Ch. 7: Supporting the Search Process
Chapter 3 discussed the various theories associated with the information seeking process, and the subsequent chapters described interfaces for supporting query specification, results presentation, and query reformulation -- the standard stages of the search process. This chapter describes interface ideas for other aspects of the search process: search starting points, search history, and interfaces that support the process as a whole. The final section discusses attempts to integrate search into the sensemaking process.
7.1: Starting Points for Search
The first step in addressing an information need is deciding which tools to use and which collections to search over, a process which is sometimes referred to as source selection. Today there are many choices, including phoning, emailing, or texting a friend, reaching for a physical book, going to a physical library, or sitting down at a networked computer and starting up a web browser.
7.1.1: Starting Points in Web Search
For those who go online, today the most common starting point is to open a Web browser and start with a Web search engine. Today, web browsers make that choice even easier by including an always-visible entry form in the browser's “chrome” or by supporting search directly in what used to be the address entry form. But people also commonly start searches from favorite information resources, such as bookmarked Web sites (Teevan et al., 2004). Web browsers have always allowed users to retain bookmarks, but today are making site revisitation even easier by showing usage history as a drop-down menu within the address bar, and matching that history as the user types. The Chrome Web browser shows a grid of thumbnail images of the user's commonly-visited sites directly on the browser's home page.
In the early days of the Web, hyperlinked directories of Web sites were quite popular as search starting points, in part because the set of interesting sites was smaller, and because for many years search engines were perceived as inaccurate and slow (Piontek and Garlock, 1996). Hyperlinked directories like Yahoo's remained popular until the Web became too large both for the editors manually editing the sites, and for users navigating through the hierarchies. (In May of 1996, Yahoo had 200,000 Web sites in its directory, out of what was estimated to be a half a million sites and 21 million pages (Steinberg, 1996).) Web directory sites also became crowded with advertising and other material in the quest to build lucrative “portals,” which led to a backlash against the “clutter.” The use of Web directories has declined markedly, replaced by search engines as starting points.
Traditionally, librarians have prided themselves on acting as information curators, selecting information collections and ensuring the authenticity and reliability of information presented within those collections. There is a general concern that Web search obliterates the distinction among sources, and provides searchers with few cues about the quality or reliability of retrieved information sources (Rieh, 2002). Web search engines do employ internal quality metrics, in part to eliminate spam, and in part to improve ranking, but they are influenced by popularity of Web sites as well as by a computed measure of “authority” for those sites. These measures in some cases correspond to what information scientists deem to be quality measures (Amento et al., 2000), but do not use their notions of quality explicitly. (For instance, a few years back, the top results for web search on the query Joe McCarthy were primarily sites that praised this political figure and claimed he had gotten an unfair treatment by historians, a view significantly outside the mainstream.) What is not in dispute is that Web search result surrogates do not provide adequate information to allow users to judge the underlying quality of the information sources (Rieh, 2002), and those credibility cues that do appear on Web sites can be easily mimicked (Fogg et al., 2001).
To address these concerns, the Google Co-op project provides vetted sites for certain types of queries. For instance, for health-related queries such as tamoxifen, some of the search results after selecting a Co-op refinement link indicate which trusted Web site they have been endorsed by (see Figure 8.1 in Chapter 8). More ambitiously, the Mahalo.com Web site is a relatively new portal that provides a curated set of Web sites and news sources.
Nonetheless, source selection interfaces are still lacking in Web search. Even the relatively easy task of providing search results limited to web site sources with educational as opposed to commercial or political sources is not supported. This may reflect a lack of demand for such kinds of results, or it may be that automatically classifying sources in this manner is not yet possible with sufficient accuracy. Furthermore, most Web users are not accustomed to searching for sources explicitly (although navigational queries are in fact queries for already known sources). Good interface design could overcome this particular barrier.
