. The talk will also touch on the regulatory measures which have been deployed, encompassing the many ethical considerations within this field and how and why this research is so permissive within the UK. A significant cross-disciplinary effort will no doubt be required to advance stem cell research into applied clinincal settings.
Human languages exhibit a combination of computational features that make them unique systems of communication in nature: large and learned lexicons, combinatorial phonology, compositional semantics, and hierarchical phrase structure. In the field of evolution of language controversies have often focused on the complexity of these computational mechanisms. These controversies include debates about innateness, whether or not language was exapted, if it is the result of a few or many mutations and if it increased in complexity over evolutionary time (see e.g. Pinker and Bloom, 1990, and the many peer commentaries and the authors' response in the same issue). We analyze these debates and find that at their core they rely in varying degrees on two implicit assumptions: (i) that complexity in the computational machinery for processing language is difficult for evolution to achieve and/or that (ii) that complexity is itself a trait which can be selected for or against.
Out of the many possible ways of studying computational complexity, formal linguistics has primarily been concerned with situating natural language processes and formalisms on various computational hierarchies. By far the most studied of these is the (extended) Chomsky Hierarchy. We ask the questions: how do the two assumptions outlined above fare when analyzed under this notion of complexity, and how does this apply to the debates in the field? Such a formal definition would potentially resolve conflicting intuitions about complexity (exemplified e.g. in Lewontin's and Piatelli-Palmarini's commentaries on Pinker and Bloom, 1990).
We argue that complexity in the automata theoretic sense is in fact very common in natural systems. We find it plausible that genes can code for systems with small numbers of elements interacting with simple rules. There is increasing evidence that these sorts of systems are in fact often computationally universal (e.g. Wolfram, 2003). Furthermore, certain classes of neural network models have been shown to be Turing equivalent (Siegelmann and Sontag, 1991), and capable of efficiently encoding phenomena such as hierarchical phrase structure (Pollack 1988). We suspect that the reality is that brains in many kinds of animals are already implementing algorithms and computations which are sufficiently complex to represent and process language in the strict automata theoretic sense.
Furthermore, we go on to argue that these grammars and automata are not well suited to be used as phenotypes in biological models. They do, of course, expose interesting differences in grammatical classes on the hierarchy. For instance, the word recognition, or parsing problem increases in time complexity as one makes certain moves up the hierarchy. Likewise, differences in the hierarchy can be understood in terms of increasing relaxation of memory limitations, e.g. finite to stack based to stack based with less restrictive push procedures, etc. But it is difficult to see how these differences satisfy various evolutionary constraints or can affect fitness. We argue that instead of looking at these formalisms in terms of their place on the hierarchy we must look deeper at the properties of language that they are meant to abstract.
We summarise that it is not the physical constraints of the general neural architecture that restrict natural language to a specific complexity class. Rather, the requirements of learnability and population coherence as well as the interface conditions of interpretability and producability under realistic time and noise constraints choose specific classes of computational mechanisms. These mechanisms restrict any language that is to survive either cultural or genetic transmission. We discuss the implications of this for the debates outlined in the introduction and reach some general conclusions. For instance, is it theoretically useful to describe the evolution of language as climbing the Chomsky hierarchy? (as do, e.g. Hashimoto and Ikegami, 1996). Finally, we conclude that while the Chomsky hierarchy is a bad model of phenotypic complexity, it is a very good model of language. This suggests a way of rescuing it as a tool for evolutionary theory.
Hashimoto, T. and Ikegami, T. (1996). The emergence of a net-grammar in
communicating agents. BioSystems 38, 1-14.
Pinker, S. and Bloom, P. (1990). Natural language and natural
selection. Behavioral and brain sciences 13, 707-784.
Pollack, J. B. (1988). Recursive auto-associative memory: Decising
compositional distributed representations. In: Proc. of the Tenth
Annual Conference of the Cognitive Science Society. Lawrence
Erlbaum.
Siegelmann, H. and Sontag, E. (1991). Neural networks are universal
computing devices. Technical Report SYCON--91--08, Rutgers Center
for Systems and Control.
Wolfram, S. (2002). A New Kind of Science. Champaign, IL, U.S.A.:
Wolfram Media.
