information access tasks and evaluation for personal lifelogs

The Second International Workshop on Evaluating Information Access (EVIA), December 16, 2008, Tokyo, Japan

Information Access Tasks and Evaluation forPersonal Lifelogs

Gareth J. F. Jones Cathal Gurrin Liadh Kelly Daragh Byrne Yi ChenCentre for Digital Video Processing

Dublin City University, Glasnevin, Dublin 9, [email protected]

Abstract

Emerging personal lifelog (PL) collections containpermanent digital records of information associatedwith individuals’ daily lives. This can include mate-rials such as emails received and sent, web contentand other documents with which they have interacted,photographs, videos and music experienced passivelyor created, logs of phone calls and text messages, andalso personal and contextual data such as location(e.g. via GPS sensors), persons and objects present(e.g. via Bluetooth) and physiological state (e.g. viabiometric sensors). PLs can be collected by individu-als over very extended periods, potentially running tomany years. Such archives have many potential ap-plications including helping individuals recover par-tial forgotten information, sharing experiences withfriends or family, telling the story of one’s life, clin-ical applications for the memory impaired, and funda-mental psychological investigations of memory. TheCentre for Digital Video Processing (CDVP) at DublinCity University is currently engaged in the collectionand exploration of applications of large PLs. We arecollecting rich archives of daily life including textualand visual materials, and contextual context data. Animportant part of this work is to consider how the ef-fectiveness of our ideas can be measured in terms ofmetrics and experimental design. While these studieshave considerable similarity with traditional evalua-tion activities in areas such as information retrievaland summarization, the characteristics of PLs meanthat new challenges and questions emerge. We arecurrently exploring the issues through a series of pilotstudies and questionnaires. Our initial results indicatethat there are many research questions to be exploredand that the relationships between personal memory,context and content for these tasks is complex and fas-cinating.Keywords: lifelogging, human digital memories(HDMs), personal information management, user in-formation needs, personal lifelog access evaluation

1 Introduction

Early information retrieval (IR) evaluations wereconcerned with library records and academic sources,while much work on summarization (or abstracting)focused on scientific documents. More recently thedevelopment of standardised evaluation IR workshopssuch as TREC, CLEF and NTCIR focused initially onretrieval from newspaper document collections. Sub-sequently they have expanded the scope of their in-vestigations to other tasks which have individual char-acteristics. For example, web data, patent and legalcollections, and multimedia content. Work on infor-mation extraction and summarization, such as that re-ported at MUC and DUC workshops has similarly fo-cussed on well defined published texts such as newsreports. A common characteristic of all of these datasources is that, subject to suitable distributional agree-ments with copyright owners, this data can be dis-tributed to the signed up participants of the evalua-tion exercises. Further it has generally been possibleto capture and study the information needs or searchrequirements of expected users of these types of col-lections. These have then been used to develop ex-perimental search topics or other user requirements toenable exploration of the task effectiveness for thesecollections. The success of these systems then beingassessed based on manual post hoc assessment of thedata by analysing documents retrieved or submitted forassessment by workshop participants. This frameworkfor evaluations makes a number of assumptions aboutthe way in which the data can be used and the test eval-uation collections designed.While these methods work well, albeit with lim-

itations, for these information sources which are insome sense published, formally structured or publi-cally available, other types of collections introducenew challenges for the design evaluations for informa-tion access tasks. Of particular interest to this paper isthe emergences of Personal Lifelogs (PLs), sometimesreferred to as Human Digital Memories (HDMs)1.

1Although referring to them as HDMs is somewhat misleading

― 75 ―


Recent developments in data capture and storagetechnologies mean that it is now becoming possibleto capture and store electronic records of many of anindividual’s life activities. PLs can most obviouslystore materials such as emails that an individual hassent and received; text documents from the web andelsewhere that they have read, written, or downloaded;multimodal footage from their life experiences suchas photographs taken or received, videos recorded orviewed, and music created or listened to. But can alsoinclude details of places visited derived from locationdata (e.g. harvested from GPS sensor tracklogs), de-tails of people met and objects present (e.g. via Blue-tooth sensors), and details of the subject’s physiolog-ical state (e.g. via biometric sensors) when informa-tion was encountered. This information can be cap-tured from a range of increasingly ubquitous devicessuch as personal computers, mobile phones, cameras,video recorders, audio recorders, personal GPS sen-sors, Bluetooth sensors embedded in various devices,and small portable biometric sensors.A number of research projects are underway ex-

ploring the collection of large heterogeneous PLs,probably the best known of which is the MicrosoftMyLifeBits initiative [5]. While the development ofintegrated PLs is still not commonplace, many indi-viduals are already collecting personal archives of oneor more of these data streams from their lives. It isquite reasonable to expect such archives to grow mas-sively in size, scope and complexity in the comingyears. The ability to gather and store all this personaldata is a considerable achievement in itself. How-ever, it is probably only worthwhile if these collectionscan be usefully employed for some purpose. Fortu-nately such archives have many potential applicationsincluding search activities helping individuals recoverpartial forgotten information, sharing experienceswithfriends or family, telling the story of one’s life, clinicalapplications for the memory impaired, and fundamen-tal psychological investigations of memory. In order toprovide these applications in an effective way new ap-proaches to information access and extension of exist-ing ones are required, and in order to understand theirbehaviour and effectiveness, issues of evaluation needto be addressed and applicable test procedures defined.To date little work has been appearingwhich addressesthis need for new evaluation strategies.In this paper Section 2 first examines the similari-

ties and differences between PL and more traditionalinformaiton collections, Section 3 looks at some of thetasks for which PLs might be applied, while Section4 describes some of the metrics which can be usedto measure performance over these tasks, Section 5summarises work in PL applications to date, Section6 describes current work in our group at Dublin CityUniversity, and finally Section 7 concludes with some

in terms of the accepted definitions of memory.

thoughts about future initiatives in evaluation activitiesfor PLs.

