Presentation on theme: "Quality assessment of a current awareness system Thomas Krichel LIU & HГУ 2007–1023."— Presentation transcript:
Quality assessment of a current awareness system Thomas Krichel LIU & HГУ 2007–1023
acknowledgments Thanks to organizers. I am grateful for comment by –Bernado Bátiz-Lazo –Joanna P. Davies –Marco Novarese –Christian Zimmermann I thank everybody involved in RePEc and NEP, as well as JISC.
current awareness Current Awareness aka Selective Dissemination of Information is a simple idea: a user is informed about new documents in her area of interest. Current awareness generate a double classification in –subject... –time... matter.
why bother? It is niche activity that has been neglected by the search engines. I have registered with Google and Amazon. They give me tips but these are generally poor. We can not trust computers to do it. –Neither on subject matter –Nor on time
types of current awareness personal (amazon) vs collective (Google news) machine generated vs human generated? Actually I claim the only human-generated current awareness service for academic documents is NEP.
computer-based + keywords In computer generated current awareness one can filter for keywords. In academic digital libraries, since the papers describe research results, they contain all ideas that have not been previously seen. Therefore getting the keywords right is impossible.
computer-based + categories It is possible to classify documents based on categories say football vs tennis. It works fine when the vocabulary used in different categories is quite different. For some academic areas the differences are just too subtle.
computers and time problem In a digital library the date of a document can mean anything. The metadata may be dated in some implicit form. –Recently arrived records can be calculated. –But record handles may be unstable. –Recently arrived records do not automatically mean new documents.
we need humans Catalogers are expensive. We need volunteers to do the work. Junior researchers have good incentives. They –need to be aware of latest literature; –are absent from the informal circulation channels of top level academics; –need to get their name around among researchers in the field..
introducing NEP There is only one large freely-available human-based current awareness service. It is NEP: New Economics Papers at http://nep.repec.org Remainder of this talk is about NEP.
NEP: New Economics Papers NEP is a current awareness system for the working paper in the RePEc digital library about economics. Published articles are excluded because they are way too old.
NEP service model There is a basic model behind this service we could call the "NEP Service Model". –two stages... –flat report space...
general two stage setup First stage: A general editor compiles a list of all new papers. This forms an issues of the allport. Second stage: A group of subject editors filter the new allport issue into subject reports. Each editor does this independently for the subject reports she looks after.
a flat space There is a series of reports. Each reports has a number of issues over time. There is an allport a report that contains all papers that are new in the period covered by the issue.
first stage in NEP General editor compiles a list of recent additions to the RePEc working papers data. –Computer generated –Journal articles are excluded –Examined by the General Editor (GE, a person) This list forms an issue of nep-all.
first stage in NEP nep-all contains all new papers Circulated to –nep-all subscribers –Editors of subject-reports
second stage Each editor creates, independently, a subject report for her subject. She does this by removing from nep-all. A subject report issues (sri) is the result of this process. There have been over 47,000 sris issued through the lifetime of NEP to date.
history There are basically two phases in NEP the pre-ernad 1998 to 2004 and the post ernad phase. I will deal with pre-ernad history here. Some research on NEP has been conducted in the pre-ernad phase. This has informed the work that went into ernad.
early history System was conceived by Thomas Krichel Name NEP by Sune Karlsson Implemented by José Manuel Barrueco Cruz. Started to run in May 1998.
starting setup First the system was all email based. The nep-all was composed as an email. It was sent to editors as an email. Editors used whatever tool they used to compose the email.
web interface John S. Irons issued the first web interface for report composition on 2000-02-01. This would just compose the report. Editors would still cut and paste the results of the form into email clients.
historic mail support First mail support was given by mailbase.ac.uk. When this was closed in 2000-11, NEP moved to jiscmail.ac.uk. Since the mailing list service was only supposed to be for UK academic community, it was deemed not sustainable. Thomas Krichel started hosting lists on 2002-11-16. It is a nightmare.