The way microorganisms sense and move about their world is important to us. If we can understand these processes, then we might be able to manipulate the environment to guide and direct the movement of pathogenic microorganisms, for example, away from the site of an infection. Our understanding of the microbial world is largely a chemical one; science has focused most intensively on the way that microorganisms sense chemicals in their environment, in much the same way that we smell chemicals in ours. However, smell is not our only sensation, and the same is true for microbes: they can detect gravity, some can detect light, and most can detect electric fields. We believe that electric fields are an important source of information for microorganisms and that they sense electric fields in much the same way that we sense light fields. We are testing for this by generating electric fields of different properties in order to mimic the electric fields that microorganisms would experience in the wild, and then we observe their response. We are creating virtual electric worlds in order to understand microbial perception of electric fields and their behaviour in them.
In this talk I will discuss three topics. The first is Lee's General Tau Theory of Sensorimotor Control. In short, this theory predicts the movements of an organism towards a goal state. This is achieved by determining the organism's changing perception of the goal-state as it moves towards it. Secondly, I will touch upon Activity Theory (of which General Tau Theory can be considered to be a piece) and the apparent 'scalability' of living processes. Finally, I will delve into my own work that applies these ideas to the movements of the single-celled ciliated protozoa, Paramecium caudatum. I will discuss work underway here to define and describe the very complex movements of these animals, and we will see whether or not they respond to their environment in a manner comparable to the way we respond to ours.
I plan to talk initially about my PhD work which is looking at active perception in robotic agents, but also hopefully broaden it into a discussion on psychological models of attention. For this latter part I'll be hoping for audience input as my ideas are not well formed.
The aim of my PhD work is looking at a class of problems know as Partially Observable Markovian Decision Process (POMPD) and trying to learn them using reinforcement learning. POMDP are known to present problems for reinforcement learning, which can solved in the general case by using memory of previous state and actions to compliment the agent's observations of the world. However my approach differs in that I am considering agents who attempt to resolve ambiguities in the world by directing their sensory input in order to obtain observations that are not ambiguous.
Before we can address ourselves towhat we think concerning the relationship between 'intuition' and 'creativity', it seems appropriate to address the question tohow we think. Prompted originally by Bergson's theory of 'intuition' in Creative Evolution (and Russell's heckling, in A History of Western Philosophy), the paper draws on Saussure's linguistics and connectionist models of cognition as articulated by Andy Clark to propose a fruitful ordering of symbol, sign, pattern and pathway.
The job of an engineer is create a design that translates from a problem specification to an implementation. The design process is difficult because this mapping is neither simple nor one-to-one. If the specification changes, then the translation must be redone; this is why software reuse and maintenance is so difficult, and why software engineers are forced to constantly rewrite large amounts of code from scratch with each new revision of a product.
Biological organisms face a similar problem. An organism has a phenotype: its physical body, which is completely described by its genotype: the organism's DNA. What makes evolution interesting is that genomes are evolvable designs -- they are capable of constant, incremental change, and do not seem to suffer from the code maintenance issues that afflict software development.
I will discuss some recent advances in software engineering that are designed to alleviate the code maintenance issue -- such as domain specific languages, aspect-oriented programming, and feature composition. These ideas are directly related to certain biological processes -- such as sexual crossover, genetic switching networks, and morphogenesis. I will further show that the design process in both cases can be modeled as a knowledge representation problem, and I will use this model to explain why current programming languages are poor tools for design, and what we can do to fix them.
According to Church's thesis, all models of computation are equivalent in power. However, the everyday intuitions about computation itself vary widely among disciplines, leading to difficulties in communication. For example, how can the mind be like a computer, if we have no grasp of what kind of thing a computer actually is? Brian Cantwell Smith has analyzed the similarities and differences of these construals, such as Formal Symbol Manipulation and Effective Computability in the Foundations of Computing. I will describe briefly his findings and how these effect the interface of syntax and semantics, the semantics of natural language, and the World Wide Web.
