"1. Previous research on motivation recognition based on learner-computer interaction. {1} The research was based on captured screen activity. Within this framework, the two objectives of this study are: a) To identify variables related to the three motivation dimensions, which are computable from data stored in the system log files; b) To cluster these variables, based on empirical data, for classifying them according to the three motivation dimensions. 3 Methodology. 3.1 Procedure. The study was carried out in three consecutive phases using different methodologies: Phase I – Constructing the Research Framework. This literature-based phase was used to conceptualize the terms to be used in our motivation research. Within the framework, three dimensions were defined (Engagement, Energization, Source); see previous section. Phase II – Identifying Variables. In order to choose and define motivation-related variables, Learnograms – visual representations of learning variables over time – were used as the main research tool (as described in [13]). Learnograms presenting students' activity (N=5) in an online vocabulary course for adults (see section 3.2) were observed, in order to identify the relevant computable variables. The compatibility of the variable to previous empirical research in this field was taken into consideration, as well as their association to our framework. At the end of this phase, seven variables were identified. Phase III – Classifying the Variables According to the Motivation Dimensions. An empirical study for the evaluation of the identified variables was conducted (with the same learning environment used in phase I). Log files of a large population (N=2,162) for one month (April 2007) were collected and preprocessed. Students using the researched system belong to different courses (varied by length, intensity, starting date and proximity to the exam), however this logged segment was analyzed regardless the student's learning stage. A filter was applied for keeping students with at least 3 active sessions (N=1,444). Algorithms for calculating the variables were formally written and implemented using Matlab. The dataset was preprocessed and the final set of cases to be analyzed was defined (N=674). Finally, Hierarchical Clustering of the variables was applied using SPSS, with Pearson Correlation Distance as the measure and Between-groups Linkage as the clustering method. 3.2 The Learning Environment. A simple yet very intensive online learning unit was chosen as the research field. This fully-online environment focuses on Hebrew vocabulary and is accessible for students who take a face-to-face preparatory course for the Psychometric Entrance Exam (for Israeli universities). The online system is available for the participants from the beginning of the course and until the exam date (between 3 weeks and 3 months in total). The system includes a database of around 5,000 words/phrases in Hebrew and offers varied instructional strategies: a) Memorizing, in which the student browses a table of the words/phrases along with their meanings; b) Practicing, in which the student browses the table of the words/phrases without their meaning. The student may ask for a hint or for the explanation for each word/phrase; c) Searching for specific word/phrase; d) Gaming; e) Self testing, in the form of the exam the students will finally take. Throughout the learning process, the student may mark each word/phrase as ""well known"", ""not-well known"" or ""unknown"". This information is stored and being used by the system. 3.3 Log File Description. The researched system logs the students' activity, thus each student is identified by a serial number. Each row in the log file documents a session, initiating by entering the system and ending with closing the application window. For each session, the following attributes are kept: starting date, starting/ending time, number of words marked as ""known"" at the beginning/end of the session, ordered list of actions and their timestamps. 4 Results. 4.1 Motivation-related Variables (Phase II). The authors have examined the Learnograms of a few students (N=5), searching for interesting patterns and irregularities, while considering learning behavior which may be related to motivation. Seven variables were identified and calculated (see Table 2). 4.2 Classifying the Variables (Phase III). The variable distributions were examined over the 3-sessions filtered dataset (N=1,444), see Figure 1. Three of the variables had a significant 0-values noise: wordMarkPace, examPC, gamePC, thus cases with 0-value in them were cleaned for focusing on the positive-value cases. Since the variables were skewed in the final dataset (N=674), we used transformations of log (timeOnTaskPC, avgSession, wordMarkPace, examPC, gamePC) and square-root (avgActPace, avgBtwnSessions). Table 2. Motivation-related variables. Figure 1. Distribution of the variables before cleaning and transformation were applied (N=1,444). The clustering process is described by a dendrogram (from the Greek dendron ""tree"", -gramma ""drawing"") presented in Figure 2. The vertical lines determine which variables/clusters were grouped together and at which stage of the algorithm (from left to right). For example, the first coupled variables were timeOnTaskPC and avgSession, and next examPC and gamePC were grouped. The resulting clusters appear in Table 3, their relation to the motivation dimensions is given in the Discussion below. Figure 2. Dendrogram of the hierarchical clustering process. Table 3. The resulted clusters and their mapping to the motivation dimensions. 5 Discussion. In this study, an empirically-constructed tool was developed for log-based measuring of online learners' motivation. Motivation is measured by three dimensions – Engagement, Energization, and Source – and by seven computable variables corresponding to these dimensions (see Table 3). The classification of the clustered variables to the three dimensions is based on previous research in this area. The variables timeOnTask and avgSession, which form the first cluster, might be related to the extent of Engagement, as it was previously suggested that working time might be a measure for attention or engagement [3, 17]. examPC and gamePC - grouped together in the second cluster - reflect the student's Source of motivation; it may be reasonable to hypothesize (inspired by, e.g., [9, 15]) that students who frequently tend to take self exams (related to performance-goal orientation) have extrinsic motivation to learn, while those who tend to game applications (related to learning-goal orientation) are intrinsically motivated. The variables avgActPace and avgBtwnSessions are also clustered together with the previous two, but their closeness to Source of motivation is yet to be established. The variable wordMarkPace, indicating the word marking speed, forms the third cluster. According to a diagnosis rule found in de Vicente and Pain [5], fast speed of activity together with high quality of performance (when staying in similarly-difficulty exercises) suggests increasing motivation. Since an increase in the number of words marked is an indication of the student's perceived knowledge (i.e., a reflection of the performance), wordMarkPace might be related to the direction of motivation, i.e., Energization. The tool developed in this study enables to measure online learners' motivation by using solely information stored in log files. However, there are three limitations to this innovative tool. First, variables were identified based on a specific learning environment; it might be useful for similar systems, but for different environments (varied by, e.g., learning domain, instruction modes available) these variables should be converted, and their clustering should be re-examined. Secondly, the classification to the motivation dimensions within the framework has not been validated yet, as well as their actual scales; it is within the authors' agenda to continue in the direction of validation. Third, the tool might not be complete; we only focused on seven variables, however others might be considered. Identifying these variables based on a segment of the learning makes it possible to employ this tool during the learning process; that way, intervention when needed might be possible, and changes in motivation may be analyzed. Further to the developed tool, the process used in this study – i.e., constructing a literature-based conceptual framework, using Learnograms for identifying variables, clustering and classifying these variables within the framework - is of great importance, since it is a procedure which might be transferable to other domains (e.g., anxiety, self- regulation) for developing measuring tools. Measuring the online learner's motivation has a major role in the instruction-learning cycle. Monitoring the learner's motivation might enable the instructor to interfere when needed (e.g., when student's motivation is decreasing), and should help in developing of intelligent tutoring systems which react not only to the learner's cognitive behavior but also to her or his affective situation. The overall objective of this underlying approach is to increase the efficiency of the learning process."