Query-based Navigation in Semantically Indexed
`Hypermedia
`
`Daniel Cunliffe
`Hypermedia Research Unit,
`Department of Computer Studies,
`University of Glamorgan,
`Pontypridd, CF37 1DL, UK.
`Tel: +44 1443 482718
`E-mail: djcunlif@glamorgan.ac.uk
`
`Carl Taylor
`DK Multimedia,
`9 Henrietta Street,
`Covent Garden,
`London, WC2E 8PS, UK.
`E-mail: cdtaylor@dkmm.co.uk
`
`Douglas Tudhope
`Hypermedia Research Unit,
`Department of Computer Studies,
`University of Glamorgan,
`Pontypridd, CF37 1DL, UK.
`Tel: +44 1443 482271
`E-mail: dstudhope@glamorgan.ac.uk
`
`ABSTRACT
`This paper discusses an approach to navigation based on
`queries made possible by a
`semantic hypermedia
`architecture. Navigation via query offers an augmented
`browsing capacity based on measures of semantic closeness
`between
`terms
`in an
`index space
`that models
`the
`classification of artefacts within a museum collection
`management system. The paper discusses some of the
`possibilities that automatic traversal of relationships in the
`index space holds for hybrid query/navigation tools, such as
`navigation via similarity and query generalisation. The
`example scenario suggests that, although these tools are
`implemented by complex queries, they fit into a browsing,
`rather than an analytical style of access. Such hybrid
`navigation tools are capable of overcoming some of the
`limitations of manual browsing and contributing to a smooth
`transition between browsing and query. A prototype
`implementation of the architecture is described, along with
`details of a social history application with three dimensions
`of classification schema in the index space. The paper
`discusses how queries can be used as the basis for
`navigation, and argues that this is integral to current efforts
`to integrate hypermedia and information retrieval.
`
`KEYWORDS: Hypermedia, semantic index space, semantic
`closeness, query-based navigation, cultural heritage, museums
`
`1 INTRODUCTION
`The conventional boundaries separating hypermedia from
`related disciplines are dissolving to yield a confluence of
`hypermedia, information retrieval, and multimedia database
`technologies [21, 27]. Traditionally, query-based retrieval
`methods have been considered a very different kind of
`interaction to the browsing commonly associated with
`hypertext or hypermedia. Queries require prior planning and
`a more analytical approach with greater cognitive overheads
`Permission to make digital/hard copies of all or part of this material for
`personal or classroom use is granted without fee provided that the copies
`are not made or distributed for profit or commercial advantage, the copy-
`right notice, the title of the publication and its date appear, and notice is
`given that copyright is by permission of the ACM, Inc. To copy otherwise,
`to republish, to post on servers or to redistribute to lists, requires specific
`permission and/or fee.
`Hypertext 97, Southampton UK
`© 1997 ACM 0-89791-866-5...$3.50
`
`for the user. Recent research in information retrieval (IR)
`has endeavoured to break down this barrier and build
`hypertext interfaces to IR systems. The separation of media
`items or information content from a link base is now widely
`accepted in hypermedia systems [22, 24]. In IR-oriented
`applications, an important distinction is that between the
`document space and some form of index space [18, 30, 2].
`Terms in the index space are descriptors for media items in
`the document space—text passages, images, etc. Many
`hypertext-IR retrieval systems have been based on statistical
`analysis of terms automatically extracted from free text, [17,
`4]; the index space is automatically created and ‘vector
`space’ similarity coefficients measure degrees of match
`between queries and media items, or between two media
`items. A different approach to IR is based on (usually)
`manually created index spaces, where semantic relationships
`exist between index terms [35, 25], and this is the approach
`we follow. In the cultural heritage domain, traditional
`practice includes the use of thesauri or classification systems
`as controlled vocabularies to index the media items in the
`collection. It is also possible to combine the two IR
`approaches [2, 47] with a thesaurus used to expand query
`terms.