The traditional approach to cosmology is based on Einstein's theory of relativity, where one assumes that spacetime is a four-dimensional manifold. During the past two decades there has been much progress in combining relativity with quantum mechanics, leading up to the so-called string/M-theory. There is a natural cosmological model based on string theory, known as the brane world scenario. The Universe is assumed to be a membrane-like object that lives in a higher-dimensional bulk - much like E. M. Abbot's Flatland. This scenario may offer a solution to the hierarchy problem of the Standard Model of particle physics - the seeming disparity between the electroweak energy scale and the quantum gravity scale - and new mechanisms for cosmological inflation.
For the configuration of large computing sites to be effectively managed, it is important for the configuration to be specified using a language which provides a higher level abstraction of the configuration data. Having the configuration of all machines specified in a declarative language provides greater scope for reasoning about groups of machines and the validation of the configurations.
Current configuration management systems are limited in terms of scalability, validation, support for devolved management, and specification of high level and dynamic properties.
The talk will outline the limitations of the current configuration languages and discuss language features which may be used to overcome these limitations. These include constraint satisfaction, aspect composition, use of a well defined semantics, and peer to peer communication between nodes.
The Unified Modeling Language (UML) is a semi-formally defined standard for modelling the design of object-oriented software systems. It provides different diagram types each of which is used for modelling a particular aspect of the system. There exist many design tools which help the designer to draw UML diagrams but there is very little support for the design activity as such. The hard task of choosing a design that is consistent, fulfils the requirements and contains enough detail is still almost completely left to the designer.
I will talk about how formal games can be used as concept for UML design tools which allow the designer to explore and evaluate different design possibilities. In contrast to the two-player games that are traditionally used for verification, the rules of the game are not fixed in this approach but change when the design is modified during a play. I will discuss some of the problems that arise with games and UML, and sketch how prototypical tool support could look like.
I will talk about what I have been doing over the last couple of months here in Edinburgh, the main focus of which has been tutorial interviews with undergraduate subjects on Wason's selection task. This is a very well-known task in the psychology of reasoning literature, and has formed the basis for a lot of theorising about reasoning processes, due to the somewhat surprising and extremely robust results which it generates. Most of this theorising agrees that student subjects are highly irrational.
Stenning and van Lambalgen have recently proposed that the observed results are best understood as student subjects' rational but not always successful struggles to interpret the conflicting information they have about the task and materials.
My work here has largely followed on from their proposals, using tutorial dialogues to elicit students' interpretations. I will discuss some motivations for this approach and some preliminary observations from the work so far.
The purpose of automated Question Answering (QA) is to retrieve an answer to a question expressed in natural language. Whereas Information Retrieval provides a set of documents covering the topic expressed by a keyword query, QA focuses on retrieving fine-grained and exact answers. The Text REtrieval Conference (TREC), organised each year by NIST, has offered since 1999 a specific track to evaluate large-scale open-domain QA systems. For TREC 2003, three kinds of questions were distinguished: factoid, definition and list questions. While the first type expects only one exact answer, the latter two require an advanced analysis of the potential answer space to evaluate which potential answers fit the expected definition (``What are fractals?") or are part of the answer list (``Who were the great opera tenors of the twentieth century?"). Providing multiple answers is challenging, especially when the source of such answers is untyped and heterogeneous free text.
In this context, we are developing a new technique based on information fusion to allow answer comparison and improve answer selection. The main use of information fusion in language technology has so far been in summarisation, to organize information coming from heterogeneous sources. Here we show that information fusion can also improve automatic question answering (QA). We describe an experiment on TREC where-questions and suggest directions for future work.
This is a practice for a conference talk to be presented at the 2nd CoLogNET-ElsNET Symposium.Recent developments in our understanding of the underlying mechanisms of ageing, and practical successes of intervention in those mechanisms in animal models have, for the first time in human history, realistically opened the possibility of dramatic increases in life-span, amounting possibly to functional immortality in humans. This talk will briefly outline the practical research underpinning this assertion, and will then focus on the ethical, social and legal issues to which this prospect gives rise.
My aim is to expose a pernicious attitude that plays a significant role in the discourse surrounding the animal issue.
It is commonly held that mastery of language justifies the ascription of certain mental capacities to humans. Language is considered a necessary condition for certain mental states, such as thoughts and beliefs. However, we are faced with a conceptual problem when we turn to animals: they lack language, and thus, supposedly, the necessary conditions for mental capacities such as thoughts and beliefs. The idea is that the capacity for these states presupposes language. Such an argument is based on a priori reasoning.
