Abstract
In previous research by Trepess et al (1998) a framework for erroneous actions in Computer Supported Co-operative Work (CSCW) environments was proposed. This paper will extend this research to see how the framework can be applied to erroneous actions found in CSCW use. Experiments were conducted to allow the observations of users performing a task using video conferencing and shared workspace software. This paper will describe the experimentation that took place and look at how the results can be mapped on to the framework proposed in Trepess et al. (1998).

Keywords
Erroneous actions, framework, Computer Supported Co-operative work (CSCW), human error classification.

1. Introduction
Research into human errors has been progressing for many years now including studies by Norman (1981), Rassmussen (1987),  Hollnagel (1993), Maurino et al. (1995), Fields (1995) and Trepess et al.(1998). However, very little research has been completed in the occurrence of human errors or erroneous actions in CSCW applications.

Previous research by Trepess et al. (1998) examined different human error classifications by Norman (1981), Reason (1987, 1995) and Hollnagel (1993) and then applied them to a CSCW context by using CSCW frameworks by Mantovani (1996) and Dix (1994). The result of this research was the proposed framework that will be examined in more depth in this paper.

The following sections will describe the framework, the experimental method and then analyse the results in the context of the framework. The aim of this research was to test the framework on real errors that are found in the use of CSCW applications.

2. The framework
The proposed framework (fig. 1) is an extension of the framework described in Mantovani (1996). In the proposed framework there are three products which are obtained at each of the three levels. These products consist of goals, plans and actions. There are three main classifications of errors that can occur in the formulation of these products, these are:

(a) social conflicts,
(b) planning conflicts,
(c) interaction conflicts.

Social conflicts are conflicts of socially contextualised goals of different groups or individuals participating in the collaboration. Planning conflicts are when there is a conflict in plans arising from differences in opportunities and interests. Interaction conflicts are when errors at the level of interaction are caused due to inappropriate planning. Fig. 1 depicts errors in respect to their cause, consequence, detection, recovery and the event.
 

The cause of an error can only be found in levels one and two. This is because at level three errors, even slips, are caused through alterations in the opportunity and interest of the actors. For example a sudden distraction causes a change of interest for a split second thus causing a slip or memory lapse. A slip can be explained in the inappropriateness of a plan for a particular actor. The actor may devise a plan to type a word quickly but may not have the typing skills to accomplish this goal and thus the probability that an error is made is high.

The consequence of errors is initiated at level 3 where the interactions are made. Erroneous actions at this level are active failures and can only have single user consequences (SUC) in a similar form to those seen in single user systems. Errors with multiple user consequences (MUC) are latent failures and become erroneous at level 2 because they are only erroneous in the context of other users. As far as the user who made the error is concerned the interactions made are understandable and correct.

Error detection is made at level 1. This is because an error can only be detected when its consequences are perceived to deviate from the goal set at this level. The process of error recovery also begins at this level. The error recovery plan has to start with the formulation of a goal followed by the formulation of the plan followed by the actions necessary to carry it out.

The event loop illustrates that erroneous actions can occur and have consequences at any level of the model. For example a social conflict at the level of social context will cause revisions to the plan at the level of interpretation of the situation which in turn will effect the local interactions with the interface. Likewise an error at level 3 which has MUC’s will cause errors in the plan at level 2 and result in deviations from the goal at level 1.

3. The experimentation
The experiments were set up in order to allow observations of participants behaviour when completing a task using CSCW technology. Four experiments were carried out each containing three participants. The Participants had to communicate using email and video conferencing technology and had to complete the task using a shared white board. The situation was that one participant (Actor A) had email as their only means of communication and had a full, paper based, copy of a data flow diagram (DFD). The other two participants (Actor's B and C) had email, video conferencing and a half completed version of the DFD on the shared white board. The task was that Actor A had to give instructions to Actor's B and C on how to complete the diagram who then had to discuss the emailed instructions and complete the diagram. Actor A had no awareness of the contents of the half completed diagram. Seed errors were incorporated in the experimental design to examine how people attempt resolve them and the impact that they have on the task. Each group had one hour to work on the task. Failure to complete in the time allowed was considered an erroneous situation. Data was collected and analysed through video recordings and email logs.

4. Mapping of the results into the framework
This section will examine a selection of erroneous actions that occurred in the experiments in the context of the framework elements. Each erroneous situation will be described in relation to their cause, detection, consequence and recovery.