`
`One possibility that a semantic index space provides is an
`organised set of browsable nodes and links as a navigation
`aid to the associated layer of media items [1, 33]. The user
`can browse the index space, ‘beam down’ [12] to view
`media items of interest, and conversely ‘beam up’ to the
`index space from media items. The inclusion of semantic
`information in the index space, however, provides the
`opportunity for knowledge-based hypermedia systems that
`can provide intelligent navigation support and retrieval, with
`the system taking a more active role in the navigation
`process rather than relying purely on manual browsing. The
`navigation via query described here offers an augmented
`browsing capacity based on measures of semantic closeness
`between
`terms
`in an
`index space
`that models
`the
`classification of artefacts within a museum collection
`management system. This paper contributes
`to
`the
`integration of hypermedia and information retrieval from a
`hypermedia perspective that focuses on the notion of
`
`87
`
`PETITIONERS
`EXHIBIT 1015, Page 1
`
`

`

`navigation and hybrid query-navigation tools. All forms of
`navigation in our system are ultimately based on queries. We
`have been particularly interested in reasoning over the
`relationships in the index space (see also [15, 26]).
`Traversal of transitive semantic relationships makes possible
`imprecise matching between query and media item, or
`between two media items. This paper gives examples of
`such traversal algorithms in different types of tools and
`navigation techniques in both index and document spaces.
`
`1.1 What is navigation anyway?
`In the most extensive discussion of navigation to date,
`Marchionini
`[27]
`first distinguishes analytical
`from
`browsing strategies, and then identifies navigation as one of
`four browsing
`techniques. Traditional IR queries are
`considered an analytical strategy, requiring planning, greater
`cognitive overhead, more goal-driven
`than browsing
`techniques and more batch-oriented. Browsing strategies, on
`the other hand, are more heuristic, interactive, data-driven,
`and opportunistic (discovery as opposed
`to search).
`Marchionini recommends that retrieval systems should
`support both strategies. Much of the HCI and hypermedia
`literature equates hypermedia navigation with browsing.
`Marchionini however, distinguishes
`the four browsing
`techniques: Scan, Observe, Monitor, and Navigate with
`different degrees of interactivity, cognitive effort, goal-
`orientation and
`information structure. Navigation
`is
`characterised by high
`interactivity
`in a
`structured
`environment with the destination seldom pre-determined.
`Navigation is often a compromise between user and system
`responsibility; an incremental process with the user making
`choices from directions and feedback provided by the
`system. This aspect of discovery emphasised by Marchionini
`finds echoes in recent calls by museum professionals for a
`different approach to information retrieval in future public-
`access systems: “pathways to discovery instead of answers
`to questions ... Users seek information that is unknown,
`often not knowing what they want until it is seen” [38], a
`guide to “the serendipity of information” [39].
`
`Another notion intrinsic to navigation is the idea of
`movement through an information space. The concept of a
`current position, or context, is crucial. The next navigational
`move is based on the results of previous navigations;
`hypermedia navigation is an example of an inter-referential
`mode of interaction [16]. This typically occurs when output
`(results of a query, destination of a link) forms part of the
`next input to the system. However in some cases, a previous
`input to the system is involved (navigation history, relevance
`feedback). Waterworth and Chignell [46] describe a model
`for information exploration with dimensions that distinguish
`target orientation (the degree of goal directedness in the
`user's cognitive attitude) from responsibility (the proportion
`of the responsibility for controlling the search shared by user
`versus system). Our query-based navigation corresponds to
`'mediated browsing’ in their model. They observe that rather
`than being dichotomies, these dimensions are continuum,
`
`88
`
`and that systems should offer a smooth transition between
`querying and browsing.
`
`Hypermedia has tended towards the manual navigation of
`explicitly authored links, while IR has traditionally been
`oriented to the construction of analytic queries returning a
`ranked set of documents. Recent work in both disciplines
`has attempted to integrate these two approaches, although
`they still tend to appear as distinct phases of a user session.