7.1.2: Starting Points in Online Library Catalogs
For many years, online library catalogs required users to begin by looking through a list of names of sources and choosing which collection to search on (Dempsey, 2006). Often the user had to repeat the same search across different collections within an online library catalog, in part due to technology limitations, and in part due to ownership restrictions over the different collections. More recently, library catalog interfaces have become integrated, allowing the user to issue their query once and see the results for multiple resources in one place. But this produces another problem, as search engine interfaces tend to remove distinctions between sources and place the user into the middle of a Web site or other resource with little information about context. To counter this, library catalog interfaces have recently begun adopting faceted navigation interfaces (Hearst et al., 2002) that will eventually allow the user to group and select search results according to source, along with other attributes (see Figure 7.1). Faceted navigation interfaces have been shown to exhibit good usability results in online library catalogs (Olson, 2007), and are discussed in detail in Chapter 8.
An even more recent innovation can be seen in the Lexis-Nexis interface for statistical databases. This system provides a standard forms-based starting-points interface, but unlike systems of old, allows for searching over a huge collection of tables of data, including allowing search over specific fields such as table titles and table text (Figure 7.2 shows the query forms and search results). This is a powerful interface, but because it uses a parametric design (where users have to select all the fields up front) the user often ends up with empty results sets (as opposed to a faceted interface with query previews, which prevents the user from selecting fields that would yield empty results sets). For instance, if the query shown in Figure 7.2 is modified to select annual data with hits in table titles, no results are returned. An even more radical interface is shown in Figure 7.3 in which the data from the datasets is exposed directly in the search starting point. The user can select alternative views within datasets and can do comparisons, on the datasets directly. This is a strikingly different form of search starting point.
(a)
(b)
7.1.3: Interactive Dialogues as Search Starting Points
Another approach for getting started with search is the interactive dialogue, which provides support for a series of question-answer interactions between the user and the system. Dialogue-based interfaces have been explored since the early days of information retrieval research, in an attempt to mimic the interaction provided by a human search intermediary (e.g., a reference librarian). Early work in the THOMAS system provided a question and answer session within a command-line based interface (Oddy, 1977). Others have defined quite elaborate dialogue interaction models (Belkin et al., 1993). More recently, interactive question-and-answer interfaces have been developed for specialized information seeking tasks, such as choosing a laptop computer (McSherry, 2003). Dialogue-style interactions have not yet become widely used, most likely because they are still difficult to develop for robust performance.
7.2: Supporting Search History
Another way to support the search process is to help users retain the context of their queries, sources, and results sets. The most straightforward way to do this is to record queries and search history and allow users to re-access these records. As seen in Chapter 4, even the primitive TTY-style interfaces of early search systems like Dialog allowed users to name queries and results sets and use those to build up complex Boolean queries. In web-based search systems, query history is sometimes shown as a chronologically-ordered list of most-recently issues queries (as is done in the PubMed interface, as shown in Chapter 1). And as discussed above, Web browser address forms now show previously visited Web pages in a drop-down history list. Search engine toolbars, such as those supplied by Yahoo, Google, and Microsoft search, also support a facility to record and view search history.
Many researchers have experimented with incorporating thumbnails as memory aids in browser and search history. Chapter 10 discusses the use of thumbnail images in search results (suggesting that at best they do not improve upon textual search results displays, but using them in page history seems to be useful).
A particular kind of history mechanism used primarily in vertical web sites, is the breadcrumb. These are especially common on e-commerce sites and sites that use faceted navigation. A distinction is often made between path breadcrumbs that reflect the sequence of links that a user has clicked on since beginning a navigation session (hence earning their name, after the Grimm's fairy tale of Hansel and Gretel), and location or information structure-based breadcrumbs that reflect the site structure, indicating where in the Web site's information architecture the current page is situated. (An example of the path version is shown in Figure 7.1 below the search entry form.) Nielsen, 2007 argues in favor of structural breadcrumbs, noting that showing the site structure is useful for when users arrive at a page directly from a search engine, rather than when navigating to it, and because the Web browser back button works very well for allowing searchers to retrace their actions. Nielsen, 2007 states:
“Despite their secondary status, I've recommended breadcrumbs since 1995 for a few simple reasons: Breadcrumbs show people their current location relative to higher-level concepts, helping them understand where they are in relation to the rest of the site. Breadcrumbs afford one-click access to higher site levels and thus rescue users who parachute into very specific but inappropriate destinations through search or deep links. Breadcrumbs never cause problems in user testing: people might overlook this small design element, but they never misinterpret breadcrumb trails or have trouble operating them. Breadcrumbs take up very little space on the page.”
