More specifically, it addresses the placement problem in informal learning settings in which there is no clear and structured information on the individual learning resources available, a scenario that is in increasing demand in professional learning
In our state of the art analysis we have seen that LSA is on the one hand successfully used in several educational applications but on the other hand most implementations work with much larger text corpora as we expect to have in learning networks. Another constraint for our experiment was that most of the experiments are done the englisch text corpora. Since we were planning to develop we-services for the learning environments of the Open University of the Netherlands we have used Dutch text corpora For this purpose we first did a methodological study to see if LSA can discriminate between similar and dissimilar documents.
After a review of existing solution we have developed a web-service around several LSA libraries. This web-service can import documents, preprocess them with weighting, stopwords etc., execute the SVD and finally query document to document comparison.
In the TENCompetence project we have developed a so call hybrid personalization prototype. This prototype combines so called “top-down”-approaches using metadata/ontologies with “bottom-up”methods like LSA and collaborative filtering.
A placement web-service for lifelong learners
A Placement Web-Service for Lifelong Learners Marco Kalz, CELSTEC, Open University of the Netherlands [email_address]
Motivation/Problem Description <ul><li>How to find learning activities that fit to the prior knowledge of learners? </li></ul><ul><li>….in informal learning environments/learning networks? </li></ul><ul><li>….with the absence of metadata or ontological descriptions? </li></ul><ul><li>This is what we coin the “positioning” or “placement” problem. </li></ul>
Approach: Latent Semantic Analysis (LSA) <ul><li>Assumption: We can model the importance of learning activities via the similarity between the content of the learner portfolio and the content of learning activities </li></ul><ul><li>Method for Similarity Calculation: Latent Semantic Analysis (Deerwester, Dumais, Furnas, Landauer, & Harshman, 1990) </li></ul>
Latent Semantic Analysis (LSA) I <ul><li>LSA is a theory and method for extracting and representing the contextual-usage meaning of words by statistical computations </li></ul><ul><li>Complex method involving several steps: </li></ul><ul><ul><li>Preprocessing: </li></ul></ul><ul><ul><ul><li>Stopwords, Weighting, Local Frequency, Global Frequency </li></ul></ul></ul><ul><ul><li>Term-Document-Matrix (TDM) </li></ul></ul><ul><ul><li>Singular-Value-Decomposition (SVD)->LSA space </li></ul></ul><ul><ul><li>Querying: Terms similar like term X, Documents similar like document X </li></ul></ul><ul><ul><li>Advantage to keyword-based approaches: Latent relations </li></ul></ul>
Latent Semantic Analysis (LSA) II San Miguel 2006
Evaluation Studies <ul><li>Study 1: Methodological study to evaluate the general applicability for our purposes. Special focus: small corpora </li></ul><ul><li>Study 2: Performance Evaluation of using LSA as a classification method for “meaningful” content for an individual learner </li></ul>
Evaluation Study I Important Results: The application of up to 50% stopwords and the identification of the ideal number of dimensions to retain in LSA are important factors for the application of LSA with smaller corpora
Evaluation Study II Table 1: Confusion Matrix for LSA as classifier for “meaningful” content (n=504) 424 12 Irrelevant 40 28 Relevant LSA rating Irrelevant Relevant Human rating
Evaluation Study II AUC (Area under the ROC curve) value= .81 (95% CI, Std Err. = 0.0262)
Architecture of the Placement Web-Service Layer for interfacing with Controlling Environment Sparse Matrix Library <ul><li>Lsa Engine </li></ul><ul><li>import </li></ul><ul><li>preprocessing </li></ul><ul><li>decomposition </li></ul><ul><li>query </li></ul>SvdLibC WinGtp ALGLIB Document s Web Service Layer with SOAP Interface Output
Architecture of the Placement Web-Service Iduser = Integer Learninggoal = Array (Strings) EventType On request. Format DATA Fields Frequency 2-dimensional array of floats. Each UoL with its calculated cosine values. Iduser=xx Return a list of UoL annotated with cosine values Get getPositionValues Output Input Description Method Name Placement Web-Service Interface (API)
Role of the Placement Web-Service in the TENCompetence infrastructure Drachsler, Herder, Kärger, Kalz 2008 Curriculum Planner Navigation Web Service Preference Engine Placement Web Service Curriculum -based Service Preference -based Service Navigation -based Service Positioning -based Service SOAP SOAP LPath creator LPath Provider Learner metadata Learning Paths getLocation(UoL) getLocations(UoLs) createLearningPath, getLearningPath(ID) Hybrid Personalizer Configuration Framework Units Of Learning Metadata
Contact Thank you for your attention! Questions? Marco Kalz, M.A. Centre for Learning Sciences and Technologies Open University of the Netherlands P. O. Box 2960 6401 DL Heerlen, The Netherlands [email_address] Presentation will be at http://dspace.ou.nl & http://slideshare.net/mkalz