The IntFilter Project
Eva R Fåhræus
Technical Report 97-003
Department of Computer and Systems Sciences
Stockholm University/Royal Institute of Technology
January 27, 1997
When I go out fishing on a lake, I also find many fishes that I don't want; some are poisonous, others are full of bones, or they just taste of mud. I bring home in my net bag all those that I know taste good, but also some that I think might taste good. I intend to ask an expert. As I am rowing my boat over the lake, I also see many things I find beautiful, interesting or odd. I settle down on an islet and enjoy the view over the lake. I see a forest on the other side and a clump of reeds near the shore and I think there might be plenty of fish there too.
Searching for interesting information on the global electronic network Internet, e g about filtering techniques, I also find things I am not interested in, either because they speak about quite different things, I have already read them, or they are too advanced for me. The net is a real treasure-trove filled with valuable information. The problem is just that the valuable information is mixed with stuff that is less valuable, plain rubbish or even misleading. I miss the outlook, the place where I can survey the supply.
If it were possible to get help from the computer to find exactly what I am interested in, then I would easier find the valuable things and thus have more time to read them. I would then be better informed about what is happening, get hints and ideas and wouldn't have to redo what others have already done.
Internet contains many possibilities and tools to help me find information. Here are some examples:
Search engines, e.g. Alta Vista, are used to search for texts in World Wide Web, using my own search key I get a recommendation about an interesting home page, from where I find links to information in the same area I can subscribe to a News group and also make my own contribution in a certain subject area and get answers I can be on a distribution list which is used to deliver information in a specific task
Information I get this way still contains things that don't interest me. I don't always know what I am searching for because I try to find new ideas or events. To be sure that I read everything that is interesting I have to read an unreasonable amount of material.
As a good help to get just the fishes I want, I can chose a net with the right size of the meshes or the right kind of bait, but that is not enough to avoid all the rubbish.
One way to find the interesting things and to avoid too much and not interesting information is to use a filter. The problem is to adapt the filter to my interest and my situation. Also, my interests change over time.
At the Department of Computer and Systems Sciences (DSV) a project is presently carried out called "Methods and Utility in Group Supporting Systems". It is financed by the Swedish National Board for Industrial and Technical Development (NUTEK), a part of the programme Co-operation and Technology. Its aim is to get utility products for individuals and groups that communicate via the Internet (Palme, 1995).
In project "Intelligent filters in group communication systems" we wanted to develop and implement a prototype and let users try it (Kilander, Fåhræus & Palme, 1997a). The present report describes a user study which explored the usability of the chosen filtering techniques and ways to improve it. The aim was not to test how easy the systems are to use in a general sense; the filtering techniques were in focus. We chose to perform the investigation on Usenet News users.
Usenet News is composed by a large, rapidly growing number of groups, on more than 10 000 different subjects going from scientific such as history, trough practical technology such as printers, to hobbies such as gardening. You can subscribe to the groups you are interested in. When you connect you get a list on every contribution made within the last weeks. You can choose which contributions on the list you want to read. The next time you enter the system the contributions you have already read are not shown.
Anyone can make a contribution on the Usenet News, thus the quality of the contributions vary a lot. There are often chains of comments, i.e. a contribution is an answer or a comment on a contribution made earlier. This can usually be identified because the contributions have the same subject line.
In 1993 and 1994 Ann Lantz made two studies with the aim to see how experienced users of the Usenet News choose the information they want to read (Lantz, 1993, 1995). In the first study she let the users think aloud while they were searching the contributions they wanted to read. In the other study, Lantz sent an electronic inquiry to members in Swedish speaking News groups. Her results showed that the users usually made their choice on keywords in the subject line.
I believe it is possible to make a computer system that helps people find interesting information. Different people think in different ways when they choose what to read. They also handle their received information in different ways. Consequently, Different people may need different types of support. But it may not be enough to ask people what support they want. You can not choose until you have had the chance to try several alternatives. And not until you have tried for a long time and found your own way of operating the tool, you can decide how much it will be able to help you.
The best way to use the computer may be to make a rather simple and easily understood sorting of the texts and then present them to the user in a way that gives a good overview.
Perhaps it is enough to make a RATHER good sorting of the texts to give the user a good help in searching for interesting information.
