Database Research Group Events

One approach to integrating knowledge bases is to based on finding assertions that relate the expressions in the constituent terminologies. For knowledge bases with many terms this task requires computer support. The authors set up formal framework for merging Description Logic TBoxs, and then explore the limits of a purely structural approach to the problem of finding inter-relationshops between knowledge bases. Some theoretical notions are empirically examined in a real medical ontology (GALEN).

(For Schema Integration Background:)
A Comparative Analysis of Methodologies for Database Schema Integration, C. Batini, M. Lenzerini, S. Navathe, ACM Computing Surveys, 18(4), 1986

(For More Detailed Work of the Authors:)
"What's not in a name?" Initial Explorations of a Structural Approach to Integrating Large Concept Knowledge-Bases, A. Borgida, R. Kuesters, Technical Report DCS-TR-391, CS Rutgers Univ. 1999 

The problem of Question Answering (QA) as used in the TREC can be formulated as follows:Given a collection of natural-language (NL) documents find an answer to given NL query that is a short substring of one of the documents and found in a relevant context.

I will:

• discuss the problem in the context of the problems of IR, IE, and the classical QA, and why it is important,
• give a description of the TREC QA task,
• present some related systems: AskJeeves.com, FAQ Finder, predictive annotation, and ExtrAns, and
• present a novel approach based on the stochastic unification-based grammars.

For TREC 9, we completed some of the experiments that we didn't get done in time for TREC 8. We parsed the questions, and used the parse to enhance the set of search terms and to locate the answer within candidate passages. I will present some results and ideas for further development.

This talk will explain how DB2 is evolving to meet the extensive demands of e-business. Starting with an assessment of the role databases play in the application development process in the world of e-business, I will show you how users leverage the core e-business standards optimized in DB2 application development. Next, I`ll address the expansion of the new functions in DB2 V7 for UNIX, Windows and OS/2 that address business intelligence environments in the e-business world. This includes features to support rolling OLAP functions, enhancement to OLAP cube support, advanced statistical analysis, complex query cache management, and how DB2 helps OLAP and mining tools, and ERP applications. The session will also include an analysis of XML in the handling of business objects and business-to-business and a demonstration of DB2`s capabilities in this domain as it relates to business intelligence.

I will also give a quick overview of some successful applications of the model for data retrieval, transformation, and hypertext creation. I will then discuss the potential for applying the model to XML data and introduce some interesting reseach problems.

Reference: A.Salminen and F.W.Tompa, "Grammars++ for Modelling Information in Text," Information Systems, Vol. 24, No. 1 (1999) 1-24. 

Here are some references:

1. Quilt paper
2. XML Query Requirements
3. Quilt solutions:
• http://www.almaden.ibm.com/cs/people/chamberlin/usecase_xmp.quilt
• http://www.almaden.ibm.com/cs/people/chamberlin/usecase_tree.quilt
• http://www.almaden.ibm.com/cs/people/chamberlin/usecase_seq.quilt
• http://www.almaden.ibm.com/cs/people/chamberlin/usecase_r.quilt
• http://www.almaden.ibm.com/cs/people/chamberlin/usecase_sgml.quilt
• http://www.almaden.ibm.com/cs/people/chamberlin/usecase_text.quilt
• http://www.almaden.ibm.com/cs/people/chamberlin/usecase_ns.quilt
• http://www.almaden.ibm.com/cs/people/chamberlin/usecase_parts.quilt
• http://www.almaden.ibm.com/cs/people/chamberlin/usecase_ref.quilt

Spurred by the potential of discovering interesting and useful knowledge, substantial research has been done on data mining. Most previous work essentially falls into one of two "generations". In the first generation, the focus was on identifying which patterns are interesting and significant, and devising fast algorithms for mining them from large data sets. In the second generation, the importance of integrating data mining with the other key components of the knowledge discovery (KDD) process has been recognized. One such component is the underlying DBMS and strategies and architectures for integrating association rule discovery with the DBMS have been studied. 

In this talk, I will focus on the other, equally (if not more) important component of KDD -- the human user. Many mining algorithms devised in the first generation implicitly assume data mining is a one-shot exercise, as opposed to the iterative exploratory process it really is. This is a significant shortcoming since mining algorithms tend to be computationally intensive. Thus, the concerns in integrating the user in the loop are: (i) how can the user control the nature of the mining computation undertaken by the system at any point? (ii) how can the user enforce focus on mining based on his knowledge of application semantics? (iii) how can the user migrate from specific mining tasks performed, to issuing ad hoc mining queries? I will discuss how constraints can play a significant role in addressing these concerns, specifically in the domains of frequent sets and clustering. Many a time, finding {\em when} a pattern holds in a data set can be as interesting as the pattern itself: e.g., when is an itemset frequently purchased? I will briefly discuss algorithms and structures that facilitate this kind of dual mining. 

Finally, the data mining process tends to involve multiple mining tasks, such as classification, frequent sets, data cube, etc. I will briefly discuss our recent work on a model and an algebra for data mining that neatly integrates many mining tasks into one framework so the input of one mining task can be the output of another.