Aeroflot document The Aeroflot document was a thinking piece that Thomas Krichel wrote as early as 2001. http://openlib.org/home/krichel/work/aeroflot.html This paper already sets out ideas for what would be ernad. At that time the Siberian RePEc team promised help with building such a system.
discover disaster In 2002-2003 Jeremiah Cochise Trinidad Christensen and Thomas Krichel were the first people to try to get a systematic picture of how NEP works. They discover that this is exceedingly difficult.
mail log parsing Logs were not moved from to Maibase to JISCMail. Mailbase removed the logs in 2002-11. Thomas Krichel got them just before they were destroyed. The mail logs were the only source for historic NEP information.
parsing targets handles: severely compromised by cut-n- paste operations, editor locales, etc. date of issue: editors were free to set dates, nep-all dates may not be preserved time of issue: an email is almost impossible to time.
state of pre-ernad data After a regular expressions orgy, we can get some approximate idea about the handles that were used. Thus the thematic component is roughly intact. We have a problem with a bug in the discovery program that made many papers appear several times in nep-all. This makes it difficult to associate subject and allport issues.
state of pre-ernad data Timing of emails is extremely difficult, even with full headers. The logs of the Mailbase system only have times for when the email client said it sent the mail. This is the local editor's PC time, can be years out of whack! We still have some data for research...
research conducted on NEP Most of the research conducted on ernad has been done in the pre-ernad phase. The difficulties of some of this work has informed the construction of ernad.
Chu and Krichel (2003) Heting Chu & Thomas Krichel (2003) NEP: Current awareness service of the RePEc digital library. http://www.dlib.org/dlib/december03/chu/12 chu.html vaguely talks about NEP. Notes that there is a problem of timeliness in the subject report issue, despite the very shaky data.
Barrueco Cruz et al. (2003) Jose Manuel Barrueco Cruz, Thomas Krichel and Jeremiah Cochise Trinidad- Chrisitensen Organizing Current Awareness in a Large Digital Library http://openlib.org/home/krichel/papers/espo o.pdf have two themes –overlap between reports... –coverage ratio... as well as history and suggestions.
overlap Barrueco Cruz et al (2003) argue that overlap occurs not when two papers are appearing in the two reports, but when the two reports are read by the same readers. They have data on pairwise overlap between reports, based on crude membership data.
overlap puzzle Here is a puzzle to think about –If a person will be interested in two subject areas because they are close, she will subscribe to both reports. –But since they are thematically close, she will sometimes receive the same papers twice. With mail technology and asyn-chronous issue generation, this appears difficult to solve.
coverage ratio We call the coverage ratio the number of papers in nep-all that have been announced in at least one subject report. We can define this ratio –for each nep-all issue –for a subset of nep-all issues –for NEP as a whole
coverage ratio theory & evidence Over time more and more NEP reports have been added. As this happens, we expect the coverage ratio to increase. However, the evidence, from research by Barrueco Cruz, Krichel and Trinidad is –The coverage ratio of different nep-all issues varies a great deal. –Overall, it remains at around 70%. We need some theory as to why.
Krichel & Bakkalbaşı (2005) Thomas Krichel and Nisa Bakkalbaşı Developing a predicitve model of editor selectivity in a current awareness service of a large digital library. http://openlib.org/home/krichel/papers/bosto n.pdf
coverage ratio theories Krichel & Bakkalbaşı (2005) build two theories of the observations of Barrueco Cruz at al. (2003) They are –Target-size theory –Quality theory descriptive quality substantive quality
theory 1: target size theory When editors compose a report issue, they have a size of the issue in mind. If the nep-all issue is large, editors will take a narrow interpretation of the report subject. If the nep-all ratio is small, editors will take a wide interpretation of the report subject.
target size theory & static coverage There are two things going on –The opening new subject reports improves the coverage ratio. –The expansion of RePEc implies that the size of nep-all, though varying in the short-run, grows in the long run. Target size theory implies that the coverage ratio deteriorates. The static coverage ratio is the result of both effects canceling out.
theory 2: quality theory George W. Bush version of quality theory –Some papers are rubbish. They will not get announced. –The amount of rubbish in RePEc remains constant. –This implies constant coverage. Reality is slightly more subtle.