When I started this study I hoped to find new ideas and I was willing to abandon my preunderstandings if necessary.
This study was a field investigation using a filter prototype to read Usenet News in order to give the project the base for two parallel analyses:
1) The behaviour, thinking and experiences of users is examined and analysed.
2) Texts that have been categorised by the users and stored in the computer will serve as test exemplars to try different filtering algorithms. This analysis is presented in Kilander, Fåhræus & Palme (1997a).
A method for inductive creating of knowledge was chosen because it was difficult to find a theory for such a new phenomenon as reading Internet. The analysis was done according to 'Grounded Theory' (Strauss & Corbin, 1990, p.23).
A grounded theory is one, that is inductively derived from the study of the phenomenon it represents. That is, it is discovered, developed, and provisionally verified through systematic data collection and analysis of data pertaining to that phenomenon. Therefore, data collection, analysis, and theory stand in reciprocal relationship with each other. One does not begin with the theory, then prove it. Rather, one begins with an area of study and what is relevant to that area is allowed to emerge.
The goal was not to formulate a theory as such, but to try to get new ideas as how to support users searching texts, through the work with the analysis.
The study can be looked upon as a complement to the one made by Lantz (1995). Together they would constitute a triangulation since Lantz used the so called "think-aloud-method".
Experimenting with the prototype had the advantage that the users could do their reading and judging, without disturbance, at times that suited them best. At the same time the classified texts were saved in computer-readable form. This meant that it was possible to go back together with the users and discuss how they did the judgements and why.
By giving the users a chance to write their comments directly when judging the messages, the intention was to get both spontaneous, immediate thoughts and more reflected ones. That is, when the users have a certain distance to their behaviour and survey their sorting of messages at a later occation. Of course there was also a risk here that reactive thinking may occur, i.e.that the users made a later rationalisation of their own behaviour. The design chosen was an attempt at a reasonable balance between work done and the connection to the real world.
To describe the most important aspects of the study a conceptual framework was built (Miles & Huberman, 1994, p. 18), shown in Appendix 1; it will be commented here.
The main actor in the study is the user, who is supposed to be deluged with information and is supposed to be anxious to find what is worth reading. The user applies a number of strategies for chosing texts to read.
The main object is the contribution which reaches the user. A contribution can consist of several elements, for example sender, heading, introduction, and text. The contributions have properties like length, form, and content. Further, contributions can have relations to each other in the shape of threads or trees.
A tool that was used in the communication between the user and the filter was a system of baskets. The user chooses which baskets to use and which categories they are to stand for. Here the user puts typical examples of interesting or not interesting contributions of different categories. The user applies strategies on the elements, properties, and relations of the contributions in order to sort the contributions into the baskets.
The third main component is the filter. The filter constructs a user model from the contents of the particular user's baskets. With the aid of its algorithm the filter sorts incoming unread contributions into the user's in-baskets, so that the user can choose which basket to start with and within that basket the user begins with the most interesting contribution.
The research problems of the filter project can be stated in the following way:
What assistance does a reader of electronic messages want in order to be able to sift the reading matter and only read what is interesting?
I have chosen to start by asking the following question:
1) How do readers proceed when they sift the material unassisted?
Starting from this point I intended to look for such strategies and methods that can be automated. But the computer might be able to make a sifting that is not simple to make manually and therefore seldom is used; still this way of sifting the material could be relevant. Therefore I also asked:
2) What kind of assistance would readers like to get from the computer when sifting electronic messages?
As the overall project title, to which this study belongs, is Usefulness of Approaches in Group Support Systems I am also interested in the role of filters in co-operation.
3) In which way can a filter be used in co-operation between people?
The focus of the study is on the selection strategies that can be automated and on the user interface for experienced Usenet News readers. The design was made to catch this by letting users read Usenet News and by asking them questions about how they chose what they wanted to read.
For a potential further development, it was also interesting to study such selection strategies as could be automated in the future or strategies on which you can apply so called social filtering, i.e. when a reader puts a judgement on a text which guides another reader's reading. Further it would be of interest to expand the domain to other areas than Usenet News, above all electronic mail and less experienced readers, but this would be outside the definition of this project. I was not asking specific questions about this but I noted spontaneous comments and observations that touched upon the boundaries.