2 Personal Lifelogs

Information access applications for PLs presentnew challenges and opportunities for technologies andevaluation. In this section we explore the differencesbetween PLs and more traditional content archives. Akey feature of PLs is that they will often incorporatea combination of many types of media, including au-dio, video, images, and many types of textual content.There is also the potential for a large percentage ofnoisy data in these archives. This can arise since muchof the content is captured passively by computing de-vices without active participation by the user. The re-sult of this approach to capture will mean that whilethere will be high coverage of potentially interestingand important life experiences, it is also likely thatmuch of the captured material will be of no future in-terest to anyone. In addition, since people often returnto the same information many times, and in the caseof some items they will be revised incrementally overtime, many items in the archive may be very similar,repeatedly covering the same topic. Other differencesbetween PLs and traditional collections include thatthe user may have forgotten about an item or event,and they may not even be aware that a particular pieceof data was captured and is available in the PL.When performing information access tasks from

within a PL users may not be able to describe clearlywhat they are looking for. In many cases items maynot have formal textual descriptions, meaning that theycannot be retrieved using standard text or meta-tagbased retrieval methods. Another is that items maynot be joined by inter-item links, meaning link struc-ture cannot be utilized in the retrieval process. How-ever, one feature with intriguing implications for in-formation access is that items can be linked based ona variety of content and context attributes. All thesefeatures mean that there are new technical challengesassociated with the PL domain.One further crucial difference is that a PL is ob-

viously personal to the individual collecting the datawhich raises many issues. Since they record the per-sonal experiences of the individual whose life theycapture, this user will often have unique memories as-sociated with the items in the archive. While it is ac-knowledged that memory and various cues play a vitalrole in retrieving an item from human memory, currentIR systems make little attempt to exploit what peopleremember about items. For example while an indi-vidual may work through the following thought “I re-member finding the file on my laptop before when Iwas in the cafe drinking coffee with Pete when it wasraining about a month ago” and use some of this infor-mation to find the file, we are not aware of any current

― 76 ―


IR systems which allow a user to search for an itembased on a complex combination of features includingsuch details as the people in their vicinity or the pre-vailing weather conditions.The rich variety of context data sources that are po-

tentially available in PLs makes a compelling case forthe exploration of their exploitation to improving theaccuracy in information access applications. Beyondthis observation, coupled with rich context sources isthe additional information that can be obtained aboutitems by linking them to related content. In standardIR, the exploitation of the web’s linked structure hasproven to be a important component in high qualityweb page retrieval [30]. The link structure is used todetermine the quality and potential relevance of a pageby looking at the number and quality of pages linkingto it. It is interesting to consider whether exploitationof existing or inferred links within a PL could also beused to enhance the performance of information ac-cess applications. These links may be either content(e.g. linking based on the content similarity of items)or context (e.g. linking temporally based on user ac-cess patterns) based. To date only very limited workhas been reported within PLs, for example [36] whereresult sets were linked based on user access patternsand the resulting linked structure was used to rerankthe result set using PageRank type algorithms. Webelieve that there is significant scope to extend theseideas to newmethods for information access from PLs.Having considered some of the important distinc-

tive features of PLs we next turn our attention to con-sidering what user tasks they might actually be used tosupport.

3 Information Access Applications forPLs

We envisage PLs being used as part of a rangeof information access applications, some really beingstraightforward extensions of standard IR to a new in-formation source, others being modifications or en-hancements of existing user activities, and others be-ing potentially new tasks. Some specific possibili-ties making use of a number of underlying informa-tion access technologies include: standard interactivesearch; proactively reminding the user of related expe-rience from the past (e.g a previous visit to the sameplace), an advanced application of this sort might takethe form of a type of prosthetic extended memory; al-lowing users to reminisce about experiences (a sort ofpersonal browse or surfing service); social use (wherethe user might share their experiences with colleagues,friends and relatives) or, in a rather more complex sce-nario, multiple users might link sections of their PLs tocreate “community” memories; and their use in ther-apeutic treatment of memory impaired patients. An-other strand of applications, which we don’t explore

further here, is their potential application to basic psy-chological memory studies, where they could be usedto explore areas such as: success and failure of recallfrom memory, errors in memory recall, false memorysyndrome, and learning and rehearsal studies.Realising some of these possibilities is relatively

simple, and indeed work to develop several of themis well underway, while others are much more ambi-tious, requiring extended periods of research. How-ever, if they are to be explored in a scientific manner,they need to be scrutinized carefully in an attempt toproperly understand the basic scientific questions andissues raised, and to develop appropriate experimen-tal strategies to evaluate them. In order to do this weneed to consider the available sources of data, classifyit into well specified types depending on its features,and consider the means of capture. In terms of experi-mental design we need to be aware of the issues associ-ated with developing evaluation strategies for informa-tion access tasks with traditional information sources,but also be aware of the unique features of PLs. Theseinclude issues such as the fact that individuals will of-ten consider their PL data private to themselves. Thusthe data cannot be distributed to researchers (as is gen-erally the case for other experimental information ac-cess search data sets), and since much of the meaningand interpretation of content will also be personal, theindividual will need to be involved in developing thetest data for their PL and interpreting the results. Evenwhere the user can at least to some extent be sepa-rated from the these activities, many people will notbe willing to allow users to look at data from their PLeven for the purposes of tasks such as relevance as-sessment. The requirement that the “owner” of the PLmust be heavily involved in the collection of the data,the development of test data and the evaluation of ex-perimental results, will necessarily limit the numberand types of people who are able and willing to beinvolved in these activities. It is perhaps likely thatpeople willing to make this level of commitment willnot be properly representative of typical user groups,meaning that one has to design and interpret experi-ments with care. Also the extensive involvement ofthe individual may affect their behaviour at any stageof the experimental process, which may again call intoquestion the validity of experimental results. For ex-ample, their behaviour during data collection may notbe typical, and repeatedly gathering a range of exper-imental results may affect their knowledge of the col-lection restricting their ability to behave in a normalway when interacting with the data or biasing their in-terpretation of results.In addition to the personal use of PLs, alternative

scenarios are clearly possible, for example one usercould give another permission to access their PL, per-haps within a family or among relatives making theirPL available to their descendents. While the possibil-

― 77 ―


ity to explore the life history of a grandparent directlyfrom their PL offers fascinating prospects for the fu-ture, assuming grandparents are willing to be quiteso open about their experiences, at present such his-torical collections do not exist, and our experience isthat users are not willing to give this level of access totheir associates. At this point, third party use of PLshas largely been outside the scope of the experimentalstudies.At this point we turn our attention to consideration

of some specific tasks in a little more detail.

