ABSTRACT

The TileBars interface is a graphical tool for users of information access systems, that shows the relationship between the terms in a query and the documents that are retrieved in response to that query. TileBars simultaneously and compactly indicate relative document length and query term overlap, frequency and distribution. The patterns in a column of TileBars are meant to help users make fast judgments about the potential relevance of the retrieved documents. An unexpected benefit of the interface is that because it requires users to specify their queries as a list of topics, better rank orderings can be obtained than with standard information retrieval ranking mechanisms.

Keywords

Information retrieval, Information Access, Full-length text, Visualization.

Note: A longer version of this videotape appeared in the video proceedings of the International Joint Conference of Artificial Intelligence (IJCAI) 1995.

THE TILEBAR INTERFACE

Information Access research at Xerox PARC focuses on amplifying the users' cognitive abilities, rather than trying to completely automate them. This framework emphasizes the participation of the user in a cycle of query formulation, presentation of retrieval results, and query reformulation. The information presented is typically not document descriptions, but rather intermediate information that indicates relationships between the query and the retrieved documents. We have developed information access tools intended to supply some of this functionality. This video demonstrates one of these, called TileBars [3].

In a typical information retrieval system, documents satifying the query are returned and are rank-ordered according to some function of the number of hits for each term [5]. We argue that this kind of ranking is opaque to the user, in part because it is not clear how well each term is represented in the retrieved documents.

By contrast, the TileBars graphical interface allows the user to make informed decisions about which documents and which passages of those documents to view, based on the distributional behavior of the query terms in the documents. The goal is to simultaneously and compactly indicate (i) the relative length of the document, (ii) the frequency of the term sets in the document, and (iii) the distribution of the term sets with respect to the document and to each other.

The TileBar interface requires the user to type queries into a list of entry windows. Each entry line is called a termset since it is intended to contain a set of terms representing one topic. Typically, the query is treated as a conjunction of topics, and the topics are listed in order of importance.

In order to perform retrieval over full text (as opposed to just titles and abstracts), we impose some very simple structure on the documents in the collection. Each document is partitioned in advance into segments, the size and extent of which can be determined several different ways. The orthographic structure provided by the author, in terms of paragraphs, pages, or sections, can be used if available. A still simpler alternative is to use contiguous blocks of text of a fixed size. These two approaches have drawbacks: author- provided structure can vary greatly in length and meaning, and fixed-length blocks are not intuitive units to show to users. We prefer to use a robust, statistical segmentation algorithm called TextTiling [2] to subdivide documents into multi-paragraph subtopical units. This algorithm allows for the creation of meaningful, or motivated, segments even if the author has provided none, and allows for customization of the average segment length.

Figure 1 shows an example of the TileBars display on a query about automated systems for medical diagnosis. The graphical representation works as follows. Each rectangle represents a document. Each row of the rectangle represents the corresponding termset in the query display, i.e, the top row corresponds to patient, medicine, or medical, the second row to test, scan, cure, or diagnosis, and the third row to software or program. The rectangles are also subdivided into columns, where each column represents a text segment, as described above. Thus, the leftmost column indicates the first segment, or paragraph, of the document, the column to the right of this indicates the second segment of the document, and so on.

Each square represents the number of hits for the corresponding termset in the corresponding document segment. The darkness of the square indicates the number of times the query occurs in that segment of text; the darker the square the greater the number of hits (white indicates 0, black indicates 8 or more hits, the frequencies of all the terms within a term set are added together).

Thus the user can quickly see if some subset of the terms overlap in the same segment of the document, and can see at what position of the document this overlap occurs. For example, the bottommost TileBar in Figure 1 shows that all three termsets overlap only in the very last subtopical segment; from this the user can assume that there is only a passing reference to medical applications at the end of the document. As another example, in the fourth TileBar, the title implies that the article discusses computer automation of medical admissions (to a veteran's hospital), as opposed to automation of diagnosis. However, the TileBar reveals that terms from the diagnosis termset are indeed well represented in the document.

 
Figure 1:   The TileBar Display on a query about automated systems for medical diagnosis.

TileBars allow users to indicate which part of the document to view. by mouse-clicking on the corresponding part of the representation. that symbolizes the beginning of the document. For example, a user may go directly to a segment in the middle of the text where termsets overlap, knowing in advance how far down in the document the passage occurs. In a newer version of the interface [1], each termset is color-coded and the terms of the termset are highlighted with the corresponding color when the text of the document is displayed.

The version of the interface demonstrated in the videotape allows the user to adjust constraints: minimum number of hits for each term set, minimum distribution (the percentage of tiles containing at least one hit), and minimum adjacent overlap span. In Figure 1 the user has indicated that the diagnosis aspect of the query must be strongly present in the retrieved documents, by setting the minimum term distribution percentage to 30% for the second termset.

A system that simply ranks the documents does not make these kinds of distinctions available to the user. The TileBar ranking order is different than that of standard systems. In the implementation demonstrated here, documents are ranked first by how many segments have overlap of all three termsets, second by the overall frequency of the query terms in the document. We have recently [4] obtained strong improvements in precision at high document cutoff levels by requiring the query to be specified in terms of a list of topics, and then applying two constraints that take advantage of the query structure. Our initial user studies [1] suggest that users seem to find the query format easy to use.

References

1

Marti Hearst, Jan Pedersen, Peter Pirolli, Hinrich Schüetze, Gregory Grefenstette, and David Hull. Four trec-4 tracks: the xerox site report. In Donna Harman, editor, Proceedings of the Fourth Text Retrieval Conference TREC-4. National Institute of Standards and Technology Special Publication, 1996. (to appear).

2

Marti A. Hearst. Multi-paragraph segmentation of expository text. In Proceedings of the 32nd Meeting of the Association for Computational Linguistics, June 1994.

3

Marti A. Hearst. Tilebars: Visualization of term distribution information in full text information access. In Proceedings of the ACM SIGCHI Conference on Human Factors in Computing Systems, Denver, CO, May 1995. ACM.

4

Marti A. Hearst. Improving full-text precision using simple query constraints. In Proceedings of the Fifth Annual Symposium on Document Analysis and Information Retrieval (SDAIR), Las Vegas, NV, 1996.

5

Gerard Salton. Automatic text processing: the transformation, analysis, and retrieval of information by computer. Addison-Wesley, Reading, MA, 1988.

About this document ...

This document was generated using the LaTeX2HTML translator Version 95 (Thu Jan 19 1995) Copyright © 1993, 1994, Nikos Drakos, Computer Based Learning Unit, University of Leeds.