The data collections had two aims: 1) to broaden the knowledge about how users spontaneously filter their information, and 2) to get material for use in testing different filtering algorithms (rules or programmes). In this essay only the first aim is treated.
Within the project a prototype was developed, that executes filtering of contributions in Usenet News (Kilander, Fåhræus & Palme, 1997b). The prototype consists of a Usenet News browser and a classifier. The browser makes it possible to read contributions from the selected Usenet News groups. The classifier is a programme where the user sorts the contributions read, into different categories. The classifier can also give new contributions scores that tell how interesting they are assumed to be to the user.
The collecting of data proceeded in the following way: Users read contributions, with the aid of the browser, in the Usenet News group they had chosen. With the aid of the classifier, the users placed each contribution in some category which they themselves had chosen. Thus the user could continually add new categories, "baskets", for sorting the contributions. There was also space for writing a comment on why a contribution had been sorted into a certain basket.
The prototype stored the categorisation made, so that they later could be identified and studied individually for each user.
The filtering part of the prototype was not used in this section of the study.
In order to find users for my study I asked through e-mail all employees in the laboratory where I work and study. This gave me only one positive response. Therefore I talked to people I found in their working rooms. Through this procedure I managed to recruit together four users.
Before the reading started, I gathered the four users for an instruction. At this time I informed them of the aims and the design of the study. The constructor of the prototype showed them the prototype and the users themselves could try it.
The users were instructed to try and make a qualitative categorisation in addition to categorisation according to content or other ways of sorting the material they themselves had chosen. Thus each user should not only decide what a contribution dealt with (e.g. artificial intelligence) but also whether it was interesting or not to this particular person.
At this stage the procedure had a relatively tight design since users were rather constrained by the prototype and by the aim of the study.
My next step was to interview the users according to a method resembling that of Holstein & Gubrium (1995). I interviewed each user individually and together we went through the classified contributions in order to decide what principle had been used and what means of assistance could be useful. The interviews were audio taped.
After the user spontaneously and from the recorded comments had described to me how this classification had been made, I asked follow-up questions like:
"How did you notice this was interesting?"
"How did you intend to use this?" "Did you also consider who was the author?" "Can you think of some way that the computer can help you in the classification?" "Would you be helped by seeing the classification made by other readers?"
Some technical problems arose in this process. Many of the comments that User 1 had made had not been saved in the correct manner so we could not view them. For User 3 a major part, perhaps 3/4 of the categorised contributions had disappeared. The remainder was only about 20 contributions in four baskets.
Before the reading I had explained to the users that I wanted to know if the filter could discriminate interesting from uninteresting contributions, but I did not want to control the users' choice of baskets. Only one user had chosen to sort the contributions in interesting and uninteresting ones. Because of this I asked the users to try to use separate waste paper baskets when they continued their reading.
During the reading the users suggested some changes in the prototype, which the constructor implemented. After the interviews the users got the possibility to use the new changed prototype, this time with the filter in operation. They were now able to use their saved examples to "teach" the filter how they wanted the contributions sorted and try it. This test was very limited, however, due to technical problems with new equipment and lack of time.
My last step of the procedure was to gather the users in a meeting, together with the constructor of the prototype. There we discussed our experiences. This took place after my analysis of the interviews.
My aim was to look for concepts, criteria and behavioural patterns that could help us to understand how the users had chosen what they wanted to read (Hammersley & Atkinson, 1983, p. 209). I was looking for concepts and patterns that appeared to recur and I tried to find out if there were contradictions and exceptions. This comparison I did within one user's material and between users. Next I tried to find out what the most central criteria was and I looked for connections between them. This procedure has been described by Strauss & Corbin (1990, p. 57) as having the following steps: a) open coding, b) axial coding, and c) selective coding.
When the users had been working on the reading of contributions for about two months, I contacted them. Then I found out that one user had not had time to read any contributions at all, one had only read a few, while two users had gathered quite a large number of contributions each. I started with interviewing those two that had gathered a large amount of material (User 1 and User 2) and interviewed the third one (User 3) when he had had more time for reading. The fourth person decided to drop out of the study.
The interviews were carried out by me in the user's room in front of her or his computer. I asked how the general situation had been and how it had been to work with the prototype. Then I asked the user to describe her/his baskets. After certain technical problems with finding the files with the categorised contributions, I asked the user to open a particular contribution and tell what she/he had done. I tried to get a clear picture of what the user had looked at, how far the user had read, and what considerations had been made. In this way we together went through most contributions in some of the baskets. The interviews took about two hours each.
