|
The Rutgers Distributed Laboratory for Digital Libraries | ||
|
|||
|
|
The Rutgers Distributed Laboratory for Digital Libraries | ||
|
|||
Two of the existing IS SROA-supported Working Groups (Human Centered Systems, and User-Centered Internet) have joined forces with others from the Newark Psychology Department who have worked specifically on digital libraries of distributed images, to continue their efforts under the single heading "Rutgers Distributed Laboratory for Digital Libraries". The Rutgers DLDL will bring together scholars from diverse disciplines including Psychology, Library Science, Information Science, Communications and Computer Science to address the enormously important issues of human access to very large distributed heterogeneous collections of digitized information, in a variety of media, and accessed through a variety of modes.
Core Team:
The extended team will likely grow to include all of the individuals named in the progress reports on the digital libraries component of the UCI activity, in the Knowledge Networking component thereof, and in the report of the HCI working group.
The challenging issues of human access to Digital Libraries are to be studied from a Rutgers perspective which asks, at all times, what difference do these enormous information resources make either for the direct users of specialized information resources, or for the much larger number of ordinary human beings, whose lives are substantially transformed by the fact that professions (such as medicine and engineering) have access to such collections of information. In other words, we will study both specialist users, of specialist information, and general users of either general or specialist information, to the extent that the latter occurs.
Groups of users to be directly studied, as we evaluate the effectiveness of new digital library components, will include: general users of the Rutgers libraries online resources (directed by University Librarian Marianne Gaunt); specialist users (depositors) of the nucleic acids database (directed by Prof. Helen Berman, or Chemistry); specialist and generalist users of the MapGarden GIS databases (directed by Prof. Tev Airola), all types of users of images of material objects, stored at the Jazz Institute (developed and maintained by Hanson and Bly). Our research will result in permanent enhancements to these diverse digital libraries, which will benefit all future users.
We will also run experiments on the TREC judged text collections (maintained locally by Belkin and Kantor at SCILS), adhoc collections of images developed in the course of image research by Dickinson, and email data collected by the AlterEgo project.
Throughout its work, the Rutgers Distributed Laboratory for Digital Libraries will maintain a consistent combination of fundamental research, into issues of cognition, and of information retrieval, and of rigorous metric-based evaluation of the resulting systems. Thus the results of the laboratory's work will be new advances in access to distributed, heterogeneous collections, and rigorous, replicable experiments to assess the specific impact of any advance, whether it be in the matching of queries to retrievable items, or in the support for multiple modalities of retrieving, or in the development of new techniques for visual navigation.
Immediate projects at the lab will include continuations of work by Belkin (on multiple modes of interaction), of Kantor (on information pheromones as an aid to search and navigation), of Dickinson, on the organization and retrieval of images), of Hirsh (on intelligent personal software assistants), of Pylyshyn (on visual principles for organization and navigation in digital libraries), and of Hanson and Bly (on the management, organization and searching of distributed libraries of complex images, such as a collection of jazz artifacts) There are many synergies among these research initiatives, and every member of the core team has collaborated in the past with at least one other member of the team, in two and three-way collaborations.
Descriptions of these research projects are presented in the Appendix.
Our theme is that we support interactions of various kinds, with distributed heterogeneous (especially includes images) media, AND we conduct rigorous experiments to assess the impact of specific support mechanisms, AND that these mechanisms are developed from a basis in either Cognitive Science, Human Behavior, Machine Learning or Information Retrieval.
To establish Rutgers as a national leader in this arena, we will use the distributed laboratory to draw in additional researchers at SCILS, at RUCCS, at the DCS, at RUTCOR, and at Psychology, Newark. In this environment a new cadre of graduate students will be prepared to work on the boundary-spanning problems which arise when technical systems are being developed to serve subtle human purposes of organizing and finding information.
At the present time, members of the core team are active in a variety of projects and proposals having to do with Digital Libraries specifically, or with the more general notion of "Knowledge Networks", which extends the ideas of digital libraries in some ways. They are also active, nationally, in defining the standards for measurement and evaluation of the coming round of Digital Library research projects.
In addition, the Rutgers DL for DL will be a strong candidate to be one of a small number of national repositories for Information Retrieval tools, which are to be established under a new Infrastructure program being set up now at the National Science Foundation. [Belkin was a co-chair, and Kantor a session leader, at the recent NSF workshop on this issue].
Proposed research into the effectiveness of Data Fusion in Information Retrieval (Kantor) will find a natural home in this laboratory as well.
The RDLDL research in the areas enumerated above will provide a continuous flow of pilot results, which will give us a demonstrated base of accomplishments from which to respond to new challenges in the exponentially growing arena of digital libraries.
Core members of the group are playing key roles right now in the development of the current Rutgers Digital Library proposal, headed by University VP for Research James Flanagan. As the researchers involved are based at the College Avenue Campus, the Busch Campus, and the Newark Campus, distributed collaborative capabilities are essential to the integration of the research.