3.1 Ad Hoc Search

One class of PL application is ad hoc search. Vari-ous scenarios are possible here. Users may perform aknown-item search for a specific item or event whichthey partially remember, either to inform themselvesof some forgotten detail or to share it with others. Al-ternatively, they may have a less well focussed infor-mation need where they either feel, know or believethat they have experience of something in their past,but are unsure of the details. Or perhaps, the user maywish to browse their PL with respect to some facet,such as experiences where person X was present. Amore advanced version of this would have the sys-tem link together related materials, potentially dynam-ically forming further links in a query dependent man-ner, to provide users with suggested related memories.Some of these the user may remembered, but the sig-nificance of others may have been overlooked or theymay have been forgotten entirely.In our work we hypothesise that contextual infor-

mation related with people’s memory of the genera-tion of and previous access to items and events storedin their PL can be used to improve subsequent accessperformance. Traditionally IR systems allow users tosearch based on their memories of keywords containedwithin items. However, over time individuals memo-ries of these keywords can fade resulting in increaseddifficulty in locating the required information. In thecase of some data such as files, it is often easier toremember non-textual elements such as location [3].Non-textual memories of items may range from peo-ple present when accessing an item, to a phone con-versation that interrupted your email, writing, etc. Inour work on IR for PLs we are working to develop andtest methods which integrate automatically recordedand derived context data types into traditional IR algo-rithms for improved retrieval in the PL domain [18].Our results so far indicate that context can indeed bebetter recalled than content as the time from PL collec-tion increases [27]. As part of this work we are seekingto better understand the processes and circumstancesin which content and context are recalled accurately,inaccurately or forgotten, and to the extent to whichthese details can be generalised across users [12].

3.2 Narratives

In addition to traditional search activities, PLs of-fer new opportunities for other domains. One of par-ticular interest to us is that of digital narratives. Intheir simplest form a user may wish to re-experiencematerials associated with some activity from the past.A slightly more formal application is the use of per-sonal visual diaries captured using a passive camera,such as the Microsoft SenseCam [24], which regularlyrecord images from the wearer’s day enabling them tosubsequently replay them. Visual diaries are a uniqueaspect of PLs that capture the actions of an individ-ual from their perspective, thereby creating a personalarchive containing a visual stream of everyday life. Asan alternative to writing a diary, the visual diary orlifelog can be recorded from the start to the end of theday. Using a conventional diary, a person can easilyreview a day or any period of time quickly and effi-ciently, however with visual lifelogs, this process ismore challenging because although the visual lifelogis a richer medium than a conventional diary, organis-ing and finding the content poses more of a challenge.For example, with a visual lifelog captured using videocontent, at normal playback speed it will take a year torelive a year’s lifelog [25], clearly it is necessary todevelop organisation and search techniques that man-age personal information archives and these need tobe adequately evaluated. Replaying events from visualrecords of this form has found application in improv-ing short termmemory recall for memory impaired pa-tients in clinical studies [24].A more challenging and intriguing scenario is the

use of PLs to construct a digital narrative describingincident(s) from the user’s life in story-form. Sincethis is a topic generally poorly explored in informa-tion access, we outline its features in some detail inthis paper. Stories, and their retelling, are fundamentalto being human, in [31] Kearney declares that “sto-ries are what makes our lives worth living”. If weconsider that PLs contain a person’s life experiencescaptured through digital technologies, then the poten-tial for telling personal and intimate stories from suchcollections is enormous [11]. A digital narrative inits simplest form is a series of causally related mediaartefacts that are brought together in some presenta-tion mode to provide a meaningful and coherent dia-logue. Brookes more clearly defines a model of digitalnarratives built from three separate functional layers,namely: the structural, representational and presen-tational [8]. Structural considerations are concernedwith describing the narrative in simple abstract terms.The representation of the story seeks to capture the re-lationships between the various story elements in or-der to facilitate the reasoning required to build the endnarrative. Finally, the presentation of a digital narra-tive applies the structural and representational knowl-

― 78 ―


edge in order to choose and sequence story elementsthrough reasoning and with a sense of aesthetic style.One of the key issues in reviewing contents from a

user’s PL is the selection of material to be included inthe presentation. Even for a visual diary simple rapidplayback of all of passively captured images from dur-ing a single day rapidly becomes time consuming andtedious. This quickly becomes a much more signif-icant problem when attempting to construct rich nar-ratives taken from over an extended period of time.Overlapping with the issue of content selection, is thetopic of presentation. Providing an effective interfaceto the data should enable the user to more rapidly andeffectively access the presented information, and canpotentially enable more information to be included inthe output without loss of comprehensibility.In our current work we are exploring the use of PLs

for both search and narrative tasks. However, to under-stand our success and failure in this work, and to moveour research forward, we need to establish measuresby which our work can be assessed.

4 Evaluation Metrics

In order to evaluate information access applicationsfor PLs experimental collections are required whichare sufficiently large and diverse to represent the ex-pected features of real user PLs. As a consequence,evaluations are wholly constrained by the number ofparticipants who have collected data. Furthermore,there must be a sufficiently long period of collectionto allow for an adequate number of personally signifi-cant and causally related events to have been captured.Issues of this sort are of course common for existinginformation access experimental work. Developmentof test sets inevitably leads to collections which havesome limitations and are open to criticisms of the ex-tent to which they are truly representative of real tasks.For example experimental results of retrieval evalua-tion tasks using crawled web search test collectionsare often cited as being too small to reliably reflectperformance on the web itself which is several ordersof magnitude larger. We explore our current work onthe collection of large experimental PLs in section 6.Assuming that suitable collections are available,

performance in search tasks can bemeasured by a vari-ety of metrics taken from traditional IR evaluation ex-ercises, and personal digital narratives can take lessonsfrom existing work in narrative studies and summa-rization.For known-item ad hoc search with ranked retrieval

one can apply the metrics of mean rank or mean re-ciprocal rank as used in existing workshop evaluationtracks, for example [26]. Where there is more thanone potentially relevant item for the search query stan-dard metrics of precision and recall can be applied.These have been applied in some of the limited exist-

ing studies in information access for PLs. For exam-ple, the Connections [37] system used precision andrecall to measure performance [37], and also obtainedanecdotal evidence of user satisfaction. Another met-ric explored in existing work is the analysis of the timetaken by subjects to complete tasks using different ap-proaches [32].While quantitative methods have been used in some

studies, questionnaires, surveys, observations of sub-jects searching behaviour and diary studies (the resultsof which can be used to note trends and peoples’ be-haviour) to establish user’s searching, refinding andfile sharing processes were the most common evalu-ation methods used by participants at the Personal In-formation Management (PIM) 2008 workshop [2].The evaluation of narrative storytelling requires a

little more extensive analysis. At present, there is nogeneral consensus on how a narrative should be eval-uated or the criteria for their evaluation. Researchersworking in this area may fail to report their evalua-tion strategy with sufficient depth for it to be useful toothers or repeatable [4], opt for a highly qualitative orinformal approach to evaluation [23, 9] or omit evalu-ation (from their reports) altogether [33]. As outlinedpreviously there are four major functions of our inter-nal ’organic’ stories which should be applied to theirdigital counterparts. These can be broadly condensedto two types of tasks. The first is the social sharing ofstories which applies to both everyday and life stories.The second is personal review of stories for reminis-cence or reflection, which is predominantly a functionof our life story. As with IR strategies for PLs, theevaluation of reminiscence based tasks with personalnarratives is problematic.In order to practically and reliably investigate re-

flection such a task can only be evaluated with theparticipant who owns the data. The sharing of suchstories is not bound by this limitation. Even if only asmall number of participants collect PLs, if they arewilling to contribute their material to a research com-munity, shared narratives may be successfully evalu-ated by distributing them to a large number of externalparticipants. .One seemingly obvious approach to the quantitative

evaluation of digital narrative output within PL contentwould be to adopt the approaches taken with the sum-marisation, as for example adopted for the evaluationof unstructured video Rushes Content in [29]. Withinthis TRECVid summarisation evaluation, participat-ing sites seek to compress raw unedited video footage,with high numbers of retakes and large amount of con-sequent repetition, to within 2% of its original lengthusing a variety of techniques. The resulting video con-tent is then evaluated using a standard set of bench-mark measures including coverage and flow or usabil-ity of the summary. Coverage of the narrative of theoriginal content is a sensible metric to be included in