`IR has gone a long way to incorporating browsing into
`essentially analytical
`strategies via
`the navigational
`interpretation of query results. If we consider hypermedia to
`have a different perspective, the issue is the incorporation of
`query
`into essentially browsing
`strategies. Modern
`hyperbase
`systems, where computed
`links can be
`implemented as queries on underlying linkbases, offer
`opportunities for rich and hybrid forms of query/navigation.
`For example, Garzotto et al [20] and Amman et al [5] have
`investigated how navigation can be based on database
`queries. Microcosm [24] supports IR free text queries
`(which do not require exact match) in its Computed Links,
`and various systems have made use of IR techniques to
`generate guided
`tours or
`trails [9, 23, 7]. Several
`applications now include content-based image matching as a
`component of navigation.
`
`A query-based interpretation of navigation should allow
`augmented navigation tools to dynamically return the results
`of complex queries as link destinations, but retain the
`essential character of hypermedia navigation. This can be
`characterised by interaction that is relatively simple, does
`not require high cognitive effort, but involves user choice of
`navigational moves. A session will have a broad direction
`but not be so goal-oriented as to leave no room for
`discovery. Next moves will be based on the current position
`and context, but the query component can overcome an
`over-reliance on the immediate neighbourhood and context
`that some see as a disadvantage of navigational browsing
`[46].
`
`2 CULTURAL HERITAGE INFORMATION SYSTEMS
`Within the cultural heritage domain the advantages of a
`controlled vocabulary with semantic relationships between
`terms are well recognised [3]. Controlled vocabularies, such
`as ICONCLASS [11], and the Art & Architecture Thesaurus
`[32], already exist for large scale multimedia information
`systems. Due to the heavy investment of both time and
`expertise in such vocabularies and their role in professional
`practice, they will almost certainly have a part to play in
`future museum based multimedia information systems. For
`example, the ongoing European Aquarelle project makes use
`of a Semantic Index System, based on an object-oriented
`semantic model [14]. Such semantic schema provide a
`controlled vocabulary used by a human indexer to impose a
`classification on a set of objects, rather than automatically
`deriving classifications directly from the features of the
`objects themselves. Although manual classification is time
`
`PETITIONERS
`EXHIBIT 1015, Page 2
`
`

`

`intensive, the skilled indexer can identify an object and
`place it within its wider context in a way that automatic
`content-based
`indexing alone cannot. In practice
`the
`classification of a complex media
`item, such as a
`photograph, is an expert task, and efforts must be taken to
`produce frameworks that encourage consistency in index
`term selection, although some variation can be compensated
`by intelligence in retrieval tools [29].
`
`In general the vocabulary is organised into some form of
`semantic
`network,
`for
`example
`a
`hierarchy
`of
`broader/narrower terms and sometimes equivalent and
`related terms. This schema provides an appropriate frame of
`reference for both indexing and search. Museum information
`systems have typically concerned themselves with the
`management of collections, but there is a growing interest in
`public access systems which allow the non-expert museum
`visitor
`to obtain
`information about
`the museum,
`its
`collections, exhibitions, and so forth. The non-expert user
`may, however, be unfamiliar with the classification space,
`term vocabulary and organisation of the material and as a
`result of this, may have difficulty navigating through multi-
`faceted index spaces. Unlike tightly focused applications,
`cultural heritage information is often multi-purpose, with an
`element of discovery in its use. There is therefore a need to
`provide advanced navigation tools to assist exploration of
`such complex information systems. This paper is intended as
`a step in this direction, describing prototypes of techniques
`that can form part of such future tools.
`
`3 THE SEMANTIC HYPERMEDIA ARCHITECTURE
`The SHA [41] can be described in terms of three main
`layers: a media-base, a semantic index space, and an
`application layer. The application layer supports the user’s
`interaction with the system, giving access to the set of
`navigation tools provided in a particular application and
`providing presentation capabilities for media items. The
`media-base contains the actual media items, such as
`photographs, sound recordings and so on, together with
`authored links between media items. To date within the
`project we have not been particularly concerned with the
`implementation of the media base, we simply assume that
`media items can be retrieved using a reference (to the media
`item as a whole—we do not currently support sub-
`component addressing).