Both User 1 and User 2 had several hundreds of contributions saved, so we did not look at all of them. We chose two or three of their baskets and looked at some contributions from the beginning until we recognised a pattern. The amount of contributions saved by User 3 was limited so we could look at all of them.
The first thing I did after completing the interviews was to transcribe the beginning of the interview with User 1.
I used the following list of codes based on the conceptual framework (se part 4):
PRE Presumptions and the user's situation
STR Strategies for the choice. Principles used to put a contri bution in a certain basket. The labelling of baskets.
TAC The tactic used for judgement. How is the user doing this told in practical terms.
CO-OPP Co-operation with the prototype
CO-OPH Co-operation with other people
CON Consequences from the choice and strategy selected and used
PAR Participation, engagement and adjustment to the study, and to filtering in general
THI Thinking around the process and the attitude towards the contributions
With this list in front of me I started to read my typed interviews with User 1 and noted where I found anything that applied to the codes as I went along. I then put questions about the meaning and importance of what I had noted and found that the codes TAC and THI where important and needed specification.
To get a more detailed image of the tactics I split TAC into six codes and completed the coding of the text with these.
TAC-SU Subject line
TAC-LI The first line(s) in a paragraph
TAC-GL Glancing/skimming through the text
TAC-AL Reading all the text
TAC-SL The text is either short or long
TAC-FO The form of the contributions
In the same manner THI was split into four groups.
THI-KE Keywords
THI-QA Question-answer, threads
THI- CU Curio
THI-? Miscellaneous, unknown
With this, these two groups had been given a more precise meaning. Tactics is about what element of the contribution the user uses, while Thinking is about the contents of the text.
The results of the open coding helped me as I tried to interpret the user's way of sifting information. Therefore, I continued coding parts of the other interviews in the same manner. Now I could find typical patterns but also contradictions and exceptions.
The results of this is described for each user, illustrated by excerpts from the interviews (appendix 2).
User 1
User 1 is a young man who works with programming at one of the laboratories at the department. He reads Usenet News on a regular basis to find hints about installations and to avoid so called bugs, i.e. mistakes at the programming. He is of the opinion that he has profited greatly from it.
1) User 1 reads the subject line and sometimes he finds a keyword that determines where to put the contribution. Example:
User 1: This. A text like that, then you see the line of subject here.
E: Mm.
User 1: Because it is Kerberos version 4.
E: "Performance of CNS versus AFS."
User 1: CNS and AFS was the server to...
E: And?
User 1: I recognise that.
E.: And then you know it is interesting?
User 1: That it is not interesting.
E: So?
User 1. This is version 4 and I am interested in version 5.
E: OK. This was the old version.
User 1 : Then I put it in the basket for the old.
E: So you don't need to read any more?
User 1. No.
2) Sometimes he can already in the first line see if the contribution belongs to a thread, i.e. an answer to a question or a comment on an earlier contribution. Normally a contribution belonging to a thread is placed in the same basket as earlier contributions belonging to the same thread.
User 1: Here, this is probably also a continuation to the same threads we have seen 'CNS versus AFSK'.
3) When the subject line does not give enough information to sort the contribution immediately, User 1 continues to read the first lines of the contribution. If that is not enough he skims the text, principally by reading the first line in every paragraph. Example:
E. And these words you see now by just glancing..?
User 1: Yes, exactly, just skimming.
E: Yes?
User 1: ...First line in each paragraph, maybe.
4) Finally he reads the complete text. He then looks for keywords and quotations. He also notices if he recognises parts of the text, e.g. questions he has read earlier, or if he associates to problems he has met himself. Example:
User 1. Let's glance through the question a bit. It is not much... Concentrate on the important part. The substance. The answer.
5) He usuallu sees quickly if the text is not interesting. Then he seldom has to read the whole text. Example:
User 1.: Especially when I'm not interested. Especially if you aren't interested then you can see it quickly. Just a few lines down. And this is a short text as well.