― 79 ―


narrative evaluation, however, it should be noted thatnarrative is not just summarisation. In fact there areclear distinctions between the goals of the two. A goodnarrative is to be emotive, expressive and most impor-tantly engaging, while a summary attempts to commu-nicate as much information in as compact a space aspossible. As such other measures must be consideredto temper the notion of coverage and more fully con-sider the goals of a narrative. We believe that there arefour major dimensions by which a narrative should beevaluated:

1. Form and content: The choices of what to showand how to show it are fundamental to a gooddigital narrative [38]. Within PL narratives mul-timodal content will typically be presented to anaudience. As such both the fluency of the mediachosen and the mode in which it is presented willhave significant impact on the communication ofthe meaning and themes of the story. We mustas such consider how to appropriately measure orevaluate and compare such factors.

2. Coherence: Coherence is integral to a good story:The quality of the end narrative is determined byhow well organised and well paced the content ispresented, and the rate at which the plot unfolds.Within interactive and digital narrative there is acareful balancing act to be maintained, betweenthe story’s coherence and the provision of agencyto the audience as to how the story unfolds [10].Coherence can potentially be measured throughthe level of control provided, the perception ofdistortions or ’jumps’ in the story, and/or attempt-ing to ascertain the overall flow of the story.

3. Expressiveness and affect: Our life stories are ex-periential and are often highly bound to our emo-tional (or affective) state at that time. As such, theresulting stories must sufficiently express theseaffective and experiential components. There aremany ways in which the affect may be repre-sented within narratives, one method is the useof colour to represent mood or emotional state[38]. The evaluation of expressiveness shouldnot focus on the mode of presentation however,but rather should consider how successfully thenarrative conveys and communicates these expe-riential aspects. Approaches to ascertaining suchmetrics might be to employ heuristics such asthose suggested in [14].

4. Engagement, Impact and feedback: Brooks as-serts that an audience should actively desire toknow the outcomes of a story, and with a betternarrative, the experience becomes much more en-gaging achieving just that [8]. This engagementis a vital component for any story, and should ide-ally be reflected in PL-based stories, and as such

we must carefully consider it within any evalu-ation. Although it is likely to be difficult to ac-curately and reliably elicit such metrics, we as-sert that evidence of personal engagement withthe story should be recorded and reported in somemanner be it qualitative or quantitative.

An alternative approach to evaluating narrativewould be to adopt the rules which apply to more tradi-tional or conversational discourse. For example, Grice[21] outlines a series of maxims under four categories:quantity, quality, relation and manner. These couldeasily be adapted to provide a suitable framework forthe evaluation of many of the properties of a good nar-rative, primarily its form, plot, and discourse.Brooks also notes that the better the representation,

the more choices in the narrative paths for the storyand this leads to a much better end narrative [8]. Assuch it may be necessary not only to evaluate the out-put and form of the end narrative, i.e. its presenta-tion, but also to examine in some way the represen-tation. Elson and McKeown [16] outline measuresfor the quality of a narrative’s representation as: ex-pressiveness (the range of stories it can encode), ro-bustness (how well it handles partial or abstractly toldstories), formality (how well its symbols lend them-selves to formal analysis or automatic processing) andusability (how intuitive the representation is for col-lecting encodings from those not versed in narrativetheory). The evaluation of the representational formof the narrative under such measures of quality couldbe included, however it is likely to be dependent on thetask under evaluation and the goal of the investigation.

5 Previous Work in Information Accessfor PLs

Although the topic of collection materials for PLsand their utilization is still very much in a develop-ment stage, a number of relevant studies have alreadybeen reported in each of the areas discussed so far inthis paper. This section summarises some of the mostnotable recent work in these areas.

5.1 Retrieval from PLs

With regard to retrieval tasks the following studiesdescribe some of the most relevant work reported inthis area to date. In [36] [37] PL retrieval algorithmsare developed and evaluated on personal archives for6 subjects consisting of computer activity over a pe-riod of 6 months. The subjects submitted 3-5 queries,which they appeared to freely generate from theirmemory. These were multi-item standard ad hoc typequeries, e.g. “locate all items associated with writ-ing of EVIA paper.” To create oracle results for thesequeries the results of this query across different search

― 80 ―


engines were pooled together, and the subject rated therelevance of the pooled result set. Other researchers,while limiting their evaluation to emails from personalarchives, moved beyond this limited free recall ap-proach to generate large numbers of tasks for known-item tasks. In [15], 30 queries with target email itemswhich had been sent to a large number of people in acompany were manually created. These queries werethen entered by test subjects into the Stuff I’ve Seen(SIS) interface, resulting in the retrieval of variousitems from their personal collections including the tar-get email. Subjects were then required to locate thetarget email, thus allowing for testing of various ver-sions of the SIS interface. While this technique proveduseful for testing features of PL interfaces, it wouldprobably not be appropriate for generating tasks forall types of PL system experiments. For example, forevaluating PL retrieval algorithms where a user is re-quired to use recalled content and context to form aquery, i.e. if we generated tasks by giving a subjecta task description of emails that had been sent to agroup, the subject may have no recollection of theseemails, and therefore could not recall content or con-text data with which to form a query.In Elsweiler’s work [17] a more personal approach