`
`Together with the media-base is a link-base of explicit
`(authored) node-to-node
`links.
`(We
`are
`currently
`investigating how to integrate the system with existing
`hyperbase technology.) However, the focus of our research
`has been the semantic index space. The index space holds
`the descriptors (index terms) of the information items in the
`media-base and the semantic relationships that give the
`terms meaning.
`
`Both the index space and the explicit link-base are
`implemented on an underlying database, a form of Binary
`
`application layer
`
`semantic index space
`
`link-base
`
`media-base
`
`Figure 1: Semantic Hypermedia Architecture overview
`
`restricted set of primitive
`Relational Store with a
`relationships [10]. Thus, the index space consists of a set of
`binary relationships of the form:
`
`<object relationship subject>
`
`where object and subject are entities, such as media item
`IDs or index terms, and relationship is one of the semantic
`relationships permitted within the model, such as A-Kind-Of.
`A
`restricted
`set of core
`relationships have been
`implemented, including notions of hierarchy, aggregation,
`and
`instantiation. The database supports
`three basic
`operations, insertion, deletion and query, all at the level of a
`binary relationship triple. Insertion simply adds a triple to
`the store. Deletion and query are based on Standard
`Associative Forms [19]. These SAFs are essentially pattern
`matching queries on the database, using wildcards. Thus a
`SAF of the form:
`
`<? A-Kind-Of Community Life>
`
`where ? is a wildcard, will match all triples with the
`relationship A-Kind-Of which also have object Community
`Life. A SAF query with no wildcards is a test for the
`existence of a specific triple within the database. SAFs can
`be used for conventional query interactions with the
`database. They can also form the basis of the query-
`supported navigation described in Section 5.1. A constraint
`manager ensures the integrity of the relationships within the
`database.
`
`The results described in this paper have been generated by a
`research prototype Lisp/HyperCard implementation. The
`database is a (memory-resident) collection of unordered
`triples, manipulated by LISP routines. A recent re-
`implementation uses a relational database with the constraint
`manager being implemented as stored procedures, and an
`expanded set of core relationships. An investigation into the
`appropriate table and index structures is underway.
`
`89
`
`PETITIONERS
`EXHIBIT 1015, Page 3
`
`

`

`3.1 Social history museum application
`We chose the domain of social history to investigate a
`particular application of the SHA, a prototype interactive
`social history exhibit of the town of Pontypridd, mainly
`consisting of photographs from
`the archives of
`the
`Pontypridd Historical and Cultural Centre, with some text
`and oral history material.
`
`For a subject index, we adopted SHIC—the Social History
`and
`Industrial Classification
`[37]. Clearly,
`other
`classification schemes could have been modelled in the
`SHA, however SHIC seemed appropriate for the subject
`area. SHIC is a museum standard which classifies objects
`according to their context of use, or the sphere of human
`activity with which they are primarily associated, as opposed
`to the more common type of object name classifications. It is
`used by many UK museums (and is currently being
`considered by the Aquarelle project) for cataloguing and
`collection management of social history material—we
`adapted it for public presentation. A small scale in-gallery
`evaluation of an early version that did not involve advanced
`navigation tools was conducted earlier in the project [43].
`The four first level classification divisions, Community Life,
`Domestic and Family Life, Personal Life, and Working Life,
`form separate hierarchies (facets). Each of these terms is
`further subdivided to form a hierarchy of more specific
`subterms five levels deep. Four of these are modelled in the
`SHA, where SHIC terms are connected by semantic
`relationships to form a hierarchical subject index space (see
`Figure 2). Since photographs are composite objects, and an
`object can have several uses, multiple SHIC
`terms
`(sometimes from the same facet) were used to index each
`item, and an earlier version of the system allowed Boolean
`queries of SHIC terms [43].
