Time: 12:00 noon
Place: 5317 Sennott Square
Speaker: Yang Xu (Information Sciences, University of Pittsburgh)
Title: "An Integrated Token-Based Algorithm for Scalable Coordination"
Abstract: Efficient coordination among large numbers of heterogeneous agents promises to revolutionize the way in which some complex tasks, such as responding to urban disasters can be performed. However, state of the art coordination algorithms are not capable of achieving efficient and effective coordination when a team is very large. Building on recent successful token-based algorithms for task allocation and information sharing, we have developed an integrated and efficient approach to effective coordination of large scale teams. We use tokens to encapsulate anything that needs to be shared by the team, including information, tasks and resources. The tokens are efficiently routed through the team via the use of local decision theoretic models. Each token is used to improve the routing of other tokens leading to a dramatic performance improvement when the algorithms work together.
Time: 12:00 noon
Place: 5317 Sennott Square
Speaker: Hua Ai
Title: "Comparing User Simulation Models For Dialog Strategy Learning"
Abstract: In this study, we explore what kind of user simulation model is suitable for developing a training corpus for using Markov Decision Processes (MDPs) to automatically learn dialog strategies. Our results suggest that with sparse training data, a model that aims to randomly explore more dialog state spaces with certain constraints actually performs at the same or better than a more complex model that simulates realistic user behaviors in a statistical way. Also, we find that the state space representation and the reward function of the MDP framework have an impact on the choice of simulation models.
Time: 12:00 noon
Place: 5317 Sennott Square
Speaker:Pamela Jordan
Title: A Collaborative Dialogue Agent for Peer Learning
Abstract: Peer tutoring and collaboration strongly promote learning [Cohen et al.,1982; Brown & Palinscar,1989; Britz et al., 1989; Rekrut,1992]; however, it has not been studied as extensively as tutoring in the Artificial Intelligence and Education Community. I will describe a project with UIC in which we seek to increase our understanding of peer learning by (1) developing a computational model of peer collaboration and (2) embodying this model within an artificial agent that can collaborate with a human student via natural-language (NL) dialogue. We are analyzing and modeling collaboration on elementary data structure and algorithm tasks, such as linked lists, stacks and binary search trees.
Time: 12:00 noon
Place: 5317 Sennott Square
Speaker:Kevin Ashley
Title:"Learning by Graphically Representing US Supreme Court Oral Arguments"
Abstract: If an intelligent tutoring system is to provide feedback on students' arguments, it must have (1) an input format for argument that is easy to learn and use for students; and (2) methods that enable it to analyze students' arguments expressed in this format and generate comments or reflection questions. We address these issues in the context of an intelligent tutoring system, LARGO, that helps students analyze oral argument made before the US Supreme Court. Using LARGO, students create diagrams of argument transcripts in a special-purpose graphical language geared toward the kind of hypothetical reasoning that typically occurs in these transcripts. LARGO provides feedback on students' argument diagrams in the form of reflection questions. It matches students' argument diagrams against anticipated patterns of incomplete, non-standard, or intricate hypothetical reasoning. To evaluate LARGO's educational effectiveness, we compared learning results of students who used LARGO to those of students who applied standard text-based annotation tools and techniques. The study involved 28 first-semester law students in the context of a first-semester Legal Process course. All students analyzed four argument transcripts dealing with issues of personal jurisdiction. All students also took a pre- and post-test measuring their ability to analyze and engage in hypothetical reasoning in a range of familiar and unfamiliar contexts. The use of LARGO did not lead to better learning overall. It did however have advantages for lower-ability students, as measured by LSAT scores. Among these students, those who used LARGO learned better to analyze hypothetical reasoning in a familiar area of the law and learned better to evaluate the import of hypotheticals with respect to proposed decision rules in familiar and unfamiliar areas.