Rogers and Chapparo, 2003 found in a controlled study with 45 participants that the structure-based breadcrumbs led to a better understanding of site structure than no breadcrumbs. Hearst, 2006b also supports structure-based breadcrumbs, noting that in faceted navigation, the path within each facet should be shown in a separate visual element, to both reinforce the understanding of the facet structure, and to allow for more flexible expansion or retraction of the query.
7.3: Supporting the Search Process as a Whole
Much less well-understood is how to devise interfaces to support the search process as a whole. Marchionini et al., 2000 describe a framework they call agileviews which consists of six kinds of views: primary views, overviews, previews, reviews, peripheral views, and shared views. The primary view is represented by the documents that have been accessed, and the results listing. Overviews show starting points and orient the user to the choices that are available. Previews show the user what will happen if a certain choice is made, allowing for informed decisions of what to do next (these include document surrogates, as discussed in Chapter 4 and query previews as discussed in Chapter 8). Reviews allow for revisiting past choices, essentially providing search history as described above, and peripheral views show information “in the background” that may become of interest but are not the current focus of attention, such as minimized windows of previously viewed information. Shared views show the state of search actions performed by other people. The agileviews framework underscores the need to support fluid interaction among the different views to support the search process.
Not many interfaces support all of this functionality. Most attempts show query history, organized results, and suggested terms in a window or set of views. The Protofoil/Infogrid interface (Rao et al., 1992, Rao et al., 1994, Rao et al., 1995) (see Figure 7.4) was an early example of a visually-oriented interface to support the search process that displayed a number of the views described by Marchionini et al., 2000. The layout of this interface was a grid divided into fixed areas used for different purposes: showing the query, showing retrieval results, viewing the current document, a side panel for storing selected documents, and a lower gutter viewing a history of retrieved documents. The results viewing section of the interface allowed for flexibly moving between different views of retrieval results. There is some evidence that allowing flexible switching between different views of retrieval results is useful (Hearst et al., 1998, Reiterer et al., 2000).
Several of the digital library research systems developed in the 1990s experimented with more elaborate ways to support the search process (Baldonado and Winograd, 1997). For example, in the Ariadne system (Twidale and Nichols, 1998), queries were run against a library catalog system, which produced textual output. The designers converted this text-heavy view into a visualization of a sequence of cards containing thumbnail outlines of the query and result screens. The cards were arranged on three horizontal tiers, where the top row's cards indicated a meta-activity (selecting a command from the system's menu), the middle row represented queries and results, and the bottom row contained documents viewed or saved. This view allowed the user to review and re-visit their sequence of actions (see Figure 7.5).
The DLITE system (Cousins et al., 1997, Cousins, 1997) presented a creative extension on these ideas (see Figure 7.6). It split functionality into two parts: control of the search process and display of results. The control portion was a graphical direct-manipulation display with animation. Queries, sources, documents, and groups of retrieved documents were represented as graphical objects. The user created a query by filling out the editable fields within a query constructor object. The system manufactured a query object, which was represented by a small icon which could be dragged and dropped onto iconic representations of collections or search services. If a service was active, it responded by creating an empty results set object and attaching the query to it. A set of retrieval results was represented as a circular pool, and documents within the result set were represented as icons distributed along the perimeter of the pool. Documents could be dragged out of the results set pool and dropped into other services, such as a document summarizer or a language translator. The user could also make a copy of the query icon and drop it onto another search service. Placing the mouse over the iconic representation of the query caused a “tool-tips” window to pop up to show the contents of the underlying query. Queries could be stored and reused at a later time, thus facilitating retention of previously successful search strategies.
A flexible interface architecture freed the user from the restriction of a rigid order of commands. However, such an architecture must provide guidelines to help get the user started, give hints about valid ways to proceed, and prevent the user from making errors. The graphical portion of the DLITE interface made liberal use of animation to help guide the user. For example, if the user attempted to drop a query in the document summarizer icon -- an operation that is not permitted -- rather than failing and giving the user an accusatory error message (Cooper, 1995), the system took control of the object being dropped, refusing to let it be placed on the representation for the target application, and moved the object left, right, and left again, mimicking a “shake-the-head-no” gesture. Animation was also used to help the user understand the state of the system, for example, in showing the progress of the retrieval of search results by moving the result set object away from the service from which it was invoked.