`
`Frequently a media item will be indexed by several terms, or
`combinations of terms used in a query. One interesting
`approach to the index space is to conceive it as a
`‘hyperindex’ [12, 13]. If connectors between index terms
`have been defined, the term list becomes an ‘index
`expression'. From this a ‘power index expression’ can be
`formed, a lattice-like structure which is the set of all sub-
`expressions. This lattice can be navigated (‘query by
`navigation’); the user interactively broadens or narrows the
`search by specifying
`the current
`focus within
`the
`corresponding multi-term query. Arents and Bogaerts [6, 7]
`have extended this approach
`to semantically coupled
`thesauri with ‘semantic hyperindexing’ and computed links.
`Here a priori knowledge of the different facets (dimensions)
`in their subject thesaurus is used to create rules for
`meaningful (or interesting) combinations of terms from
`different facets; some combinations of terms are disallowed.
`These rules can be used to dynamically generate navigation
`possibilities, in particular valid links from the current node,
`or suggested trails. We do not currently have connectors
`between terms or make use of semantics associated with the
`combination of facets, however the SHIC traversal algorithm
`
`90
`
`yields degrees of similarity between sets of terms [45] and
`this is used in the navigation via similarity tool.
`
`Empirical studies of user behaviour in the humanities and
`cultural heritage areas show that time and space are
`important access routes to information [8, 28]. Accordingly,
`we decided on an index space with three ‘dimensions',
`representing the three primary categories we chose to model
`social history: subject (SHIC), time and space. Media items
`are indexed along the three dimensions. The temporal
`schema models time as a linear sequence of discrete, non-
`divisible elements—the smallest element of time being the
`year. These elements are combined to form higher level
`intervals each with a start year and an end year. Two classes
`of interval are used within the schema, calendar intervals
`and cultural intervals. Calendar intervals define a hierarchy
`of constituent subintervals, e.g. century, decade, year, and so
`on. Cultural intervals reflect a particular perspective on
`significant subdivisions of the timeline, e.g. the Victorian
`Era, or World War I. Temporal query operators, such as pre,
`post and during are available to the user and can be applied
`to both types of interval.
`
`The spatial schema is based on the notion of a geographical
`area connected to other areas via semantic links; unlike
`Geographical Information Systems there is no geometric
`data model—in future work we wish to include co-ordinate
`information. The schema is made up of four components,
`compositional
`relationships, a geographic area
`type
`hierarchy,
`proximity
`relationships,
`and
`temporal
`relationships. Compositional Part-Of relationships are used
`to construct a regional hierarchy, where each area is
`composed of a set of non-overlapping sub-areas. Within the
`prototype the fundamental unit at the lowest level of the
`hierarchy is the street, although prominent landmarks can be
`defined as a composite part of a street. The geographic area
`type hierarchy provides a framework for the conceptual
`classification of geographical areas, e.g. towns, streets,
`bridges and so on. Although the compositional relationship
`allows the modelling of topological structure, it does not
`provide for
`the modelling of proximity relationships
`between geographic terms. To represent this, a Next-To
`relationship was added, allowing simple qualitative relations
`of proximity to be expressed. Temporal relationships are
`included within the geographical schema to model change in
`the definition or function of geographical areas over time
`[40]—a geographical term exists at a specific time.
`
`4 SEMANTIC CLOSENESS MEASURES
`Semantic closeness measures in the SHA are based on
`traversals of the relationships in the classification schema.
`Time and space are first class categories—each of the three
`dimensions of the classification schema has
`its own
`closeness algorithm. A brief description is given here, for a
`more extended treatment and references to related work, see
`Tudhope and Taylor [45]. The derivation of measures of
`closeness over semantic index spaces is a complex issue; a
`
`PETITIONERS
`EXHIBIT 1015, Page 4
`
`

`

`are possible.
`closeness
`indications of
`variety of
`Consideration must also be given to the validity of the
`measures for the intended application and target users, the
`efficiency of
`the
`algorithms, parameterisation
`and
`tailorability of the measures, and the visualisation of the
`effect of a particular choice of parameters. For the purposes
`of the navigation tools described in this paper, the semantic
`closeness measures can be viewed as replaceable modules.