The details of the first erroneous action are that Actor A sent out some instructions but the instructions were incorrect as there was an omission of a data source box. This error was an error of interaction conflict that had multiple user consequences(MUC's) according to the framework. The cause of this error in the first experiment was that Actor A was grouping the elements of the diagram according to their position on the screen to ease instruction but because the source box was individual it did not become part of any of the groupings and thus was omitted. The cause can be seen to be at level two of the framework because it is Actor A's interpretation of the situation that differs from that of Actor's B and C. The consequence of the error was the creation of confusion and uncertainty in Actor's B and C due to the differences in the instructions compared to the reality of their situation. The detection of the error occurred when the task could not progress due to the erroneous consequences. In order to recover from this a question and answer session over email was necessary for clarification of the situation. This involved forming a new plan and recovery goals at level 1 of the framework.

The next example is where there was a reversal of control. A reversal of control in this case was when Actor A, who was the logical controller, failed to control the task and thus another Actor, who did not have the proper resources to control effectively, took control of the task. This is an example of social conflicts where there is a conflict of expectations from each user. The cause of this, at level 1, was that Actor A was not sending many emails and the mails sent did not contain sufficient information. The cause also overlaps into level 2 as it ultimately leads to planning conflicts. The consequence of this was that very little progress made was. Recovery was attempted by Actor's B and C by taking control and asking for information based on their current knowledge. Recovery can be seen to begin at level 3 as Actor's B and C were required to review their goals and expectations of Actor A.

The next example is a combination of two errors, first the omission of certain instructions which resulted in the second, the incorrect deletion of a data store box from the diagram. The cause of this situation was that the data store box was omitted from the instructions and was thus deleted from the diagram. This error is similar, in cause, to that seen in the first example, and likewise is an interaction conflict, but it differs in its consequences. The consequence of this error was that Actor's B and C had a discussion about whether to delete the box or not and ultimately lead to the deletion of the box. One wanted to delete the box the other did not but the box was deleted because of the dominance of the actor who thought it was the correct action. This was erroneous in that both Actors had different interpretations of the situation at level 2 and constructed goals according to these interpretations at level 1. Detection of this error was made when a reference to the data store box was made in a later email and Actor's B  and C realised that it was no longer there and thus deviated from the goal at level 1. Recovery was made on a request for confirmation on the existence of this data store box. Instructions were sent by Actor A to recover from the situation.

The final error that will be examined is where there was ignorance of social norms connected with email message which comes under the classification of social conflicts. The error occurred because the participant acting as Actor A did not understand the significance of the '>' in most email packages. The significance being that the '>' denotes text from a previous message in a reply. The consequence of this was that the participant entered some of their message in a section beginning with a '>' which meant that the recipients were not sure where the new message began and an old message ended. No detection or recovery was made because nobody informed Actor A of the social norm.

5. Conclusions
From examining these erroneous actions within the context of the framework the different levels of cause, consequence, detection and recovery can be seen. The examples also illustrate the importance of hi-lighting the elements that are present within the framework. For example the different levels of conflict that can arise in CSCW use has proved to be an important observation. The research also shows the relevance of this framework in a collaborative learning environment where computers are used as the communication medium. It can also be seen that errors in CSCW environments are complex events and consist of many factors and result in many different types of consequence. It is clear that in CSCW applications a new level of error detection and recovery is necessary for erroneous events.

Further work to be carried out in this area is to test the framework on a larger user sample and to explore different methods of detection and recovery to deal with the erroneous actions like the ones seen in this research.

References
Dix, A. J. (1994). Computer-Supported Co-operative Work - A Framework. In ‘Design Issues in CSCW’. Eds. Rosenburg, D.   and Hutchison, C. Springer Verlag, pg 9-26.

Hollnagel, E. (1993). Human Reliability Analysis: Context and Control. Computers and People series, Academic Press.

Mantovani, G. (1996). Social Context in HCI: A new framework for mental models, co-operation, and collaboration. Cognitive Science Vol. 20.

Maurino, D., Reason, J., Johnston, N., and Lee, R. (1995). Beyond aviation human factors. (Pub.) Avebury aviation.

Norman, D. A. (1981). Categorisation of action slips. Psychological Review, Vol. 88.

Rasmussen, J. (1987). The definition of human error and a taxonomy for technical systems. New Technology and Human Error, New Technology and Work, Wiley Series, Ed. Rasmussen J., Duncan, K. and Leplat, J.

Reason, J. (1987). A framework for classifying errors. New Technology and Human Error, New Technology and Work, Wiley Series, Ed. Rasmussen J., Duncan, K. and Leplat, J.

Trepess, D. and Stockman, A (1998) A classification and analysis of erroneous actions in computer supported co-operative work environment. Appearing in special edition of Interacting With Computers on human error.