to task generation is adopted. This work presented aframework for a task based approach to PL user eval-uation. Specifically, over a three week period sub-jects recorded email tasks (task = email viewed and thepurpose for which it was viewed, e.g. relocate emailwhich contains Joe Blogg’s phone number). They alsoallowed for the generation of additional tasks in a man-ner similar to Dumais’s work [15] described in theprevious paragraph. However they did not providethe query terms, instead providing a task descriptionwhich simulates a ’real world’ task the subject may en-gage in. Tasks were then categorized into three distincttypes: tasks requiring a specific piece of informationfromwithin a computer item; tasks requiring a specificcomputer item; and tasks requiring information frommultiple computer items. Tasks recorded by a subjectwere then presented to the subject for them to retrievethe task using a provided interface. This approach al-lowed for comparison of retrieval performance acrossdifferent task types, and importantly the evaluation ofa PL retrieval system, in a structured manner usingthe PL owners themselves. This approach could po-tentially be extended and used as a first step towardsa much needed standard for evaluation in the PL do-main. Elsweiler et al. examined this task generationapproach only on emails and web pages, and thus itsportability to other item types is not guaranteed.By linking PL items, we believe textual descriptions

from linked to items may be obtained which couldprove beneficial in search, both for items of a non-textual nature such as photos or recordings of music,and items for which the user cannot recall the tex-

tual contents in sufficient detail. A number of exist-ing studies have begun to explore the use of linkingand annotating items. For example the PHLAT sys-tem allows individuals to add tags to their computeritems [13]. However, in general the retrieval meth-ods used for this data at present use existing backendalgorithms and appear to operate on a simple booleanpremise. For example [15] [20] where the queried con-text data is either present or not present in the databaseitems. We believe that to truly harness the power ofitem linking and exploitation of related context data,items should be annotated with many rich sources ofcontext data to support linking and search, and thatnovel retrieval algorithms tuned to PL tasks need to bedeveloped which exploit these annotations.

5.1.1 Pilot Studies in PL Retrieval at DCU

We are interested in investigating individuals’ memo-ries of past interaction with items in their PL, the typesof context data they recall, how recalled context datavaries across different item types, and how individu-als’ memories change over time. In particular, it is ourhypothesis that integrating remembered and partiallyremembered context information into a traditional IRalgorithm will prove beneficial for PL retrieval.In order to begin our investigation of these issues,

we conducted a pilot study to test the types of contextdata recalled by an individual and to examine the util-ity of remembered context data in retrieval. A smallPL was gathered for one individual over a 6 weekperiod from the middle of July 2007 to the end ofAugust 2007. To gather the PL, a logging applica-tion monitored activities carried out on the individ-ual’s computer during the period and additionally reg-ularly prompted the user to manually log the locationat which these happened. The retrieval experiment wasa known-item search task where the user searched fora single partially remembered item from within thedata collection. To generate the query tasks for thepilot study, following the 6 week collection period,the test participant identified a number of key eventsthat occurred during the time span of the data collec-tion. These events included: a friend’s birthday, sev-eral meetings in the office, and dinner at a restaurant.Following this, the participant was required to recallthe activities performed on their computer around orclose to these key events. Each of the 27 test casesgenerated in this manner then consisted of a remem-bered computer file (e.g. email, word document, IM)and remembered context data associated with the file.If PLs are to be recorded and accessed over an ex-

tended period of time, it is important that users are ableto reliably retrieve content recorded in the distant past.It is clear that a user is likely to remember a significantamount of detail soon after an event occurred, howeverwith time memory fades and it is anticipated that less

― 81 ―


will be remembered a significant amount of time afteran event occurred. A further free recall study on thesubject 6 months after the initial test case generationprocess was thus conducted to explore the effect on re-membered content and context after this time interval.This study provided us with remembered content andcontext from the original 27 test cases at the 6 monthinterval.The Mean Rank and the Mean Reciprocal Rank

[26] of the target documents were calculated acrossthe two query sets consisting of remembered contentwith various types of remembered context (e.g. queryset 3 = content + location context + date context). Theresults of this study suggest that the types of contextdata an individual is most likely to recall in the long-term and that this context data is beneficial in PL re-trieval. In particular, the results indicate that combina-tions of content and context can improve retrieval ac-curacy from an PL; and that context appears to becomea more important factor over time as the subject’s re-call of contents declines. Full details are described in[18] [27]

5.2 Visual Diaries

Before examining relevant existing studies in nar-ratives, we examine the simpler case of visual diarieswhich have so far been explored more extensively.Several studies have been conducted in recent yearswhich passively capture a visual record of the wearer’sdaily activities. In this overview we focus on work us-ing the SenseCam developed byMicrosoft Research inCambridge, UK [24]. The SenseCam is a small wear-able device that passively captures a person’s day-to-day activities as a series of photographs. It is typi-cally worn around the neck, and so is oriented towardsthe majority of activities which the user is engagedin. Anything in the view of the wearer can be cap-tured. The device incorporates on-board sensors to de-tect changes in light levels, motion and ambient tem-perature and then use these to determine when it is ap-propriate to take a photo. At a minimum the SenseCamwill automatically take a new image approximately ev-ery 30 seconds, but sudden changes in the environmentof the wearer, detected by the onboard sensors can trig-ger more frequent photo capture. The benefits of thisinclude: the ability for a user to record events withouthaving to sacrifice their participation, aiding memoryand recall; and providing insight into a person’s lifeand activities [24]. Wearing a SenseCam, the wearercan very quickly build a large archive of photos, ora visual lifelog. Within just one week over 20,000images may be captured when wearing a SenseCam.When worn over a year the lifelog photoset can growto over one million images.At DCU we have been gathering archives of Sense-

Cam images in various collection efforts. Multi-month

visual lifelogs have been gathered by 5 researcherswith another researcher gathering an archive of overtwo years data. This large collection has over 2.5 mil-lions images. Based on our experiences of gatheringthis archive, we have found that the photos capturedduring this extended period of continuous lifeloggingare vastly different in content to what one would ex-pect to find with conventional digital photos. Wehave found that images often do not have salient ob-jects, many of them are either low quality or useless(40%), and the types of scenes/objects captured dif-fer greatly from conventional photo collections. Thisis because the SenseCam is capturing aspects of awearer’s life that are not normally photographed, forexample 20% of the photos we observed containedthe wearer’s hands, 20% also show the LCD screenof the wearer’s office computer screen or laptop, whileonly 15% of these are taken inside of the office of thewearer. The personal nature of the lifelog photo col-lections, when taken from the viewpoint of the wearer,provide intimate details of the wearer’s work and so-cial activities. Other observations support this point,such as the fact that 6% of the photos show people’sfaces (friends, work colleagues and family members),4.5% show the details of meetings and 2% of the pho-tos show the wearer reading.

Evaluation of SenseCam derived visual diaries car-ried out so far at DCU has focused on examiningthe nature of the visual lifelogs, to better understandwhat information is contained in them. In addition tothis, we have carried out an evaluation of approachesto managing visual PLs, such as how best to organ-ise them into a series of events, how best to chose akeyframe to identify an event, and how visual featuredetection tools (such as faces, horizon, cup) performon the SenseCam photos.

