Date: 30th November 2007, 12:15, CS202J Computer Science, Univ. of York (Fri week 8)
Title: Persuasive Dialogue. Challenges for Natural Computer Dialogue.
Speaker: Pierre Andrews, Computer Science Department, The University of York.
Human Computer Dialogue has been developed and studied since almost the beginning of Artificial Intelligence research. However, the state of the art techniques are not yet able to deal properly with the complex dialogue interaction that is needed for persuasive dialogue. I will present my ongoing research on developing a general framework for human computer persuasive dialogue and put forward techniques that are well developed in AI and can help creating a better suited dialogue system for persuasion.
Date: 23th November 2007, 12:15, CS202J Computer Science, Univ. of York (Fri week 7)
Title: Information Retrieval
Speaker: Dimitar Kazakov, Computer Science Department, The University of York.
This talk will cover some of the basic principles behind information retrieval, the technology which is best known to the public through the use of search engines. While focussing on text documents, other types of documents will be touched upon. An in-house document clustering product will be used to illustrate some of the ideas. The talk is suitable for anyone interested in the subject and will require a minimum background (basic probabilities). This will be a shortened version of a 1/2 day invited tutorial presented at RANLP-07.
Date: 16th November 2007, 12:15, CS202J Computer Science, Univ. of York (Fri week 6)
Title: Rational Dialog in Interactive Games
Speaker: Maria Arinbjarnar, Computer Science Department, The University of York.
"I am fairly sure of this that none ever willingly errs". Socrates
The motivation for this research is the increased call for games that have a high degree of interaction with the player and a dynamic environment with intelligent Non-Player Characters (NPCs). The NPCs currently implemented in computer games rarely act in a rational way, although some have an emotional drive and a set of goals to chase. Their actual interactions are usually preset and/or very limited. Additionally the games themselves have preset narratives that result in games that the average player does not care to play numerous times, simply because the game is always the same.
The question addressed in this thesis is whether an NPC will interact with a player and other NPCs in a rational and goal driven way when given a past life and a decision mechanism based on a causal network like a Bayesian network. Will the NPC adopt a strategy that will maximize its pay-offs?
To answer this question I build a prototype engine that creates NPCs that have past lives, a knowledge base and tools to find a sentence to speak in a rational dialog. The knowledge base and past lives of the NPCs are created from plots that the Dynamic Plot Generating Engine (DPGE) creates. The DPGE creates continuously new plots for murder mystery games that are logically consistent. The interactive interfaces of the NPCs are modeled using Multi-Agent Influence Diagrams (MAIDs) and game theory, a mathematical method of decision-making in competitive situations.
The prototype engine created clearly indicates that there is basis to create NPCs that can participate in a rational dialog by calculating optimal sentences on the fly. The time complexity is linear in respect to number of sentences and more than half of the sentences are calculated in less than 1 minute. Moreover with some standard optimizations these results can be greatly improved.
Date: 9th November 2007, 12:15, CS202J Computer Science, Univ. of York (Fri week 5)
Title: Learning to Compose Effective Strategies from a Library of Dialogue Components
Speaker: Marco De Boni, Unilever Corporate Research
I will describe a method for automatically learning effective dialogue strategies, generated from a library of dialogue content, using reinforcement learning from user feedback. This library includes greetings, social dialogue, chit-chat, jokes and relationship building, as well as the more usual clarification and verification components of dialogue. We tested the method through a motivational dialogue system that encourages take-up of exercise and show that it can be used to construct good dialogue strategies with little effort.
Date: 31st October 2007, 14:00, CS103 Computer Science, Univ. of York (Wed week 3)
Title: Spoken Dialogue for Intelligent Tutoring Systems: Opportunities and Challenges
Speaker: Diane J. Litman, University of Pittsburgh, USA and University of Edinburgh
In recent years, the development of intelligent tutoring dialogue systems has become more prevalent, in an attempt to close the performance gap between human and computer tutors. Tutoring applications differ in many ways, however, from the types of applications for which spoken dialogue systems are typically developed. This talk will illustrate some of the opportunities and challenges in this area, focusing on issues such as affective reasoning, discourse and dialogue analysis, and performance evaluation.