Time: 12:00 noon
Place: 5317 Sennott Square
Speaker1:Min Chi
Title:The Impact of Explicit Strategy Instruction on Problem-solving Behaviors across Intelligent Tutoring Systems
Abstract: Explicit instruction of a problem-solving strategy in one task domain impacted students' problem-solving behaviors in a second task domain where it was not explicitly taught. Previously, we have found that not only did teaching students a problem-solving strategy improve their learning in the domain where the strategy is taught but also in a new domain that it was not taught. In this paper, we mainly investigated whether previous experience of learning a problem-solving strategy had any impact on students' problem-solving behavior in the new domain. Log files have shown that if students did not learn a problem-solving strategy beforehand, their problem-solving strategies would be similar no matter whether they were capable of independently solving the problem either partially or completely; however, if they learned a problem-solving strategy beforehand, their strategies would be different. When students were partially capable of independently solving the problem, they would apply the strategy they learned beforehand; but once they were completely capable of solving the problem, they would discard the strategy. Moreover, we also found that applying the previously learned problem-solving strategy not only made students more capable of problem-solving but also made them develop a different pattern of solutions from those who did not.
Speaker2:Xiaohui Kong
Title:Global vs. Local Information Processing in Visual/Spatial Problem Solving: The Case of Traveling Salesman Problem
Abstract: Human visual/spatial problem solving often requires both global and local information to be processed. But the relationship between those two kinds of information and the way in which they interact with one another during problem solving has not been thoroughly discussed. In the particular setting of solving the traveling salesman problem (TSP), we investigated into the relative roles of global and local information processing. An experiment was conducted to measure the importance of global information and the possible constraints of global information processing on search. A model was built to simulate human TSP performance and was used to investigate further the relationship between global information processing and local information processing. Our model was compared with the human data we collected and with other models of human TSP solving.
Time: 12:00 noon
Place: 5317 Sennott Square
Speaker: Sperling Martin (Information Resources Management Inc.)
Title: Search Engines at Pitt before there were Search Engines
Abstract: Back in the late 1950s and early 1960;s the University of Pittsburgh's Computation and Data Processing Center the predecessor of today's Department of Computer Science, was a hub of advanced applied research in what were then called full-text document retrieval systems - the ancestors of today's search engines. The work at Pitt led to one of the first Computer Aided Legal Research systems - some 12 years or so before Westlaw or LexisNexis. Other projects were pioneering efforts in natural language processing of law documents that formed the core of early legislative information systems and computerized law publishing. The work at Pitt in that era also informed the ANSI and ISO standards projects that led to the development of SGML, the precursor of XML. The presentation will be a brief history of the technology and information processing challenges faced by those pioneering efforts of nearly fifty years ago and how the Pitt teams solved them.
Time: 12:00 noon
Place: 5317 Sennott Square
Speaker1: Cem Akkaya
Title:Subjective Collocations of Function Words
Abstract: Subjectivity refers to aspects of language used to express opinions, evaluations, emotions, and sentiments. Subjectivity can be utilized to support various NLP applications like information extraction, question answering, and text categorization.There have been some attempts to identify phrases and collocations in text which are good subjectivity clues. However, function words are neglected in this context. We suspect that function words take part in collocations which bear subjectivity. Our work will exploit function words as subjectivity clues and their interaction with other lexical subjectivity clues.
Speaker2: Yanna Shen
Title: Bayesian Disease Outbreak Detection that Includes a Model of Unknown Diseases
Abstract: Algorithms for detecting anomalous events can be divided into specific and non-specific detection methods. Specific detection algorithms look for pre-defined anomalous patterns in the data. For example, in the context of disease-outbreak detection, a specific detection algorithm might examine health-care data for the onset of a known disease such as influenza. These methods are usually Bayesian. Non-specific detection algorithms attempt broadly to detect deviations from some model of the baseline situation. Most frequentist detection methods are non-specific. In this paper, we introduce a Bayesian approach that models and detects specific diseases (e.g., influenza, anthrax, etc.) and non-specific diseases (e.g., a new, highly contagious respiratory virus that has never been seen before). Experimental results support that this hybrid approach can improve the detection of anomalies, such as the outbreak of a new diseases in a population. Beyond anomaly detection, a contribution of this paper is that it introduces a general approach for Bayesian modeling of unknown processes.