Research on the use of visual diaries to improvememory in clincial patients with cognitive dysfunc-tion has used recall of autobiographical events as anevaluation criteria. This work compared recall usingthe SenseCam to recall using both a written diary andno memory aid [24]. Other experimentation has beenongoing to examine how effective SenseCams are asa memory aid, for example one study on user recallafter wearing SenseCams investigated differentiatingbetween different types of recall; did the user ’remem-ber’ the event, or did the user simply ’know’ that theevent existed? These experiments typically take placeover a short period of time, for example a week [34].A key question for our work is to explore the effective-ness of visual diaries as a memory aid both in itself andin terms of their impact on memory itself in terms ofrehearsed recall over the longer term. Moving beyondsimple diary PL collections, digital narratives providea much richer framework for presentation of contentfrom a PL.

― 82 ―


5.3 Narratives

From the components of a PL, two major types ofstories may be told. The first being a conversational,or ’everyday’ story. Within our social interactions ona day-to-day basis we frequently share stories with ourcolleagues, friends and family. Often they are anecdo-tal and are used to recall a moment of humour or a re-cent event. As such their function is primarily routedin socialising and sharing of experience(s). The sec-ond type is the ’life story’. Life stories are internalever evolving narratives which integrate significant oremotional events and update retrospectively to accountfor changes in our perception and remembrance of pastevents, values and beliefs [28].

Unlike everyday stories, the ’life story’ provides uswith the opportunity to examine a set of causally re-lated events, framed from the perspective of our cur-rent self and within the context of our larger life expe-riences, motivations, goals and ambitions. It providesthree main utilities to us: first it allows purposeful re-flection on one’s life events and their development inorder to affect or determine future actions (directive);second it enables us to share aspects of our lives withothers, often as a means to develop social bonds, in-timacy and/or empathy (social); or finally, it affordsa review of life experiences to search for meaning inthem (self) [7].

Due to the recent interest in PLs provided by toolssuch as the SenseCam, the composition of ’every-day’ stories from such archives has seen some lim-ited exploration in a number of small scale studies. In[19] images from the SenseCam along with associatedGPS location information was presented as a meansto recount a ’trip-based’ experience as a lightweightstory in the form of an animated slideshow composedof SenseCam images. Harper et al. [23] conducteda study with six participants into user-created digi-tal narratives composed from SenseCam captured im-ages. The study is very much focused on participants’perceptions of a manually created end narrative, andyields some interesting outcomes particularly high-lighting the usefulness of such images in reflection andreminiscing over life experiences. Of most relevance,but perhaps not wholly within the domain of PLs, isthe work of Appan et al. [4]. This investigated thecomposition of digital narratives for everyday expe-riences using media such as photos, gathered duringthe user’s day-to-day activities. While there has beensome exploration of conversational stories, we are notaware of any work exploring digital equivalents to thelife story. Given the volume, complexity and rich-ness contained within PLs there is great potential bothfor the elicitation of such stories and the evaluation ofthem.

6 PL Activities at DCU

In order to move on from our initial work on PLaccess applications, our group at DCU are currentlyengaged in collection of large experimental PLs. 4subjects are currently actively engaged in collectingpersonal data to form a PL. Ideally these collectionswould be developed over very extended periods, pos-sibly running into many years. For practical reasons,i.e. we cannot wait years to continue our research, weare currently limiting the next phase of the collectionperiod to one year. While not perhaps ideal, we shouldemphasise that the collections resulting from our cur-rent work will be much larger and more diverse thanother collections made by ourselves or others to date.The following timestamped data is being collected

for each of our subjects over the one year collectionperiod:

• Home and Work Computer Activity: every item(email, word document, web page, etc) accessedby the user, with time of access, contents of itemand path to item information is being recorded.To do this a program which combines the outputof Digital Memories software [1] and Slife soft-ware [35] running on the individual’s computerswas developed.

• Mobile phone activity in the form of call logs andSMSs are recorded. An application was readilyavailable to capture mobile phone call logs, andan application was developed to capture sent andreceived SMSs.

• GPS data, from which location names, light sta-tus and weather conditions can be derived, andBluetooth data, from which people present can bedetected, are also captured on the mobile phone,using software available within our group.

• SenseCam images and digital photographs arealso stored.

Physiological and heart rate data is also being cap-tured to monitor physiological condition and inferemotional state. It was decided that it would not bepractical to capture this biometric data over the courseof the entire year, since the sensors are quite physicallyintrusive to daily life, so this data is being collected fora one month period by the 4 subjects.The 4 PLs generated over the course of the year will

allow us carry out detailed investigationswith what ap-proach ’real’ PLs. However, it is acknowledged thatsolid experimental results necessitate a larger numberof test subjects. Thus to provide further support for ourexperimental results, more subjects will be obtainedduring the course of the year for smaller time framesto capture their own PLs.

― 83 ―


In order to collect their PL, each of the participantshas been provided the following apparatus.

Hardware

• Nokia N95 mobile phone: The N95 is a sophis-ticated mobile device which runs the Series 60platform enabling mobile applications to be de-veloped and deployed on it. It also includes abuilt-in 5-megapixel camera and large storage ca-pabilities (up to 8GBs) making it ideal for use inthe long-term collection of data.

• Microsoft Research SenseCam: Although notcommercially available, we are fortunate to havebeen provided with a number of these devices byMicrosoft Research.

• Garmin GPS Loggers: These small portable de-vices automatically regularly log the participant’slocation. Due to the limitations of GPS technol-ogy, typically these are only used when the par-ticipant is outdoors and on the move.

Software

• Desktop Activity Recording Software: Users areasked to run desktop activity logging softwareconstantly on both their home and work com-puters. For those running Windows-based ma-chines, they are provided with both the MicrosoftMyLifeBits [1] and SLife [35] software, whilethose running Macintosh OSX are only requiredto run SLife. These software suites record all ap-plication use and content access on a desktop orlaptop computer.

• Mobile Context Recording Software: The Cam-paignr software, provided to us by UCLA (USA),has been deployed on the N95 devices. Par-ticipants run this software constantly to capturecontextual factors such as their location throughWireless, GSM and GPS sniffing and the personsin proximity to them through Bluetooth devicesniffing.

• Mobile Activity Recording Software: A propri-etary piece of software enables the logging ofmissed calls, made and received calls and sentand received text (SMS) messages on the N95.

These various data sources will be integrated intoa PL for each user for investigation of PL informa-tion access applications. Part of the user studies willalso focus on determining the accuracy of our con-text derivation methods by comparing derived contextdata to that experienced by test subjects. Participantswill also complete a number of questionnaires over thecourse of the PL build up year to investigate variousfactors of PLs, such as items the participants consider

important within their collection, items that they be-lieve are likely to want to retrieve in the future, andthe types of content and context data subjects recallabout items of different type and importance level.