6) The keywords User 1 uses for sorting seem to be rather few. However, they do not clearly tell if the contribution is interesting or not. Contributions that contain only "uninteresting" keywords can still be interesting for User 1 to read even if he does not save the contribution for later use. He calls it "curio" or " to check up what's going on". Example:
User 1: It is about AFS, which is a filing system in Kerberos..and it is in version 4, it is not in version 5, but that is more of an interesting curio, not something I am working with.
7) The length of the text as well as the amount of quotations and the layout of the text does not influence the user's interest. Once he comments on a text with many quotations as a "heavy text". He seems to skip the quotations reading only the original text. Example:
User 1: Here is a further development on the same thread. This is a heavy text. You can see that from the fact that it starts with a lot of quotations and then one line as a answer. "I can agree with this assessment." Then there are a lot of quotations and then one line that gives no further information. This could almost have gone into the waste paper basket.
User 2
User 2 is a woman and a teacher of systems engineering. She does not normally read Usenet News and finds it difficult to find Usenet News groups that deals with matters of interest to her.
User 2 wants to find hints that are of use to her as a teacher or that she can pass on to her students. User 2 has chosen to read a Usenet News group that are less technically oriented than the other users.
1) When she reads a contribution, she does not pay much attention to the subject line but glances through the text and gets a hint of the intention of the author. Example:
User 2: Well, the interesting here was this: 'an object-oriented curriculum'. Then that goes down among pedagogic. Then as you read it more closely, that it was object-oriented programming and so on, the start of an abstract and so on, then you see rather quickly that, well, it was interesting, it was the right area, but it was not really what I find interesting.
2) She thinks about how she can use the text. Example:
User 2: Here is someone that wants to buy something.
E: Is it some kind of hardware.
User 2: Yes, it says '...MHz ...converter...'It felt kind of irrelevant so to say.
E: Mm, and to you that does include all hardware?
User 2: Yes it does. What I have chosen are programmes and such of different kinds. Because I have no equipment to sell , uh. So it seems kind of far away so to speak.
3) She looks for the keyword. Example:
User 2: 'Object-oriented software...education'.
4) She also often looks for the context of the text and the meaning and the feeling it gives her. Example:
User 2: This is about plans for education so to say. I did not think it sounded very interesting. It felt also a bit far away. Not anything that would be interesting in the future either.
User 3
User 3 is a male teacher and researcher within information systems but also has a general interest in artificial intelligence. He has read Usenet News earlier but tired of it because it took him so much time to find so little of interest.
User 3 has no intention of having any practical use of the information he finds on reading Usenet News, he just wants to keep himself informed in a general sense.
1) User 3 has made things rather easy for himself by choosing a well secluded group and baskets with rather clear subject lines. Example:
E: Now let us see. What have you done? You have made four baskets. -agents and action -games -natural language -data mining Now can you tell me what situation you have. That is, first you have chosen one News group.
User 3: I have read only one News group. It is called Artificial Intelligence. It seems interesting.
2) He goes by the keywords, both in the subject line and in the text. Example:
User 3: Yes, you can probably do that. You usually dig for characteristic rules in this sort of 'Data mining'
3) He glances over the text, looking not only for the keywords but also for names of people he recognises. Example:
E: Would you say that as soon as you saw that line, you knew that you wanted to save this?
User 3: No. I could see it would be interesting for this basket, but not that I wanted to save it.
E: No, and how did you realise it later?
User 3: I thought that this article and this proceeding could be interesting.
E: Why?
User 3: Because it is a well-known writer and it was a conference that also is rather well known.
E: Did you read more later or ... did you read it all when you arrived there?
User 3: No I just glanced through it and noticed that it was an interesting reference.
4) When making the decision he also regards whether the contribution can be of use to him or not. Example:
User 3: Yes. But you can say that I took all contributions that dealt with language. And this is very... yes you can quickly see it is about automatic translation.
E: It was not very long. About two pages. Did you go down to look beyond what is shown on the first screen?
User 3: Yes I think I glanced it over. But really, this is interesting only if you are considering entering it [the course]. And that is not of current interest if you are sitting here.
Guiding expressions
The following expressions have been identified as guiding the subjecs reading the texts. To each expression some dimensions and properties are told that seem to be very important.