2 versions of quality theory Descriptive quality theory: papers that are badly described –misleading titles –no abstract –languages other than English Substantive quality theory: papers that are well described, but not good –from unknown authors –issued by institutions with unenviable research reputation
practical importance We do care whether one or the other theory is true. –Target size theory implies that NEP should open more reports to achieve perfect coverage. –Quality theory suggests that opening more report will have little to no impact on coverage. Since operating more reports is costly, there should be an optimal number of reports.
results Krichel & Bakkalbaşı (2005) build a binary logistic regression analysis model. They find positive evidence for both target size and quality theory. The NEP editors don't like the results. They insist that they only filter by topic.
Bátiz-Lazo & Krichel (2005) Bernardo Bátiz-Lazo On-line distribution of working papers through NEP: A Brief Business History http://openlib.org/home/krichel/papers/kass el.pdf has an early history of NEP that covers organizational details I don't talk about here.
ernad stands for editing reports on new academic documents. Software system designed by Thomas Krichel at http://openlib.org/home/kric hel/work/altai.html. Software written in Perl by Roman D. Shapiro. Cost $2000. Started to work after 2004-12.
cut editor freedom I Editors no longer send mail to lists. Only one email address sends mail. But the mail appears like coming from the editor: From: Marcus Desjardin Reply-To: Marcus Desjardin
cut editor freedom II Editors can no longer edit report issue emails, e.g. add announcements of conferences. They are generated from XML files into standardized text and HTML files bound together by MIME multipart/alternative. They can not change dates of issue.
help editors Provide a simple-to-use interface for the composition of reports –provide an easy to scroll input –allow for easy sorting of report –do a better job at pretty-printing Get ready for the introduction of pre-sorting Actually presorting was only introduced in 2005-08.
statistical learning The idea is that a computer may be able to make decision on the current nep-all reports based on the observation of earlier editorial decisions. This is known as pre-sorting. Thomas Krichel Information retrieval performance measures for a current awareness report composition aid http://openlib.org/home/krichel/sendai.pdf deals with the evaluation of presorting.
presorting When an allport issue is created, it is presorted. In the allport rif each paper has a number in document order. That number is still reported in the presorted rif. The method is support vector machines svm, using svm_light.
pre-sorting reconceived We should not think of pre-sorting via SVM as something to replace the editor. We should not think about it encouraging editors to be lazy. Instead, we should think it as an invitation to examine some papers more closely than others.
headline vs. bottomline data The editors really have a three stage process of decision. –They read title, author names. –They read the abstract. –They read the full text. A lot of papers fail at the first hurdle. SVM can read the abstract and prioritize papers for abstract reading. Editors are happy with the presorting system.
what is the value of an editor? It turns out that reports have very different forecastability. Some are almost perfect, others are weak. If the forecast is very weak the editor may be –a genius –a prankster.
svm training set The positive examples are taking from the report up to a certain time limit, called the experience length. The negative results are taken from nep-all, from the date of the last issued subject report until the experience length. The experience length is fixed in ernad.conf. For NEP it is 13 months.
features selection We use individual words out of the contents from titles, author names, abstracts, classification data and the id of the RePEc series. We normalize the Euclidean sum of the feature weights. We run svm_light with the default settings.
presorting timeline When a nep-all issues has been created, a customized version of its rif is created in the source/us directory. This issue is then presorted. The presorted version is stored in the source/ps directory. Presorting therefore only takes account of the information available at nep-all creation time.
underlying technologies Written in Perl, using LibXSLT. Uses mod_perl under Apache 2. Runs on Debian GNU/Linux, could run on similar systems. Ernad needs to used some sort of mailing system but is not geared to a specific system. It basically just sends mail.