`The types of navigation described below do not depend on a
`particular implementation of semantic closeness.
`
`Semantic distance between two terms in the conceptual
`schema is essentially based on the minimum number of
`semantic relationships that must be traversed in order to
`connect the terms [34]. Each traversal diminishes the
`semantic closeness measure by a given factor and the level
`of diminishment determines the generality of the terms that
`will be considered semantically close. Generally, the lower
`the value the more terms will be considered close. In
`addition to determining whether or not two given terms are
`semantically close, the semantic distance can also be used to
`retrieve the set of terms that are semantically close to a
`given term. Refinements of the semantic closeness measure
`have included different diminishment values for each type of
`semantic relationship and a weighting according to the level
`of specialisation of the traversal. This weighting is based on
`an intuitive belief that terms separated by a traversal at a
`high level of specialisation are likely to be closer than terms
`separated by a traversal at a lower level, effectively
`considering the granularity of classification at different
`levels within the classification schema. Therefore the
`diminishment value is determined by both the type of
`semantic relationship and the level of specialisation at which
`the relationship occurs. Measures of closeness between sets
`of classification terms have also been implemented.
`
`A similar approach to semantic closeness is adopted for the
`geographical schema, traversals are over the Next-To
`relationships and again a diminishment factor is used. The
`temporal closeness function evaluates similarity between
`both years and temporal intervals. The measure is based on a
`weighted combination of the distance between the two
`temporal periods, the overlap between the periods, and the
`degree of separation between the periods. To measure
`similarity over the three dimensions of the schema, the
`semantic closeness measures are applied for each individual
`dimension and
`the results are
`intersected. A more
`satisfactory approach would be some form of unified
`similarity coefficient over the three dimensions [44].
`
`5 NAVIGATION BY QUERY
`A generic model for hypermedia navigation involves three
`elements: a source context, a destination context, and a
`transformation function [20]. Navigation can be viewed as a
`change in focus from the current context to another context.
`Conventionally, the contexts are media items and the
`transformation is achieved by the selection of an explicit
`
`link. In systems supporting computed links the media item
`itself is often not the context that is being transformed,
`rather links are defined via higher level abstractions, such as
`feature vectors or index terms. The resulting destination
`context may be a media item, but it could also be a new set
`of index terms. Queries can be said to be supporting
`navigation when the result of the query is mediated by the
`current location in which the request is framed.
`
`Dynamic query based navigation is more flexible than fixed
`link navigation in that the results of navigation do not have
`to be pre-processed or fixed by the author of the hyper-
`media. Changes in the structure of the link space can be
`automatically reflected in the links presented to the user
`without the complex task of re-engineering the links. Query
`as navigation also frees the user from the need to follow
`navigational trails engineered by the hypermedia author,
`instead they are able to dynamically generate their own
`trails.
`
`5.1 Query-supported navigation
`In our system, all navigation at implementation level is
`based entirely on queries to the database, by means of the
`Standard Associative Forms (SAFs) defined in Section 3.
`The navigation tool in the application layer queries the
`database via the SAFs, and the resulting triples are passed
`back to the application layer. The triples can then be post-
`processed for expression in a particular navigation tool.
`
`The queries to the database can be simple or complex. More
`complex queries result in the ‘augmented navigation’
`described below. Conventional hypermedia navigation
`techniques, including both local and global browsers, guided
`tours, and Boolean queries can be implemented by relatively
`simple underlying queries [10]. For example, a hierarchical
`local browser, conventionally implemented by authored
`links, can be implemented using dynamic queries as the user
`progresses through the application. In order to implement
`guided tours, an additional relationship, Next, was defined
`to impose an ordering on the tour components. Although
`Next is essentially an explicit link between media items, the
`links are not embedded in media items but held in the link-
`base. Four of Trigg's [42] guided tour navigation functions
`were implemented by queries on the store: Start, Next,
`Previous, and Jump.