By a priori reasoning I refer to the methodology that puts the empirical world "on hold" and purports to establish the conditions for X from pure reflection, without due reference to empirical evidence. It regards empirical evidence solely as raw material for doing philosophy, but not as an ongoing point of reference.
I contend that the prevalence of a priori reasoning, which denies certain mental capacities to animals, is unjustified. I trace the origins of this reasoning to two sources: the thesis of propositional attitudes and its manipulative authority, and a lack of commitment to empirical evidence. The resulting conclusion, namely that animals cannot have thoughts and beliefs, is derived from a philosophical misconception founded on an anthropocentric approach to the animal issue. That is, it examines animals with tools that apply, for the most part, to humans. In this respect, the familiar charge of unjustified anthropomorphism against pro-animal philosophers should be applied to philosophers who argue against animals on that same anthropocentric basis. By starting within a human paradigm they allow empirical evidence to be ignored. There would be no need, for example, to make observations about an animal being either disappointed or satisfied with the food she gets, as observations of animal behaviour are pointless from within the human paradigm. That is, anthropocentric a priori reasoning fuels the dismissal of the empirical evidence.
In dialogue, certain aspects of the language we produce appears to reflect what we have just heard. This phenomenon is known as priming and may be observed in our phonology (our accent changes according to who we are speaking t o), conceptual representations, lexical choice (e.g., lounge vs. living room), and syntax. Syntactic priming refers to the phenomenon of using a particular syntactic structure given prior exposure to the same structure. This behaviour has been observed when speakers hear, read or write sentences. I will present evidence for syntact ic priming of active vs. passive voice in dialogue. For instance, when a speaker has just heard The judge being dragged by the cowboy, s/he is more likely to produce a description like The nurse being bitten by the boxer rather than The boxer biting the nurse.
Generally speakers align with each other, but I will also present data showing what happens when one speaker refuses to align with another at the lexical level. The experiment compared the level of syntactic priming when the interlocutor aligns versus when the interlocutor refuses to align (and instead uses a different word to that used by the original speaker).
Naturalism is a view in philosophy that says everything that exists is a part of the natural world. Consciousness is a problem for a naturalistic philosophy. We are conscious creatures; hence, there is no doubt that consciousness exists, but just how consciousness can be a part of the natural world remains very much a mystery.
Naturalist philosophers have generally worried about two questions concerning consciousness: (1) Why is there something rather than nothing it is like for a subject to have experiences? In others words, why do creatures like us enjoy conscious experiences? (2) Why is it there is something in particular it is like to have experiences? This is to ask, why creatures like us enjoy the kinds of conscious experiences we do?
I will propose an answer to the first of these questions by arguing that there is something it is like to have an experience when a subject is in a state that represents, among other things, the subject having this experience. There is nothing it is like to have an experience when the subject is not in a state that represents the subject having the experience. What I shall label the feeling of ownership is the state a subject is in when she is experiencing herself having a particular experience.
Having introduced this account of what it is to be conscious, I will propose a naturalistic account of this feeling of ownership - the experience a subject has that s/he is having an experience. I will make some suggestions about how the feeling of ownership might be understood as a kind of bodily consciousness. The existence of this kind of bodily consciousness, I shall claim, is strongly supported by ecological and motor theories of perception that stress the close relationship between perception and action.
This talk discusses that diagnosing and taking appropriate action based on students' motivation, effort and other affective characteristics is an important part of effective Intelligent Learning Environments. On the other hand, issues relating to affect are really intertwined. For instance, a student's affective state is influenced by the whole environment and didactical situation. By employing a web-based environment (WALLIS) that has been integrated with the teaching and learning process in the School of Mathematics we were able to limit some factors that would influence our experiments or at least be aware that they do so. We will present preliminary results from observations on student's interactions which provide insight for the design of an affective component and discuss ideas on interface changes that could provide more information to the user model.
The editing of moving-images involves the splicing together of a series of moving images in a way that creates an illusion of continuity. This continuity can exist at many perceptual levels including narrative, time, and space, but all are a direct result of the difference between two visual scenes. The cognitive mechanics by which this illusion is created are only beginning to be understood even though the "rules" of continuity editing are taught throughout the film industry.