Time: 10:00 am
Place: 5317 Sennott Square
Speaker: Giuseppe Carenini, University of British Columbia
Title:Interactive multimedia summaries of evaluative documents
Abstract: Many organizations are faced with the challenge of summarizing large corpora of text data. One important application is evaluative text, i.e. any document expressing an evaluation of an entity as either positive or negative. For example, many websites collect large quantities of online customer reviews of consumer electronics. Summaries of this literature could be of great strategic value to product designers, planners, manufacturers and consumers. In this seminar, I will first present and compare two approaches to the task of summarizing evaluative text. The first is a sentence extraction-based approach, while the second is a language generation-based approach. These approaches have been tested in a user study. In the second part of the seminar, I will describe an interactive multimedia interface which presents the knowledge extracted form a corpus of evaluative documents not only as a natural language summary but also in a hierarchical visualization mode. The interface is interactive in that it allows the user to explore the original dataset through intuitive visual and textual methods. Results of a formative evaluation of our interface show general satisfaction among users with our approach.
Time: 3:00 pm
Place: School of Information Sciences 501 - 5th Floor
Speaker1: Jill Freyne, Postdoctoral Researcher, University College Dublin, Ireland
Title:"Collaborative Web Search - Exploiting User Activity for User Benefit"
Abstract: Search engines continue to struggle with the challenges presented by Web search: vague queries, impatient users and an enormous, and rapidly expanding, collection of unmoderated, heterogeneous documents all make for an extremely hostile search environment. Conventional approaches to Web search those that adopt a traditional, document-centric, information retrieval perspective are limited by their refusal to consider the past search behaviour of users during future search sessions. Collaborative Web Search (CWS) seeks to exploit the high degree of natural query repetition and result selection regularity that is prevalent among communities of searchers. CWS reuses the search experiences of community members, to promote results that have previously been judged relevant for queries. This facilitates a better response to the type of vague queries that are commonplace in Web search and allows a generic search engine to adapt to the preferences of communities of individuals.
Speaker Biodata: Dr Jill Freyne is a postdoctoral researcher at the Department of Computer Science in University College Dublin. She received her PhD and BSc degrees from UCD, and worked as part of the I-SPY research project which focuses on personalized, community based Web search. She has since worked as a Post Doctoral Researcher in the Adaptive Information Cluster in the School of Computer Science and Informatics working in the areas of social support, annotations and digital graffiti.
Time: 12:00 noon
Place: 5313 Sennott Square
Speaker1: Shuguang Wang
Title:Identifying Opinion Targets in Customer Review Data
Abstract: Given an opinion sentence, we would like to identify the targets in it. This is a information extraction task and it is useful to generate a automate opinion summary. We classify the targets into two types, explicit and non-explicit targets. Explicit targets are those explicitly stated in the opinion sentences and non-explicit targets are implicit or those are represented by pronominal terms. Therefore we divide our task into two sub tasks, each of which deals with one type of targets. We explore a Machine Learning approach to extract explicit targets and make use of an ontology to identify the non-explicit targets. We will present the results of our approach on a customer review data set.
Speaker2: Aditya Nemiekar
Title:Classifying classical music pieces using Graphical Models
Abstract: A daunting task in Music Perception is the concept of expressivity - elements in musicians' renditions that make them stylized and unique but perhaps more importantly enjoyable. I plan on studying just this phenomenon. However, the task of modeling expressivity as a whole is by no means trivial. So, as a pilot project, I propose we learn to identify expressivity before we learn to exhibit it - by learning each musician's style and then identifying whether a piece played belongs to a particular composer.