DLITE used a separate Web browser window for the display of detailed information about the retrieved documents, such as their bibliographic citations and their full text. The browser window was also used to show Scatter/Gather-style cluster results (see Chapter 8) and to allow users to select documents for relevance feedback. Earlier designs of the system attempted to incorporate text display into the direct manipulation portion, but this was found to be infeasible because of the space required (Cousins, 1997). Thus DLITE separated the control portion of the information access process from the scanning and reading portion. This separation allowed for reusable query construction and service selection, while at the same time allowing for a legible view of documents and relationships among retrieved documents. The selection in the display view is linked to the graphical control portion, so a document viewed in the display could be used as part of a query in a query constructor.
DLITE also incorporated the notion of a workspace, or “workcenter,” as it was known in this system. Different workspaces are created for different kinds of tasks. For example, a workspace for buying computer software could be equipped with source icons representing good sources of reviews of computer software and good Web sites to search for price information and links to the user's online credit service.
7.4: Integrating Search with Sensemaking
As discussed in Chapter 3, the term sensemaking refers to the process of interweaving the seeking of information with the interpretation of information. Russell et al., 1993 observe that most of the effort in sensemaking goes towards the synthesis of a good representation, or way of thinking about, the problem at hand. They also describe the process of formulating and crystallizing the important concepts for a given task. Search plays only one part in this process; some sensemaking activities involve search throughout, while others consist of doing a batch of search followed by a batch of analysis and synthesis. Sensemaking is most often used to refer to information-intensive tasks like intelligence analysis, scientific research, and the legal discovery process. But even more mundane tasks like researching information and making reservations for travel could benefit from more helpful interfaces than are available today.
How should sensemaking interfaces differ from search interfaces? Patterson et al., 2001's analysis highlights the many ways in which search tools are lacking for deep analytical tasks. A more supportive search tool would give an overview of the contents of the collection, would help the analysts keep track of what they had already viewed, would suggest what to look for next, would encourage analysts to try new queries, and would find additional documents similar to those already found, but would distinguish duplicates. Studies also suggest that analytical search tools should allow for aliasing of terms and concepts.
In these studies, one source of error resulted from analysts reading documents that occurred early in the sequence of events, when information is less well-established, rather than later reports that explained the phenomena more fully. This suggests that an interface that shows temporal relationships among related documents could improve performance. Another source of error occurred when analysts incorporated inaccurate information into their analysis. Some of the more successful analysts used judgements of source quality in their assessment of the selected documents. Thus, a facility to aid with various forms of quality assessment would also be useful. No current interfaces support all of these goals, but progress is being made especially in the context of supporting intelligence analysts.
As noted above, a sensemaking interface should support the ability to flexibly arrange, re-arrange, group, and name and re-name groups of information. It has often been observed that users use the physical layout of information within a spreadsheet to organize information (Malone, 1983, Nardi, 1993, Shipman et al., 1995). People tend to arrange physical papers in piles around them as an organizational and memory device (Rose et al., 1993, Russell et al., 1993, Robertson et al., 1998, Whittaker and Hirschberg, 2001, Fass et al., 2002), and as Malone, 1983 notes, a function of information organization is to help remind people of things to do, not just help them find information. Studies have found that tools for organizing partially-formed ideas must support informal interaction, in order to avoid interrupting peoples' train of thought (Marshall et al., 1991).
Thus, a number of research and commercial tools have been developed that attempt to mimic physical arrangement of information items in a virtual representation. The Aquanet and VIKI projects (Marshall et al., 1991, Marshall et al., 1994) designed and evaluated tools to help people organize their thoughts as they process information. These interfaces made use of a canvas or workspace, upon which a user could flexibly arrange snippets of retrieved and processed information. The goal was to allow the user to just point and type, without having to stop to categorize or label the information. These interfaces also allowed users to represent relational information, both according to spatial layout and using visual cues, such as using boxes of different shapes for different idea types. Marshall et al., 1994 found that once the user is ready to apply labels, the labeling mechanism should be flexible, and it should be very easy to group different items together into more structured units. It should also be easy to incorporate additional structure when needed. They also developed spatial layout templates that allowed for representations of argument structure and discussion.