The goal of my Ph.D. research is to investigate the cognitive foundations of continuity editing by concentrating on the immediate effects of an edit on eye movements, attention, and memory. Where does the Road Runner go when he runs off the screen? Why does watching MTV with the sound turned off hurt your eyes so much?
In this talk I will present a recent behavioural study in which I have operationalised three temporal and spatial factors of one type of continuity edit, a match-action shot, in which the direction of a principle action is preserved across a series of shots. The goal of this experiment was to show that the best editing conditions where those that respected the saccadic eye movements necessary to re-fixate the principle object after an edit.
We present a novel ensemble learning approach to resolving German pronouns. Boosting, the method in question, combines the moderately accurate hypotheses of several classifiers to form a highly accurate one. Experiments show that this approach is superior to a single decision-tree classifier. Furthermore, we present a stand-alone system that resolves pronouns in unannotated text, by using a fully automatic sequence of preprocessing modules that mimics the manual annotation process. Although the system performs well considering its simple architecture, further research is needed to make it effective for tasks such as question answering and text summarisation.
Disfluency has been described as a strategic device for signalling to a listener that the speaker is committed to an utterance under construction. It has also been described as an automatic effect of a cognitive burden, particularly of managing speech production during other tasks. To assess these claims, we crossed a listener feedback variable with a time-pressure variable in a variant of the map task. 24 subjects were tested in 4 conditions: a baseline untimed monologue condition and 3 other conditions in which visual feedback of a follower's gaze or time-pressure, or both were added. Both feedback and time-pressure affected the nature of the speaker's performance overall. Disfluency rate increased when feedback was available, as the strategic view predicts, but only deletion disfluencies showed a significant effect of this manipulation. Both the nature of the deletion disfluencies and the information which the speaker would need to acquire in order to use them appropriately suggest ways of refining the strategic view of disfluency.
A key issue in second language (L2) research is to determine how
syntactic re
presentations and processing differ between native and non-native
speakers of a
language. I will report a study which used a syntactic priming task to
investiga
te this. 'Syntactic priming' refers to the tendency people have to
repeat the ty
pe of sentence construction used in an immediately preceding, unrelated
sentence
. This effect suggests the existence of mental representations for
particular sy
ntactic constructions, independent of particular words and meanings.
The current study used a dialogue task to look at priming of actives
and passive
s in Spanish. The subjects were natives, intermediate L2 speakers and
advanced L
2 speakers of Spanish. Results demonstrated a significantly stronger
priming eff
ect in the L2 speakers compared with the native. The advanced group
also showed more priming than the intermediates. I would like to
discuss the implications of these results as well as ideas for future
studies.
When loanwords are added to a language's lexicon, their phonetic form is often altered considerably as compared to the donor language. This is due to conflicts between the sets of possible sounds in both languages as well as the possibilities of combining these sounds (phonotactics). In Russian, a large number of consonant combinations, clusters, are possible (e.g. /fp-/, /dv-/) at the beginnings of words, whereas in English the set of initial clusters is much more restricted. As a consequence, when native speakers of English are asked to repeat Russian words, they adapt these clusters in a number of different ways, i.e. by inserting additional vowels, deleting one of the consonants or changing the consonants such that an acceptable cluster in English is the result (e.g. /fp-/ to /sp-/). These adaptations cannot be predicted by simply comparing the phonological systems of the two languages, but factors outside this must be taken into account. In this talk I will show what roles factors such as frequency, similarity and gradient acceptability play in determining which strategies of adaptation will be used.
Our sense of smell is possibly the least understood of all our senses. Despite much study no stable topographic map akin to the one seen in the visual system has been discovered. The current study aims to discover whether olfactory information could be encoded in the firing patterns of a specific class of olfactory neurons. Recordings were taken from the olfactory bulbs of female rats under several experimental protocols. A cellular model was adapted to allow it to show the behaviours exhibited by cells in vivo under each protocol. The model is being used to test how the firing patterns of these cells may be regulated.