6.1 Structuring the Data

An important aspect of PL collections is how tostructure the data for effective information access. Forexample, the many thousands of SenseCam imagestaken as ongoing sequences should be segmented intomeaningful units for browsing and retrieval purposes.In the domain of digital video search, there is a wellknown technique called shot boundary segmentationthat identifies the camera recording boundaries thatcompose a piece of edited video. Often the video clipsbetween the shot boundaries (video shots) are used asthe unit of retrieval. Our group have carried out exten-sive evaluations of event segmentation for SenseCamcollections to identify the best approaches to integrat-ing context and content data to achieve high precisionevent segmentation [22]. Following on from the eventsegmentation process we have been exploring the judi-cious selection of a key-image or other surrogate datato represent an event for the end user [6]. In our workwe have focused on evaluating techniques to representeach event by a single image or keyframe. In addi-tion to event segmentation and representation activi-ties, we are exploring how successfully conventionaldigital image concept detection techniques can be ap-plied to visual lifelogs. Finally we are also planningto evaluate a procedure for automatic annotation ofpeople present at the time of capturing data for a PL,by evaluating how to effectively identify friends, fam-ily and work colleagues by means of Bluetooth devicelogging [3].Beyond the structuring of the SenseCam data, we

are beginning to investigate how the other sources ofdata within a PL can be linked in support of informa-tion access applications.

6.2 Search

Using our new PL collections we plan to conductextended versions of the experiments reported in [18][27]. We also plan to use other task generation pro-cesses described in [17], and the use of informationcontained in digital schedulers to help in the task gen-eration process.In addition to this, we are also seeking to better

understand circumstances in which users will remem-ber different features associated with personal infor-mation and the information itself, and how they willsubsequently describe it. As part of this investigationwe are looking at models of the underlying cognitiveprocesses in human information storage and retrieval[12]. In order to do this we are conducting a range

― 84 ―


of user experiments via questionnaires and interviewswith our collection participants and others. Althoughof course as mentioned earlier, care needs to be ex-ercised in doing this not to modify the typicality ofthe collections by requiring users to rehearse recall ofitems from their PL, with the consequence of affectingthe reliability of results of subsequent retrieval experi-ments.

6.3 Digital Narratives

With regard to digital narratives, we are just be-ginning to explore the possibilities within such collec-tions. We hypothesize that the richness of content andcontext captured in a PL will enable the approximationof our internal ’organic’ conversational and life sto-ries with digital equivalents. Additionally, we believethat such digital equivalents can also serve to support,prompt and assist the the functions of our internal sto-ries as outlined previously. The investigation will becarried out as follows.First a preliminary study will seek to ascertain how

personal life stories should be represented through dig-ital narrative. To determine this we have planned aqualitative probe using a large number of participants.Participants will be asked to construct a number ofsimple conversational stories and one more extensivelife story using media artifacts already available tothem, including photos and videos or text from emailsfor example, or relevant content that can be easilyfound online on sites such as Flickr. The stories cre-ated will then be probed during one-on-one follow-upinterviews, the results of which will be coded to deter-mine if there are commonalities in the approaches torepresenting the story which could be developed into amodel. The conversational stories created will also beused to probe the fluency of various media types andtheir affect on the communication of the story. Thesewill then be shown to new participants who will beasked to quantitatively analyse the communication ofthe story. To measure this we will develop an evalu-ation framework based on the dimensions mentionedin the previous section. This framework will proba-bly take the form of a questionnaire that will probe theparticipants’ reactions to and perceptions of the storypresented to them.Utilizing the knowledge gained from these stud-

ies, we will next build a semi-automatic narrative gen-eration engine as described in [11]. The structureof the narratives will be event-oriented as suggestedby the related work in [4]. In the authoring processof this system, users will browse the archive in anevent-oriented manner. They will locate and select theepisodes which are relevant to the story, marking themfor inclusion. After which, they will be asked to makebasic plot and aesthetic choices to determine the pre-sentation of the narrative. Following this the system

will enter a generation process in which the selectedcontent and author choices are constructed into a co-herent story.With the system built, the final activity in our study

will be to validate the output of the system. Exampleoutput will be used to evaluate the hypothesis that digi-tal life stories can be used to enable personal reflectionand sharing of meaningful life experiences.

7 Conclusions

In this paper we have introduced the emerging areaof information access for personal lifelogs. We haveoutlined the differences between PLs and more con-ventional existing document collections and envisagedinformation access tasks for them. We have also sum-marised work to date by ourselves and others explor-ing and evaluating information access for PLs. In par-ticular we highlighted their use in personal search,browsing and narrative. The paper concludes with adescription of our current work in collection of largePL test set for use in our ongoing experimental work.

Acknowledgments

This work is funded by grant CMS023 under theScience Foundation Ireland Research Frontiers Pro-gramme 2006 and a postgraduate studentship awardfrom the Irish Research Council for Science, Engineer-ing and Technology, and is supported byMicrosoft Re-search under grant 2007-056.

References

[1] MSR MyLifeBits Project - http://research.microsoft.com/barc/mediapresence/MyLifeBits.aspx.

[2] PIM 2008: Proceedings of Personal Information Man-agement, Workshop at CHI 2008, Florence, Italy,2008.

[3] K. Aizawa, D. Tancharoen, S. Kawasaki, and T. Ya-masaki. Efficient Retrieval of Life Log Based on Con-text and Content. In The 1st ACM Workshop on Cap-ture, Archival and Retrieval of Personal Experiences(CARPE 2004), pages 22–31. New York, New York,USA, October 2004.

[4] P. Appan, H. Sundaram, and D. Birchfield. Communi-cating Everyday Experiences. In SRMC ’04: Proceed-ings of the 1st ACM workshop on Story representation,mechanism and context, pages 17–24. ACM, 2004.

[5] G. Bell and J. Gemmell. A Digital Life. ScientificAmerican, March 2007.

[6] M. Blighe, A. Doherty, A. F. Smeaton, and N. E.O’Connor. Keyframe Detection in Visual Lifelogs. InETRA 2008 - 1st International Conference on Perva-sive Technologies Related to Assistive Environments,2008.

[7] S. Bluck and T. Habermas. The Life Story Schema.Motivation and Emotion, 24(2):121–147, 2000.

― 85 ―


[8] K. M. Brooks. Do story agents use rocking chairs?The theory and implementation of one model for com-putational narrative. In MULTIMEDIA ’96: Proceed-ings of the fourth ACM international conference onMultimedia, pages 317–328, New York, NY, USA,1996. ACM.