Some of these ideas for storing, organizing, and arranging information have appeared in commercial tools like Microsoft's OneNote. In the Web space, a number of tools have been developed to help users keep track of and organize “clippings” of information extracted from Web sites, as seen in the Internet Scrapbook (Sugiura and Koseki, 1998) and the Hunter Gatherer interface (Zhu et al., 2002). More recently, Doncheva et al., 2006 describe a system which allows users to select and store elements from Web pages and arrange them on an information canvas, and visually laying out the components using a combination of user labels and pre-determined templates (such as a map or a grid).
The flexible analysis tools of Marshall et al., 1994 did not focus on how to integrate information search within the sensemaking process. An early system to address the two together was the SketchTrieve interface (Hendry and Harper, 1997). The guiding principle behind SketchTrieve was to support information access as an informal process, in which half-finished ideas and partly explored paths can be retained for later use, and the results combined via operations on graphical objects and connectors between them. SketchTrieve showed sets of related retrieval results as a stack of cards within a folder and allowed the user to extract subsets of the cards and view them side by side. The Data Mountain 3D desktop interface by (Robertson et al., 1998) extended this idea by showing retrieved documents as physical-looking cards in 3D virtual space, allowing the user to group and re-arrange the cards by “pushing” them around the space. (This work was in turn an extension of the Information Visualizer Workspace and Web Forager work (Card et al., 1991, Card et al., 1996) ; see Figure 5.9 in Chapter 5).
The Sandbox system by Wright et al., 2006 takes these ideas even farther (see Figures 7.7-- 7.9). This and companion tools were demonstrated in detail on a complex epidemiological investigation of Avian Flu (Proulx et al., 2006). Sandbox allows for free-form organization of retrieved results, and automatically retains links to the original sources, which can be dragged and dropped to the workspace from its companion retrieval system, TRIST (Jonker et al., 2005) (see Figure 7.7). The tool allows information to exist in the workspace at different levels of representation, from a simple word or phrase, to a set of retrieved documents, to the results of a simulation derived from a different tool (see Figure 7.8). Users can draw arbitrary-shaped lassos to gather scraps of information into a new group. A simple up-line gesture allows for temporary zooming in for detail, or to zoom in or out of the entire view.
Sandbox has a simple mechanism for making room at any point on the workspace. In response to a simple up-arrow gesture at the desired spot on the workspace, the system automatically clears an open space, moving the existing pieces of information out of the way, but retaining the relative orientation and layout of the pre-existing items. The amount of room cleared is a function of the size of the gesture. The system also allows the user to “grasp” one item and use it to “knock away” other items (Hutchings and Stasko, 2002).
Sandbox offers a structured template showing how a given problem has been reasoned about in the past (for example, known paths for transmission of infectious disease, see Figure 7.9). The analyst is instructed to find supporting information for each part of the process, thus drawing attention to information that should be available but currently is missing. The system also allows the analyst to represent structured arguments, in the form of hypotheses, and support for and against. Transparency is used to indicate the degree of confidence in the various statements, to remind the user that not all information can be relied upon. The system displays a glyph in the corner of the window that shows a running total of supporting and refuting evidence. The visualization of retrieval results for the TRIST search system that underlies Sandbox is discussed in Chapter 11.
7.5: Conclusions
This chapter has described interfaces for search starting points, for retaining search history, for supporting the entire process, and for making sense of retrieved information as it is gathered. This technology is not as mature as the components discussed in the preceeding three chapters. Most pressing is the need for better interfaces to support complex information seeking tasks. By way of illustration, it is still a complex and time consuming task to plan a trip to a new location, even with the bountiful information available online. This difficulty lies in part with the fact that this task requires an understanding of the needs and preferences of the individual or group who is doing the search, and because the results of such a search have interlocking parts and dependencies. Better supporting these aspects of search are active areas of research, and new ideas should be expected to appear in future.