Eight second year PhD students from across the Schools of Informatics and Philosophy, Psychology and Language Sciences will talk briefly about their work. Confirmed speakers include:
There are two main visual pathways leading from the retina to the brain. They are the magnocellular pathway and the parvocellular pathway. Both of these types of neuron are specialised for detecting a particular type of visual information. The magnocellular pathway arises mostly from the area of the parafovea (which is responsible for movement detection and involuntary visual attention), whereas the parvocellular pathway mostly stems from the area of the fovea (which is responsible for high-acuity, colour vision). Both types are crucial to the process of visual word recognition. The magnocellular pathway is responsible for the transmission of coarse-grain information and the parvocellular pathway relays more fine-grain information. The magnocellular pathway may be responsible for both directing visual attention to specific points in the word, and for the parafoveal preview. Disruption of the magnocellular pathway has been implicated in certain forms of dyslexia. I will talk about the role of each of these visual pathways in the process of both normal, and impaired visual word recognition.
Affordances are a bridge between direct perception and action. Although they are a fundamental paradigm of ecological perception, they have been overlooked in the design of architectures of biologically inspired agents, whose functional knowledge has been usually hard-wired. During the talk I will introduce a first model to endow the agent with the capability of building its own functional perception and to use it for behaviour selection. Furthermore, I will also introduce a list of open questions to be addressed with regard to biologically inspired agents, and their corresponding experiments for the near future.
All languages must include some provision for expressing the location of objects. Prior to doing this, however, representations of objects. locations must be activated, so we know which spatial terms to apply to scenes. In the case of projective expressions (front, left, right, above, etc.) a reference frame, which identifies regions of space surrounding the ground object, must be applied to a scene. In English there are three potential reference frames that a person can use: absolute, relative and intrinsic, although in most everyday language only the last two are used. The absolute reference frame defines space in terms of environmental cues, such as the alignment of gravity. The relative reference frame situates the origin of the reference frame upon a viewer and thus defines space in terms of the viewer.s bodily axes. The intrinsic reference frame situates the reference frame on the ground object, and uses the ground object's intrinsic directional axes to define space.
These three reference frames can potentially be dissociated; for example in the case where an object is rotated from its canonical (usual) position. In such a situation a person wishing to describe a scene must select a reference frame from the options described above. Their addressee has to then decide which type of reference frame the speaker has used, if they wish to understand the speaker's utterance fully.
The present work uses a confederate-priming paradigm to investigate whether or not people's selection of reference frames are influenced by the reference frames used by their interlocutor.
The task of named entity annotation of unseen text has recently been successfully automated with near-human performance. But the full task involves more than annotation, i.e. identifying the scope of each (continuous) text span and its class (such as place name). It also involves grounding the named entity (i.e. establishing its denotation with respect to the world or a model). The latter aspect has so far been neglected.
In this presentation, it is shown how geo-spatial named entities can be grounded using geographic coordinates, and how the results can be visualized using off-the-shelf software. This is used to compare a ``textual surrogate'' of a newspaper story, with a ``visual surrogate'' based on geographic coordinates.
This talk describes some work carried out in collaboration with Gail Sinclair and Bonnie Webber, to be presented at the HLT-NAACL 2003 Workshop on Geographical References.
Parallel corpora (bilingual sentence-aligned texts) are vital for the creation of statistical translation models. IBM's original Candide system was trained on ten years' worth of Canadian Parliament proceedings, which consists of 2.87 million parallel sentences in French and English. While the Candide system was widely regarded as successful, its success is not indicative of the potential for statistical translation between arbitrary language pairs. The reason for this is that collections of parallel texts as large as the Canadian Hansards are rare. The reasons that using large amounts of training data ensures translation quality is simple: if a program sees a particular word or phrase one thousand times during training, it is more likely to learn a correct translation than if sees it ten times, or once, or never. Larger amounts of training material therefore leads to better quality. This presents a problem for language pairs which do not have large parallel corpora. In this talk I will present a method for semi-automatically creating parallel texts, using existing translation systems. I show how a weakly supervised learning method can be used to increase the size of a parallel corpus leading to increased translation quality. Furthermore, I will show how the method may be used to create training data for a language pair for which a parallel corpus is not available. Starting with no human translations from German to English I produced a German to English translation model with 45% accuracy. This suggests the method may be useful in the creation of parallel corpora for languages with scarce resources.