The given situation of the user, e.g.. working on a project, teaching a certain subject or being generally interested in the subject. The situation can influence the user either to add pressure on her/him or to inspire to reading. The reason to read Usenet News specifically. This could be one or several of following: keeping oneself generally informed, know what is going on, be able to help students, learn more, look for jobs, go to a conference, be able to find a certain article. Elements in contributions. It can be single words (keywords), signs that hint that the contribution is a part of a chain of contributions or is an answer to a question. The elements are sometimes very clear and easy to find, but sometimes they demand careful reading in order to be found. Associations, i.e. what a user thinks about while reading a certain contribution., e.g. a person interested in the contribution. The users seem to have difficulties in telling me about their associations. They are even more difficult for me to make out. Classification, i.e. the user's decision to put or not to put a contribution in a certain basket.
In the open coding, guiding categories and some of their properties and dimensions had been identified.
Now I wanted to make connections between the categories, and possibly see if there were new connections. Thereby I used the paradigm model, that is described by Strauss & Corbin (1990, p. 99)
A) CAUSAL CONDITIONS-> (B) PHENOMENON-> (C) CONTEXT->D) INTERVENING CONDITIONS-> (E) ACTION/INTERACTION STRATEGIES F) CONSEQUENCES
I described the meaning of what I had found out in the interviews by this model. I did not use the exact quotations but I rephrased the users' words, to make them suit the pattern given by the paradigm model.
A) When I see-> (B) the word Kerberos-> (C) I know this in interesting for my work and-> (D) belongs to the category "security systems",-> (E) and thus I save it in the K-basket.-> (F) so I can find it and read it later. (User 1.) (A) When I read here-> (B) I see that it is some kind of question, -> (C) about teaching.-> (D) I then think that I can contact this person -> (F) to hear if he has got any answer .->(E) So I save it in the basket for teaching. (User 2)
One can note certain differences between the three users. User 1 and User 3 expressed themselves more detached as they choose what to save. If a contribution was about games User 3 put it in the "games basket" disregarding his feelings for the contribution. User 2, on the other hand often expressed herself in words as: "it does not feel relevant to me".
I tried to find something that described a process or a change. The study was going on for such a short time that it was not likely that a user's interests would change due to new tasks at work or improvement in any area. But I noted that User 2 realised during the study how the contributions normally are disposed and adjusted her behaviour accordingly, e.g. she noted that an answer to a question usually starts with the question quoted. When she saw this she could quickly see that the answer should be sorted together with the question.
I made a verbal description of what the studied activity dealt with, thus finding the core category. (Strauss & Corbin, 1990, p 119).
The main story seems to be about how a user, by just glancing at a part of a text, surveys it and decides what it is about and if it is worth reading, now or later and if so under what circumstances. The decision is also whether to read more before deciding what to do.
From this description I made a preliminary choice of 'classification of the text' as my core category. I wanted to know more about this. If possible I also wanted to find methods to automatise this. The classification can largely be done in interesting or not interesting, but the group of interesting items can also be divided into smaller groups depending on the wishes of the user.
Next step was to relate the other categories to the core category according to the paradigm model. This could be done on a more abstract level than at the axial coding.
(A) Perception of -> (B) elements in a contribution -> (C) give, if they are recognised -> (D) an association -> (E) that results in a classification of the text -> (F) which decides whether the text is saved for later use.
Perception is the first condition to make a classification possible. It means that the user directs her/his eyes to the screen and sees something that is the element in the contribution that the user first sees. Usually, it is the subject line, but it can also be a familiar word in the text or the format of the contribution that catches the eye of the user.
The user must recognise the elements in the contribution to be able to make a conclusion from the perception. The recognition leads to one or several associations, e.g. a project , where the user could use the information received or a person that would like a hint. The association could also be that the user does not understand the word and therefore regards the contribution as not interesting. If the element is not recognised, the user will look for another element in the text until something is recognised or the user regards the contribution as not interesting.
Depending on the association the user will decide how to classify the text. Because of the abundance of information to judge the user usually does not spend much thought on each contribution. The classification is manifested by the user saving the contribution in a chosen basket so that it can be found later, or "throws it in a wastepaper basket".
At this meeting User 2 and 3 were present together with the Prototype's Constructor Fredrik Kilander (FK) and myself (EF). Unfortunately it was not possible to find a time when all the users could be present. Appendix 3 shows a report of the discussion.