`
`The user can browse through the SHIC classification (again,
`based on SAF queries) and collect SHIC terms, which can
`be combined with geographical and temporal terms into a
`combined query. A SHIC query can return all items indexed
`directly by a term and those indexed by specialisations of
`that term. The user can request the number of postings for a
`SHIC term (number of media items indexed by the term).
`All results, or destinations of navigational moves, are
`returned in the media viewer, through which the user can
`browse. An item in the Media Viewer can be used as the
`starting point for subsequent navigations. We have also
`
`91
`
`PETITIONERS
`EXHIBIT 1015, Page 5
`
`

`

`explored a variety of more complex queries as navigational
`tools based on the measures of semantic closeness over the
`index space discussed in Section 4.
`
`5.2 Query generalisation
`When a (multi-dimensional) query fails to return sufficient
`results, the user can ask the system to ‘Suggest Alternatives'.
`Query generalisation attempts to find similar queries which
`will return media items. Generalisation is based on a
`combination of semantic closeness with the number of
`media items (postings) classified under a term, either
`directly or indirectly. To generalise a single query term,
`semantically close terms are retrieved and ranked according
`to their measure of similarity to the query term, and the
`number of postings. The user is then able to select from the
`ranked alternatives to generalise the query. When multiple
`dimensions are involved, an analysis of the query is
`performed to identify the dimension performing poorly. The
`user can interactively indicate the importance of the
`different dimensions in the generalisation.
`
`5.3 Geographical guided tour
`In the interface to the geographical browser, the index space
`is ordered alphabetically by street name, corresponding to
`the common ‘A-to-Z’ gazetteer. The schema, however,
`models proximity relationships between the geographical
`terms. Thus, the user can select two street names and ask for
`a ‘geographical walk’ through the media items in the system
`associated with the shortest path between them. In effect,
`this is a dynamically generated guided tour based on
`automatic traversal of the spatial proximity relationships in
`the index space
`
`5.4 Navigation via similarity
`When a previous navigation has resulted in an item of
`interest, the user can request the system to find ‘similar’
`items. This is a complex query involving similarity measures
`based on the different semantic closeness algorithms in each
`index dimension. Resulting items are located ‘nearby’ in the
`index space, but need not be exact matches. In the SHIC
`index, the query involves expansion of the multiple terms
`used to index the media item (see the navigation scenario
`below). Like query relevance feedback in IR [36], the
`similarity tool involves the user selecting results of previous
`interactions as being of particular
`interest, although
`relevance feedback
`is usually associated with
`terms
`automatically extracted from free text. Relevance feedback
`is richer and more complex, in that it typically involves a
`user selecting several items to refine the previous query, and
`in the possibility of ‘negative’ examples being indicated.
`The feedback is used to refine the probabilistic relevance
`weights and possibly the terms employed in the query.
`Navigation via similarity is less goal and more discovery
`oriented, and is simpler as only one media item is selected
`and there is no continuing query to refine. It is more a
`navigational
`type of
`interaction with
`lower cognitive
`overhead, and allows the user to treat it as a browsing type
`
`92
`
`of tool. Usability studies are needed to identify the level of
`cognitive complexity that can be supported before losing the
`experience of browsing. Note that similarity navigation
`originates and ends in the document space without need to
`refer to the index space [30]. There does appear to be some
`empirical evidence that users find this an intuitive access
`method [31].
`
`5.5 Navigation scenario
`Although the following navigation scenario has been
`artificially constructed to combine the different navigation
`tools, it is intended to motivate and illustrate hybrid query-
`navigation. Note that browsing occurs in both index and
`document spaces, that at different times in the scenario focus
`is switched between index and document space ('beaming
`up/down’) and between dimensions of the index space, and
`that there is an element of discovery as the purpose of the
`user's activities evolves. We see examples of fairly simple
`browsing with low cognitive effort when the user is
`navigating through the SHIC and geographic index spaces,
`and when the user is ‘scanning’ though the media item
`results in the Media Viewer. We also see more complex
`queries, such as Generalisation and Similarity involving
`automatic traversal of terms in the index space, cloaked a