About the speaker: The speaker is a Leverhulme Visiting Professor at the University of Edinburgh and this talk is designated as a Leverhulme Lecture.
Date: 24th October 2007, 14:00, CS103 Computer Science, Univ. of York (Wed week 2)
Speaker: Barry O'Sullivan, Cork Constraint Computation Centre, University College Cork, Ireland
Title: Robust Combinatorial Auctions
Combinatorial auctions involve the sale of multiple distinguishable items where bidders can bid on combinations of desired items. This allows bidders to express perceived complementarities or substitutabilities between items, thus increasing revenue for the bid-taker and improving efficiency. Revenue-maximizing winner determination is algorithmically challenging. Solution uncertainty, e.g. bid-withdrawal or an incapacitated bidder, can leave the bid-taker exposed to significant revenue losses or supply failures. In this talk we will review combinatorial auctions, how the winner determination problem can be solved, and consider a novel algorithmic approach for finding robust solutions of near-optimal revenue that can withstand bid withdrawal. We will also consider how auctions can be applied to the problem of large-scale industrial procurement. Procurement can be seen as a reverse auction in which the bidders are the suppliers and the bid-taker seeks to minimize its purchasing costs. Economic efficiencies can be achieved by allowing suppliers to state their preferred combinations of items to supply.
About the speaker: Dr. Barry O'Sullivan is Associate Director of the Cork Constraint Computation Centre, a Senior Lecturer in Computer Science at University College Cork, Chairman of the Artificial Intelligence Association of Ireland, and President of the Association for Constraint Programming.
Date: 19th October 2007, 12:15, CS202J Computer Science, Univ. of York (Fri week 2)
Title: Applying algebraic statistics to probabilistic-logical representations
Speaker: James Cussens, University of York
Apologies for the scary looking title! This talk concerns two areas - probabilistic-logical representations (PLRs) and algebraic statistics - which will not be familiar to most people, and so the talk will be at an introductory level. Necessarily, this involves skating over some of the issues.
The patterns of conditional independence in PLRs are generally more complex than are found in graphical models such as Bayesian networks. So, although it useful to translate between PLRs and graphical models where possible, a representation more expressive than graphs is needed. Here, 'polynomial ideals', coming from the area of algebraic statistics, constitute the proposed alternative. Algebraic statistics is a sub-branch of algebraic geometry, and its geometrical aspect makes it easier to 'see' how it works. Using the polynomial ideal representation we show how Buchberger's algorithm for computing Groebner bases can be used to perform computations useful for PLR.
Date: 12th October 2007, 12:15, CS202J Computer Science, Univ. of York (Fri week 1)
Unsupervised Learning of Natural Language Semantics (Thesis Seminar)
Ioannis Klapaftis, Computer Science Department, The University of York.
Lexical semantic ontologies such as WordNet are widely used for Natural Language Processing. However, their development as well as their enrichment with new information requires a high degree of human supervision. This problem has motivated the appearance of knowledge acquisition methods for building ontologies automatically. Current state of the art has mainly focused on pattern-based and set theoretic techniques to develop a taxonomy of concepts and less on learning the semantics of words. In this talk, we present our work on learning natural language semantics and particularly our research on Word sense Induction, Word Sense Disambiguation and Term Recognition.
Date: 22th June 2007, 12:15, CS202J Computer Science, Univ. of York (Fri week 9)
Analysing Heuristic Performance with Response Surface
Models: Prediction, Optimisation and Robustness
(A talk to be given at GECCO 2007)
Enda Ridge, Computer Science Department, The University of York.