** This will act as a practice talk for a workshop presentation at the NAACL, so feedback is appreciated. **
With the introduction of side-channel attacks, security at the physical level of cryptographic hardware is of concern. At the same time, low-power cryptographic hand-help devices have become ubiquitous. A critical question is how to secure the physical layer against these attacks without degrading energy efficiency or performance.
The most powerful side-channel attack, DPA (Differential Power Analysis) relies on the correlation of power consumption measurements. It has been proposed that adding non-determinism to the execution of the cryptographic algorithm would prevent side-channel attacks. We propose using an asynchronous architecture, seen as a network, in which power traces are decorrelated by introducing timing shifts in the execution of instructions. Detailed simulation results demonstrate that the execution is randomised to counter DPA without unduly compromising the performance.
The background to this talk is a minority language development project in Tanzania, developing the Rangi language of Kondoa District. As an orthography has just been established, Rangi speakers are now looking to the preservation of their traditional stories and asking for being taught "good writing skills". As the features of a good story in Rangi are probably quite different from the features of a good story in English (or any other language, for that matter), it will first have to be explored what makes a good Rangi story before holding Rangi writer's workshops. In the talk, I will take one short Rangi story (16 sentences only) and look at it through a variety of narrative analytical models. Which of these provides insights potentially applicable to the teaching of Rangi story writing? Watch out: after the talk, you might never look at a story again in the same way ...
Tools for producing, accessing and modifying digital video are enabling new educational practices. I will present an approach to developing the tools and the practices by involving learners as designers. The approach is informed by field studies in different educational settings about tasks of video composition through a tangible interface in which videos are embodied in cards.
I will focus on the problem of interpretation and use. In particular, how to favour the 'open' interpretation of the representations and facilitate learners in realising closure by 'framing' them. The discussion will be concerned with two main questions: how do learners transform and appropriate the multimedia objects as 'artefacts', and how to support the accountability of learners. practices.
To be presented at the Computer Support for Collaborative Learning conference.
`Group discussion' is face-to-face, multi-party conversation, generally of a non-casual, purpose-oriented sort. Recent experiments by Fay et al. (2000) found that the interactive and communicative process in small groups (of up to 7 participants) is very different from bigger ones: small groups interact as in a series of 2-person dialogues, whereas bigger groups tend to interact in a more formal and distant way, akin to monologues broadcast to the group as a whole.
The aim of my PhD research is to build simulations of various aspects of the interaction of the small group discussions only, in order to (hopefully) provide an explanatory model of their participative and turn-taking process. The contents and intentions of the participants in a conversation are abstracted away (and simulated by probabilistic methods) so as to concentrate on the interactive aspect. Thus, the simulations only represent when participants are talking or not, and what other behaviours they do to help coordinate the interaction, such as backchannels, gestures, etc.
I am drawing on the appropriate social and linguistic literature for the facts and hypotheses to inform the models. They will then be evaluated against a corpus of video and audio recorded 5-person discussions. The methodology is to use a `baseline' simulation to adjust the attributes of the simulated participants to, however possible, match this real data. Then, using the same adjusted attributes in further simulations with specific additional elements in order to verify how much they affect, improve (or not), the comparison with the real data.
The baseline simulation has the `broad' characteristics of conversation put forward by Sacks, Schegloff and Jefferson (74): that it involves mostly a one-at-a-time speaker, with common but brief `overlaps' only at speaker-transitions, and that the first to start to speak at certain `transition-relevant' places in the speaker's discurse (characteristic `junctures' of one's talk) tends to `take' the next turn of talk, unless a later starter accuses any priority activity, such as a problem of understanding.
Building on this baseline model, further simulations will test specific elements of group discussion separately, in order to gauge their specific `contribution' to the interactive process. These include nonverbal behaviours such as gaze, gesture and posture shifts, and feedback behaviour that can also be used as turn precursors, such as "uh-huh", nods or "eh?" (ie. a problem of understanding). Also to be investigated are: pauses (and hesitations such as "uh") that attract `continuer' signals ("uh-huh", "yeah"), and simultaneous talk that may occasionally extend to inordinate lengths when participants compete for the `floor' to talk, as discussed by Schegloff (2000).