I started by asking the users to describe their thoughts and felings during the investigation. After that we had a short discussion of how to get help from the computer in finding readable texts in a News group.
The users questioned the feasability of the chosen filtering technique and its possibility to discriminate between interesting and not interesting texts. They seemed to have a greater trust in their own ability to do a visual judgement of texts.
What the user wants, according to the meeting, is "the well written, correctly spelled test, those with a long survey and nice layout." Can the computer be taught to automatically detect such properties of a message?
Therefore, the users believed in a solution where the computer presents contributions visually, thus easing the user's judgement.
Asking about how a filter can be used in co-operation I got the answer that reading can be shared between members in a group (social filtering).
During the discussions as the prototype was developed, at the preparatory studies, at the interviews, and at the final meeting we have had many ideas about how the computer best can support the user. It is impossible to tell exactly who got which idea, but theory and real life have worked together, and one idea has made way for another.
I will now systematically try the ideas for each category I found essential by processing the interviews.
The perception can be supported by the system presenting the elements in the contribution more clearly. The subject line and the keywords can be displayed brighter, in bold style or in different colours. Questions and answers can be shown together and perhaps in a different colour or style. Connections between contributions in a chain can also be shown graphically. By showing the complete contribution in a smaller scale, the system can quickly give an impression of the outline of the contributions, format and size. These measures will make it easier for the user to observe recognisable elements valuable for the classification.
To support the association, the system could show the baskets next to the contribution, thus reminding the user of chosen classifications. The users shall also be able to look into a basket and see the contributions already put there, reminding them of chosen classification strategy.
Saving a contribution can be made simpler by direct manipulation, i.e. the user can click and drag the contribution over the screen and drop it into the basket.
To make it easier for us to bring home those fishes we want to eat, we could have a system showing pictures of the fishes that are common in the water we are fishing in and that we like. The system could also give us advice about suitable tools. This would make it easier for us to recognise and catch the fish we want.
Through this study I have received support for and further developed two of my presumptions that I presented in part 2:
1) The computer should be used to present the contributions in a clear way.
2) Different users have different wishes and needs. There are differences between different users and the same user on different occasions.
But I have not got any support for my theory that even a RATHER good sorting of the contributions can be of great help. The study lasted too short a time to make the users really able to use the prototype efficiently. But a rather good sorting combined with a smart visual display would help the user to do their own choice.
The study confirms the conclusions drawn by Ann Lantz (1995), i.e. that keywords in the subject line are very important as to how the users judge the contributions. My study completes it with a more detailed image of how the users act.
The study does not give any reason to reject the filtering technique chosen but the users say they question the possibility to "teach" the computer what is interesting or not interesting. Instead they suggested some criteria to use.
Filters can very well be used by groups in co-operation, by letting different categories of contributions be sent to different individuals watching different areas for one another. The idea of letting users put a judgement on contributions, that others can benefit from (so called social filtering) is only superficially investigated in this study.
Hammersley, M. & Atkinson, P. (1983). Etnography, Principles in Practice.
London: Tavistock Publications.
Holstein, J.A. & Gubrium, J. F. (1995). The Active Interview. London: SAGE Publications.
Kilander, F., Fåhræus, E., & Palme, J. (1997a). Intelligent Information Filtering. Technical report 97-002, Department of Computer and Systems Sciences, Stockholm University/KTH.
Kilander, F., Fåhræus, E., & Palme, J. (1997b). PEFNA The Private Filtering News Agent. Technical report 97-004, Department of Computer and Systems Sciences, Stockholm University/KTH.
Lantz, A. (1993). How do experienced Users of the System Usenet News select their Information? Technical report, Department of Computer and Systems Sciences, Stockholm University/KTH.
Lantz, A. (1995). Useful Criteria for Intelligent Filtering? Technical report 95-042, Department of Computer and Systems Sciences, Stockholm University/KTH.
Miles, M.B. & Huberman, A.M. (1994). Qualitative Data Analysis. London: SAGE Publications.
Palme, J. (1995). Usefulness of Approaches in Group Support Systems.
Application to the SAMT-programme. Department of Computer and Systems Sciences, Stockholm University/KTH.
Strauss, A. & Corbin, J. (1990). Basics of Qualitative Research: Grounded Theory, Procedure and Techniques. Newbury Park: SAGE Publications.