This proposal brings together core researchers who were members of two different IS strategic initiatives. The earlier efforts by members of this team received total of $194,000 in Information Science SROA-2 funding. The effectiveness of those funds is documented in the progress reports of the UCI group, and of the HCI group - that are submitted separately.
We believe that to securely establish DL capability at Rutgers we must begin to train graduate students for work in the cross-disciplinary problems characteristic of DLs. Generally speaking, a DL problem will involve both a generic problem (distributed retrieval; organizational impacts; mixed or multiple media; methods for improved collaboration) and a domain specific issue (nucleic acids; financial databases; technical and engineering data, etc.).
We propose, as a group, to begin to train graduate students for this type of research. We seek start-up funding for 7 graduate fellowships, to be awarded to students who will work jointly with two or more members of the DLDL team. We undertake, collectively, to generate sufficient grant funding, by the second year, to continue these students on normal Research Assistantships. As a condition of their fellowships, these Rutgers Digital Library Fellows will be expected to attend group meetings every two weeks, at which they discuss their work with each other, and with members of the DLDL faculty. This obligation to remain in contact will continue through their graduate years, even though SROA funding will terminate at the end of the first year.
Finally, we propose to begin development of undergraduate courses in the Digital Library area, with development of a lower level undergraduate course, "Introduction to Digital Libraries" to be developed by members of the DLDL faculty, and to be taught for the first time in Spring 1999. There is groundwork for this in the graduate level course in Networked Information Retrieval, co-taught by Hirsh and Kantor in the Department of Computer Science. In addition, a good deal of material has been developed in a variety of courses at SCILS, which can serve as a basis for development of the introductory undergraduate course to be coordinated by Belkin.
These programmatic components define the core of our budget, and will, when coupled with the broad range of interests of the DLDL faculty, give Rutgers an excellent capability to respond effectively to the growing number of RFPs dealing with human-centered access to very large, heterogeneous, distributed collections of information. For example, at this moment there are two large NSF initiatives: the Knowledge and Distributed Intelligence program, and the Digital Libraries Initiative Phase II, each with budgets of some $50,000,000. While the work of this SROA group will be reviewed for funding after submission of the first multi- department, multi-discipline responses to each of these programs, there will be subsequent deadlines in early 1999, for which work in the DLDL will prepare Rutgers to be a very successful competitor.
The sites to be developed as the core of the Distributed Laboratory for Digital Libraries are: RUCCS, on the Busch Campus, where a usability lab has been started; the union of the Alexandria Project Lab (a nationally recognized center for the evaluation of libraries and their impacts) and the Information Interaction Labs (a nationally recognized center for studying multiple modes of user interaction with information systems), at SCILS, which are housed in the same space; and a site in Newark, possibly at the "Center for Informational and Instructional Technologies" (CIIT), which is based in the Dana library, directed by Lynn Mullins, in the development of which members of this group (Hanson and Bly) have been active.
Visualization of large complex image based data sets, their navigation and access. Application of human perception and cognition studies to the design of human centered information systems for navigating in large distributed databases.
This research places special emphasis on the use of spatial information to help in navigating through large data sets and solving the recurring problem of users' feeling "lost" as they explore such rich data sources as the Web.
Specifically we will study coding, storage and transmission of gestures and other location indicators which can help participants to individuate and refer back to items and ideas using movements such as pointing. Ultimately we want to develop a room-sized virtual spaces to locate both abstract and concrete ideas discovered in the course of a search. This research continues a major theme in the current SROA grant to the HCI (Human Centered Systems) group at NB. We have adopted the Digital Library as the application area to which to direct our studies of the optimal distribution of tasks between human and computer in the design of information systems, and of the design of systems for human- computer interaction with large rich datasets.
The Belkin project relates to the SROA efforts by providing the baseline work on highly adaptive systems for supporting multiple types of interactions with information within a single environment. The SROA funding will help us extend this work to more explicit concern with visualization and navigation in information spaces, and with integrating support for information interaction within the user's task environment. This work will connect with what Pylyshyn has proposed; indeed, it will be done jointly. Concrete plans for support in this area include proposals in preparation for both the immediate and next deadlines for KDI and DLI-2 funding, in Belkin's case proposals for: 1) investigating the design and usability of highly adaptive intelligent interfaces (building upon the DARPA work; and 2) investigating explicit methods for supporting iterative and cumulative interaction with many different, distributed and large databases (building on a PhD thesis presently being conducted in the Information Interaction Lab).
Digital libraries, including the ever-growing world wide web, contain increasing amounts of image and video data. While there are good tools for textual libraries, content-based image retrieval (CBIR) technology has not kept pace. Image statistics (color histograms, edge histograms, texture, ...) retrieve images that appear similar to the query image. In many important situations what the user really wants are images that contain similarly-shaped objects. Hence, generic object recognition is a key component of successful content-based image retrieval.