[9] K. M. Brooks. Metaliner cinematic narrative : theory,process, and tool. In Thesis (Ph.D.). MIT, 1999.

[10] A. Bruckman. The Combinatorics of Storytelling:Mystery Train Interactive. Unpublished paper, 1990.

[11] D. Byrne and G. J. F. Jones. Towards ComputationalAutobiographical Narratives through Human DigitalMemories. In SRMC 2008 - 2nd ACM Workshop onStory Representation, Mechanism and Context, 2008.

[12] Y. Chen and G. J. F. Jones. Integrating Memory Con-text into Personal Information Re-finding. In SecondSymposium on Future Directions in Information Ac-cess (FDIA 2008), 2008.

[13] E. Cutrell, D. C. Robbins, S. T. Dumais, and R. Sarin.Fast, Flexible Filtering with Phlat - Personal Searchand Organization Made Easy. In CHI 2006: Con-ference companion on Human factors in computingsystems, pages 261–270, New York, NY, USA, April2006. Montreal, Quebec, Canada, ACM Press.

[14] E. de Lera and M. Garreta-Domingo. Ten EmotionHeuristics: Guidelines for Assessing the User’s Af-fective Dimension Easily and Cost-Effectively. InProceedings of the HCI07 Conference on People andComputers XXI, page 40, 2007.

[15] S. Dumais, E. Cutrell, J. Cadiz, G. Jancke, R. Sarin,and D. C. Robbins. Stuff I’ve Seen: a system forpersonal information retrieval and re-use. In SIGIR’03: The 26th annual international ACM SIGIR con-ference on Research and development in informaionretrieval, pages 72–79, New York, NY, USA, July2003. Toronto, Canada, ACM Press.

[16] D. K. Elson and K. R. McKeown. A Platform for Sym-bolically Encoding Human Narratives. In Proceedingsof the AAAI Fall Symposium on Intelligent NarrativeTechnologies, pages 29–36. AAAI Press, 2007.

[17] D. Elsweiler, I. Ruthven, and C. Jones. TowardsMemory Supporting Personal Information Manage-ment Tools. Journal of the American Society for In-formation Science and Technology, 2007.

[18] M. Fuller, L. Kelly, and G. J. F. Jones. Applying Con-textual Memory Cues for Retrieval from Personal In-formation Archives. In PIM 2008: Proceedings ofPersonal Information Management, Workshop at CHI2008, 2008.

[19] J. Gemmell, A. Aris, and R. Lueder. Telling Storieswith MyLifeBits. Multimedia and Expo, 2005. ICME2005. IEEE International Conference on, pages 1536–1539, 2005.

[20] J. Gemmell, G. Bell, and R. Lueder. MyLifeBits:A Personal Database for Everything. Communica-tions of the ACM. Personal Information Management,49(1):88–95, 2006.

[21] H. P. Grice. Logic and Conversation. In P. Coleand J. L. Morgan, editors, Syntax and Semantics: Vol.3: Speech Acts, pages 41–58. Academic Press, SanDiego, CA, 1975.

[22] C. Gurrin, A. F. Smeaton, D. Byrne, N. O’Hare,G. J. F. Jones, and N. E. O’Connor. An Examination

of a Large Visual Lifelog. In AIRS 2008 - Asia Infor-mation Retrieval Symposium, 2008.

[23] R. Harper, D. Randall, N. Smyth, C. Evans, L. Heledd,and R. Moore. Thanks for the Memory. In Proceed-ings of HCI 2007, The 21st British HCI Group AnnualConference, page 10. BCS, 2007.

[24] S. Hodges, L. Williams, E. Berry, S. Izadi, J. Srini-vasan, A. Butler, G. Smyth, N. Kapur, and K. Wood.SenseCam: A Retrospective Memory Aid. In Ubi-Comp 8th International Conference on UbiquitousComputing, 2006.

[25] T. Hori and K. Aizawa. Context-based Video RetrievalSystem for the Life-Log Applications. In The 5th ACMSIGMM International Workshop on Multimedia Infor-mation Retrieval, MIR 2003, pages 31–38. Berkeley,California, USA, 2003.

[26] P. B. Kantor and E. M. Voorhees. The TREC-5Confusion Track: Comparing Retrieval Methods forScanned Text. Inf. Retr. (IR), 2(2/3):165–176, 2000.

[27] L. Kelly, Y. Chen, M. Fuller, and G. J. F. Jones. AStudy of Remembered Context for Information Accessfrom Personal Digital Archives. In IIiX 2008 - 2ndInternational Symposium on Information Interactionin Context, 2008.

[28] D. P. McAdams. The Psychology of Life Stories. Re-view of General Psychology, 5(2):100–122, 2001.

[29] P. Over, A. F. Smeaton, and P. Kelly. The TRECVid2007 BBC Rushes Summarization Evaluation Pilot. InTVS ’07: Proceedings of the international workshopon TRECVID video summarization, pages 1–15, NewYork, NY, USA, 2007. ACM Press.

[30] L. Page, S. Brin, R. Motwani, and T. Winograd. ThePageRank Citation Ranking: Bringing Order to theWeb. Technical report, January 1998.

[31] R. Kearney. On Stories. Routledge, London/NewYork, 2002.

[32] M. Ringel, E. Cutrell, S. Dumais, and E. Horvitz.Milestones in Time: The Value of Landmarks in Re-trieving Information from Personal Stores. In Interact2003, 2003.

[33] W. Sack. Stories and Social Networks. In AAAI 1999Fall Symposium on Narrative Intelligence, MenloPark, California, 1999. AAAI Press.

[34] A. Sellem, A. Fogg, S. Hodges, and K. Wood. Do life-logging technologies support memory for the past? anexperimental study using sensecam. In Conference onHuman Factors in Computing Systems, CHI ’07, 2007.

[35] Slife Labs - http://www.slifelabs.com.[36] C. A. N. Soules. Using Context to Assist in Per-

sonal File Retrieval. PhD thesis, School of ComputerScience, Carnegie Mellon University, Pittsburgh, PA,USA, 2006.

[37] C. A. N. Soules and G. R. Ganger. Connections: UsingContext to Enhance File Search. In 20th ACM Sym-posium on Operating Systems Principles (SOSP’05),pages 119–132. Brighton, United Kingdom, 2005.

[38] J. Tanenbaum and A. Tomizu. Affective InteractionDesign and Narrative Presentation. In AAAI 2007Fall Symposium on Intelligent Narrative Technologies.AAAI Press, 2007.

― 86 ―

information access tasks and evaluation for personal lifelogs

Documents