Multiword Expressions (MWEs) cause considerable problems for any semantically-grounded Natural Language Processing (NLP) application (including applications where semantic information is implicit, such as information retrieval) because their meaning is often not simply a function of the meaning of the constituent parts. There has been considerable recent interest in the representation of such items from various grammar development projects, but any treatment faces a considerable acquisition bottleneck.
This presentation will describe the implementation and evaluation of some statistical techniques for inferring the semantics of one type of MWE, the verb-particle construction, from corpus data.
This will be a practice run for a talk to be given to a workshop at the conference of the Association for Computational Linguistics.
The aim of this paper is twofold: firstly we want to describe the status of the Arbëresh spoken in Molise, a southern region of Italy, and secondly we want to test the validity of the recent introduction of Standard Albanian in the Arbëresh community. The data collected through a sociolinguistic survey in Portocannone, one of the Arbëresh enclaves in Molise, clearly show a high degree of variation in speakers' linguistic uses, as well as the age-related linguistic choice in the speaker's repertoire. The relevant generalisation is that there is a clear and marked trend with age: the older the speaker, the more Arberesh is the preferred language, and the less Italian is the preferred language. The picture showed by the youngest speakers, age 3 to 10, is not encouraging: Arbëresh is absent in their linguistic repertoire both in the form of active and passive competence. We observe a conscious interruption of the language transmission from one generation to another, which is the core point of the language death process.
However, some elements of revitalisation are recognised in the language minorities' status under the recent Italian laws, introducing the minority community languages in public life and in schools. What creates a problem is that Standard Albanian is being introduced. The results of the present analysis show as the adult Arbëresh speakers do not identify themselves with Albania and refuse Standard Albanian, stressing the relevance of the introduction of Arbëresh against Standard Albanian. While the youngest speakers, mostly monolingual in Italian, identify themselves neither with Arbëresh people, nor with Albanians. Another clear sign of decline and death of this language.
Fragments, i.e. non-sentential utterances which despite their `incomplete' syntactic form are intended to convey propositions, questions or request, are pervasive in dialogue. Perhaps the most prominent type of such utterances is the short answer, as in ``A: Who came to the party? --- B: Peter.'', but there are many other types as well.
In this talk we will present an analysis of the syntax, compositional semantics and contextually-situated interpretation of fragments. Our main thesis is that the resolution of the intended content of fragments can be modelled as a by-product of the establishment of coherence in dialogue, which (following much of the work on discourse) we define as the establishment of certain connections of the content of the current utterance to the content of its discourse context. We show that certain constraints on the form and content of fragments follow from how they are connected to the context.
Our analysis of the syntax of fragments is couched in HPSG, while the contextual resolution is framed in a theory of discourse called SDRT. Comparing our approach to the non-modular one of Ginzburg and Sag, we will argue that this division of labour is advantageous.
This is joint work with Alex Lascarides.
This will be a practice run for a talk to be given to the Conference on HPSG in Michigan -- a shorter version of the talk will also be presented at a workshop at the conference of the ACL. The talk will be approximately 30 minutes long.
This talk describes log-linear parsing models for Combinatory Categorial Grammar (CCG). Log-linear models can easily encode the long-range dependencies inherent in coordination and extraction phenomena, which CCG was designed to handle. Log-linear models have previously been applied to statistical parsing, under the assumption that all possible parses for a sentence can be enumerated. Enumerating all parses is infeasible for large grammars; however, dynamic programming over a packed chart can be used to efficiently estimate the model parameters. We describe a parellelised implementation which runs on a Beowulf cluster and allows the complete WSJ Penn Treebank to be used for estimation.
This talk will be presented at EMNLP next week in Japan.
Organisers: Tiphaine Dalmas, Annabel Harrison, DeLesley Hutchins, Vera Demberg, Tom Kwiatkowski
|
Informatics Forum, 10 Crichton Street, Edinburgh, EH8 9AB, Scotland, UK
Tel: +44 131 651 5661, Fax: +44 131 651 1426, E-mail: school-office@inf.ed.ac.uk Please contact our webadmin with any comments or corrections. Logging and Cookies Unless explicitly stated otherwise, all material is copyright © The University of Edinburgh |