For example, with researchers in the Department of Biomedical Engineering (Dunn) and the Rutgers Center for Cognitive Science (Dickinson, Feldman), the Department of Computer Science (Dickinson), and the Boston University Department of Computer Science (Sclaroff), we are exploring a tool that allows dental technicians to outline a lesion on a dental radiograph, and retrieve patient records containing radiographs with similar-shaped lesions (and their diagnoses). In another project, involving the Department of Computer Science (Dickinson, Stevenson) and the MIT Media Lab (Pentland), we are exploring the problem of viewpoint-invariant shape indexing, in which the query shape can be viewed from a different viewpoint from its occurrence(s) in the image database. Each of these two projects raises a number of important human-computer interface issues concerning how query shapes, in either 2-D or 3-D, are specified by the user
In work with CAIP (Marsic) and Computer Science (Dickinson, Shokoufandeh), we are looking at a new class of structured image descriptions based on detecting salient regions in a multiscale wavelet transform of a 2-D image. This image description lends itself nicely to content-based image retrieval. In a project involving the Department of Computer Science (Dickinson, Stevenson, Hirsh), we are exploring the interaction between image and text retrieval. Specifically, if a textual query returns documents that contain images, can those images, in turn, be used to query other images whose containing documents are relevant to the original textual query?
For many of these problems , 2-D generic shape description and matching are critical. In a project involving the Department of Computer Science (Dickinson, Shokoufandeh) and the Yale University Department of Computer Science (Siddiqi and Zucker), we are exploring a new class of shape description and matching techniques based on a skeletal-like description of an object's shape, and supporting very fast, novel matching algorithms. In a project with Computer Science (Kulikowski, Dickinson) and Chemistry (Berman), we plan to apply our shape description and matching algorithms to the indexing of biological structures from large databases. We will look at returning molecules or proteins whose coarse structure resembles the query shape. This addresses an important class of problems in biological digital libraries.
The above projects involve the searching of digital libraries that contain image data, text data, and biological shape data. All share the common theme that effective indexing requires generic shape description, indexing, and matching. In addition, each project, due to the human in the loop, must develop effective interaction techniques to allow users to specify shapes, provide relevance feedback, and to view the results of a search.
DIPs, now being developed in a project by Kantor, Boros and Melamed, are a compressed way of storing information about the value of links in a network, to searchers with specific purposes. If later searchers can "sniff out" the values judgments assigned by earlier searchers *with similar purposes* then they can more effectively find their ways to valuable nodes and the information that those nodes contain. This notion moves beyond the conventional assumption that "information is either relevant or not", to support judgments of what the information adds, given what the searcher already knows. Furthermore, the addition of pheromonic capabilities to a knowledge network transforms it to an Adaptive KN, which can learn about the needs and preferences of its users, as it is used. Specific research issues include (1) encoding search purposes in condensed bit strings (the pheromones) (2) matching the purposes of a current searcher to the most similar purposes in the store of previous information (3) effective elicitation of the judgments from the users and (4) assessment, using objective metrics, of the impact of the overall design of the pheromone system on the usefulness and value of the knowledge network or digital library.
This research has a collection component: digitizing various archives in the library, including audio, some video, video of artifacts (e.g. Hats of jazz musicians, Trumpets etc..) and various text sources (sheet music, posters etc..) which has immediate overlap with the research on content-based image retrieval. Other research components having multiple overlaps with the other threads of research, including pheromonic navigation and adaptation, multiple modalities, and spatial reasoning/cognition for navigation.. The collection will provide a huge volume that would support experimentation and measurement, and which would be used by a substantially different sample of users and user types than the other parts of the DLDL. ( This collection, and the associated research tasks have unique potential for corporate sponsorship, now being explored by Hanson)
This research targets the development of principles and algorithms to
inject the functionality of a "personal assistant" into computer
systems. The AlterEgo project is exploring the use of machine
learning to generalize from a user's past email-reading behavior to
predict how the user will want to read email in the future. Our
initial system has been able to replicate a real user's email-reading
preferences with 86.8% accuracy, and future work will develop,
implement, and test a new class of learning methods that address the
remaining shortcomings of this system, most notably incrementality as
well as the ability to use a range of contextual information in the
decision process. The Ilash project attempts to predict what actions
a user would be most likely to perform next and make it available to
the computer system. Our initial recently published results on 2-6
months of data for 77 real users show that the system can predict a
user's next command with over 40% accuracy extrapolating from only a
user's history of previous commands, and our future efforts target the
more effective use of a wider range of information sources, such as
the nature of the computer the user is using (via a fast network or
slow modem, for example). The MovieValet project (with colleagues
from Bellcore and AT&T) takes collections of movie-ratings data plus
background information about movies (cast lists, reviews, etc.) and
forms movie recommendation rules personalized to each user. In our
initial recently published results on over 45000 movie ratings for
more than 200 users we already outperform a commercial system
developed at Bellcore, and our future efforts target the more
effective use of a wider range of information sources. Two main
themes being explored in all these projects are the use of machine
learning to learn from and generalize user's behaviors, and the fusion
of a diverse set of sources of information in performing such
learning.
April 14, 1998
| RDLDL Home | | SCILS Home | | Rutgers Home |