underlying information AMF is a format for description of academic documents and academics. http://amf.openlib.org/doc/amf.html It is based on XML Schema, itself based on XML. Report data and issue data is encoded in AMF.
ernad.conf Ernad uses one single configuration file ernad.conf. It has a simple attribute = value structure.
affordances and domains There are basically four things that an ernad-based current awareness system provides for. For each of these affordances, we have a separate domain. This allows for distinct affordances to be run by distinct domains.
the composition domain This is the domain used by the report issue composition interface. This is the virtual domain that the ernad apache is running under. The ernad process creates files so the apache server is best run as the user ernad. Recall ernad requires mod_perl, which in turn is incompatible with suexec.
the service domain This is where potential reader look at information about the ernad service –what reports are available –who edits them This domain is fixed through the reports configuration file report.amf.xml.
the list domain This is the domain where the mailing lists are under –domain of web interface –domain of the mailing lists Each report report has a list report@listdomain, where listdomain is the list domain. This domain is fixed through the reports configuration file report.amf.xml.
delivery domain The links to the full text use the encode the identifier of the paper and the identifier of the report. This allows to see what report readers requested. It is imperative that these links are not further disseminated. There should be no archives of nep lists. It is fixed in ernad.conf
reports.amf.xml The first part has the definition of NEP itself. (next slide) The second part as the definition of a report (slide after) The allport is the first listed reports.amf.xml fixes –report handles –editor information, (incl. here: editor ids) –list domain –service domain
NEP: New Economics Papers ftp://all.repec.org/RePEc/wop/conf http://nep.repec.org
Central Banking http://lists.repec.org/mailman/listinfo/nep-cba... Alexander Mihailov Alexander Mihailov http://econpapers.repec.org/RAS/pmi59.h tm......
operation A rif always as a name yyyy-mm-dd_tist, –where yyyy is the year, mm month, dd day of the nep-all issue. –tist is a UNIX time stamp, i.e. the number of seconds that have passed since the first of January 1970 rifs are never deleted. When an operation is made, a new version of the rif with a new tist is written.
creating a sri 0 After login in to ernad, an editor sees a set of allport issues to work on. This is the report selection stage. If there is no allport issue needs working on a sorry message is displayed.
creating a sri 1 When a subject report issue is created, it is copied from the source/ps or source/us directory, depending on whether the editor chooses with presorted to work with the presorted or the unsorted version of the report.
creating a sri 1.5 If there is no paper worth to be included in the report, the editor can move back from the paper selection to the issue selection stage. There she can delete the issue. The rif of the issue is not deleted. Instead an empty issue is created in the final sent directory.
creating a sri 2 Once papers have been selected, a rif is created in the into the director selected. This rif only contains the selected papers. If there are changes in the selections, new rifs are created, as soon as the editor moves to the next screen.
creating a sri 3 Once papers have been ordered, a rif is created in the into the director ordered. If there are changes in the order, new rifs are created, as soon as the editor moves to the next screen.
creating a sri 4 Once the sri has been previewed the editor can click the send button. The rif is stored in the sent directory. Ernad created a mail file containing a HTML and text version of the sri and places it in the mail directory. This file can be sent again if there is an email problem.
a lot of data ernad@khufu:~$ date Thu Oct 18 05:18:37 EDT 2007 ernad@khufu:~$ du -s ernad/var/reports/ 24147956 ernad/var/reports/ ernad@khufu:~$ find ernad/var/reports/ - type f | grep -c ^ 76043
maintenance Thomas Krichel has written a technical guide, mainly for the director and the general editor. It is at http://nep.repec.org/technical. It illustrates well that the technical maintenance is still quite heavy. A lot of maintenance scripts still have to be written by Thomas.
assessment How well does NEP work? Some criteria are already discussed the literature –delay –coverage ratio –overlap
coverage to lossage There is a web site http://nep.repec.org/lossage that does show the papers that have not been sent, as well as the coverage ratio for each issue. It appears that coverage has improved.