This research uses a Design of Experiments (DOE) approach to build a predictive model of the performance of a combinatorial optimisation heuristic over a range of heuristic tuning parameter settings and problem instance characteristics. The heuristic is Ant Colony System (ACS) for the Travelling Salesperson Problem. 10 heuristic tuning parameters and 2 problem characteristics are considered. Response Surface Models (RSM) of the solution quality and solution time predicted ACS performance on both new instances from a publicly available problem generator and new real-world instances from the TSPLIB benchmark library. A numerical optimisation of the RSMs is used to find the tuning parameter settings that yield optimal performance in terms of solution quality and solution time. This paper is the first use of desirability functions, a well-established technique in DOE, to simultaneously optimise these conflicting goals. Finally, overlay plots are used to examine the robustness of the performance of the optimised heuristic across a range of problem instance characteristics. These plots give predictions on the range of problem instances for which a given solution quality can be expected within a given solution time. A preprint will be available from
Date: 15th June 2007, 12:15, CS202J Computer Science, Univ. of York (Fri week 8)
Title to be announced
Rania Hodhod,Computer Science Department, The University of York.
Educational computer-based games (edugames) are games that promote the acquisition of skills and knowledge in a pleasant interactive way. It is well known that not all the users share the same preferences or styles when interacting with a game and solving game-problems. This leads to the importance of adaptation in the sense that behavior of each play-instance of a game depends on the actions of an individual user/player. The major aim for an adaptive game-based learning system is to support and encourage the learner/player/user by considering his needs, strengths and weaknesses. One of the main problems encountered in educational games is the real time adjustment of the background story and being adaptable to individual users. This leads to the emerging of the idea of integrating the interactive narration with the educational games. We think this can lead to the desired adaptation of the background story, and helps the user being more engaged and immersed, also leading to more emotional and educational outcome.
Date: 9th March 2007, 12:15, CS103 Computer Science, Univ. of York (Fri week 9)
Learning Probabilistic Sequence Models for Uncovering Gene Regulation
Mark Craven, University of Wisconsin (currently visiting Cambridge University)
A central challenge in computational biology is to uncover the mechanisms and cellular circuits that govern how the expression of various genes is controlled in response to a cell's environment. In this talk, I will discuss two aspects of our work on learning probabilistic grammars to identify gene-regulatory elements in genomic sequences. First, I will describe a method we have developed for learning expressive models of cis-regulatory elements (CRMs). A CRM is a configuration of sequence patterns that controls how a set of genes responds to specific conditions in a cell. Second, I will discuss an approach we have devised for learning mappings from sequential data to real-valued outputs. This learning process involves inferring the structure and parameters of a conventional hidden Markov model, while simultaneously learning a regression model that maps features characterizing paths through the model to continuous responses. We also evaluate this approach in the context of explaining the expression levels of genes in terms of sequence patterns.
Date: 8th February 2007, 12:15, CS119N Computer Science, Univ. of York (Thursday week 5)
Efficient Evolutionary Algorithms with Extended Fitness
Daniel Cook, Final year student, University of York
The notion of extended fitness has been assigned various meanings, of which only the oldest, expressing individual reproductive success has been used in Evolutionary Algorithms so far. This work builds on recent experiments suggesting the use of other well-known definitions borrowed from biology that are based on the success with which genes replicate and propagate themselves in the gene pool could be beneficial to EAs by replacing ad-hoc scaling techniques with means of maintaining genetic diversity that draw clear analogies with nature. Techniques are implemented to reduce the extended fitness overhead and performance is compared with other state-of-the-art Evolutionary Algorthims.
Using WordNet and Short-Term Memory for Contextual Disambiguation
Matthew Burke, Final year student, University of York
Natural language disambiguation is complex and many different methods have been applied to varying degrees of success, including several using WordNet, a database of syntactic sets (synsets). George Miller proposed that 5-9 concepts can be simultaneously held in the working memory. This work will explore the use of a memory model to hold concepts triggered from corpora, represented as WordNet synsets, as a feedback device to aid disambiguation.