overlap There is no script to compute current overlap data. There is quite good historical subscriber for most of the post-ernad period. Thus it is possible to calculate overlap of reports for various nep-all issues.
to improve coverage It would be interesting to redo the work of Krichel and Bakkalbaşı (2005) For English papers, we can try to presort nep-all issues for a virtual nep-no report. This could help to identify thematic gaps. We open language-specific reports?
delay The site http://nep.repec.org/delay shows average delays of editors. Half of the editors appear to do a good job. A good job is when the average delay is below a week.
editor activity There is a web site containing activity data of editors. http://nep.repec.org/editor_activity. There appear some minor problems but overall it appears ok. Date is only available from 2005-06, because of a misunderstanding between Roman D. Shapiro and Thomas Krichel.
downloads This is the ultimate measure of success. Downloads from a report can be measured because of a GCI parameter identifying the report. Parsing the logs and matching the handles with handles in reports is difficult.
downloads from issues A lot of downloads are made by editors when they compose the report issue. A lot of others are made by robots. As for the rest, data is at http://nep.repec.org/downloads At this time this is preliminary.
the Kiev framework This is a framework I want to discuss today to assess NEP. Objective: maximize downloads of papers through NEP per paper announced. Means: –targeted report –large and targeted audience both can be influenced by the editor
unit of assessment The unit of assessment is the report issue. This is not an assessment of –the report –the editor The independent variable is related to dependent variables through simple linear regression.
dependent variable It is the number of downloads made by users from a report. We try to get to the true user data. We only look at data after pre-sorting was introduced, say 2006 and 2007. In the following, I am looking at the independent variables (i.v.)
i.v. of normalization i.v. 1: issue_size –This is the number of papers in the report. i.v. 2: membership_size – This is the number of members that the get the report just before the issue is mailed.
i.v. of membership One indication of membership quality it that is is dispersed. i.v. 3: concentration_1 –A measure of concentration of subscribers' top level domain. i.v. 4: concentration_2 –A measure of concentration of subscribers' top and second-level domain.
i.v. of timeliness i.v. 5: all_time –the time between nep-all and the current subject issue i.v. 6: neighbour_time –the minimum of the time between current issue and next issue the time between the current issue and and the previous issue
i.v. of composition i.v. 7: composition_duration –the total time of composition i.v. 8: ordering_step – the total number of times the report was ordered i.v. 9: trailer –the position of the last paper selected i.v. 10: all_size –the size of the corresponding nep-all
i.v. of season We know that the activity of RePEc is seasonal. i.v. 11 to 22: m1 to m12 –dummies to indicate the month
anything missing? the report? the editor?
Leonardo Fernandes Souto is a Brazilian PhD student working on current awareness services. He has (correctly) identified NEP as the model that all should follow ;-) His questionnaire is at http://nep.repe c.org/research/NEP_questionnaire_2007- 10-06.doc.
future extensions of ernad 1 Use editor identities to build a customized experience length. Use a collection of multi-word RePEc keywords to aid pre-sorting.
future extensions of ernad 2 Deal better with duplicate papers under different handles. –use lists before ordering as a basis for inclusion into pre-sorting. This will allow editors to delete duplicates without confusing the SVM –potentially detect duplicates at allport composition time.
future extensions of NEP Use RSS as an alternative dissemination method.
tough problems Filtering for new papers is deficient as the date on papers is not mandatory. Presorting for age seems impossible. The fight with spam is a problem for anyone who sends out a lot of mail.
finally: stop the press... 2007-10-17 Christian Zimmermann wroteAt http://ideas.repec.org/i/ e.html, I have attempted to classify registered authors by field. For this I used their papers as disseminated by NEP, and if at least one fourth are in a report, authors are considered to be a specialist of that field.
http://openlib.org/home/krichel Thank you for your attention!