Annotation (and search) of Lecture Videos

Supervisor: Erica Melis, Eric Andres

The goal of this thesis is the development of a tool through which students can annotate lecture videos by slices/titles, comments, and metadata (als Nacharbeit der Vorlesung), potentially modifying a Canadian tool. The annotations will be used for searching the pool of videos.

Automated Annotation (and search) of Lecture Videos

Supervisor: Erica Melis, Paul Libbrecht

The goal of this thesis is the development of a tool/data mining technology for automatic annotation of lecture videos in order to enable (semantic) search in a pool of videos.

Replay of Students' Actions (auch Lehramt Studenten)

Supervisor: Erica Melis

The goal of this work is to realize the replay of the history of students' actions in ActiveMath which is available through the event-system and stored in a/the GUI of the ActiveMath system. The reason is that this will enable humans to watch the history of actions of a (indicated) student and judge it. This implementation requires an interpretation of events in the GUI and possibly an extension of the event taking if it is found that important information is not yet stored.

It has to be tested with human judgers whether this replay is sufficient for judging.

Teacher Tool (auch Lehramt)

Supervisor: Stefan Winterstein

The goal of this master thesis is to further develop and integrate a 'teacher tool' for the web-based learning environment ActiveMath. Based on use cases the needs and interests of teachers/tutors for information, the presentation of figures about their students' activities and content's usage will be specified. The informaiton will be extracted from ActiveMath's logfiles (extension of what exists already) and displayed in a dedicated user interface.

Learner Model: Rating

Supervisor: Ahmad Salim Doost, Erica Melis

In order to provide a more intuitive value on the progress of a student, the learner model should not only compute an estimation of the learner's current mastery, but provide additional information. The current state of proficiency can be divided into (at least) three different variables:

1. how well did he/she perform in previous exercises (mastery)
2. how much of the available exercises have been worked on (breadth)
3. how sure can the learner model be about it's mastery estimation (depth)

Enhancing machine classifiers with NLP techniques

Supervisor: Bruce McLaren

We have developed a series of machine classifiers to help us analyze, in real-time, the interesting discussion moves made by collaborating students in an e-Discussion. An essential step here is the extraction of features from discussion contributions which are amenable to machine analysis. Given the nature of the task, these features comprise structural aspects of the e-Discussion graphs (e.g., the number of in and out links of nodes in the e-Discussion graph) as well as language characteristics which are currently extracted by means of shallow text processing (e.g., unigrams, bigrams, punctuation). The task in this Masters Thesis project is to explore and test possibilities to improve the classifiers by more advanced language processing techniques (e.g. use of a semantic lexicon like WordNet, Named Entity Recognition, etc).

The task would involve:

1. Reading and reviewing the relevant literature (e.g., text and dialogue act classification)
2. Finding appropriate NLP libraries and tools
3. Pre-processing the existing data so that the tools and libraries can be applied to the data
4. Running the experiments with different settings (baseline: the already existing classifiers)
5. Interpreting the results, especially indicating whether the new approaches have been successful
6. Describing possible future work

Using LSA to analyze e-Discussions

Supervisor: Bruce McLaren

One major problem in the analysis of natural language texts is the existence of an almost unlimited number of ways to express the same message. Latent Semantic Analysis (LSA) is one approach to remedy this problem by mapping single words and texts to a semantic space in which texts with similar content appear near to each other, even if both texts do not share a single common word. Prior work in the area of dialogue act classification harnessed LSA to classify new dialogue moves using a nearest neighbor approach. Other usages of LSA include the identification of correct and incorrect aspects of student answers in human-machine dialogues, and the automatized grading of student essays. The task in this Master Thesis project is to apply the same principles to the area of our interest, namely the automated analysis of e-Discussions. Can we find occurances of prototypical arguments? Can we determine whether crucial arguments, which we would expect for a given topic, are missing? Can we assess the overall quality of a discussion? The task is to research these and related questions.

The task would involve:

1. Reading and reviewing the relevant literature (LSA basics and applications)
2. Finding / Implementing appropriate tools and libraries to build a LSA space and to run controlled experiments with varying conditions (different background corpora, different LSA parameters)
3. Finding / Collecting appropriate text corpora (general and topic-specific)
4. Pre-processing the existing data so that the tools and libraries can be applied to the data
5. Running the experiments with different settings
6. Interpreting the results, especially indicating whether the approaches have been successful
7. Describing possible future work

Learning by knowing different solution paths

Supervisor: Erica Melis

In ActiveMath students can exercise interactively and study work examples. In some cases it is impossible to determine the solution path a student has chosen just by looking at the (correct) result. Therefore, the attribution to the correct competency is hard to determine.In order to overcome this limitation, and further extend students' ability to see and use different solution path to solve the same problem, students' can be asked to specify their solution path from several possibilities provided. Additionally they can be requested to solve the same problem using an alternative solution path. The task:

1. design/select exercises that allow for the multiple solutions,
2. extend the exercise and worked example representation so that it allows for alternative solutions,
3. develop an extension of ActiveMath that provides students with alternative solutions of worked examples and exercises upon request as well as unsolicited and possibly asks the student to compare the solutions,
4. test the extension of ActiveMath with students,
5. use the resulting log files and Machine Learning techniques to evaluate the benefit of the extension.

1. Does this approach lead to better perfomance?
2. Does the confrontation with additional solution paths increase knowledge transfer to similar problems?

Peer Authoring

Supervisor: Erica Melis

The goal of this master thesis is to develop a simplified authoring tool that can be easily used by students to:

1. vary a given interactive exercise or exercise template,
2. author an exercise, a text, add links for peer students,
3. add or author explanations to a worked solution of an exercise,
4. add or author feedback in exercises,
5. add useful visualizations and facilities to put the new content into a content collection so that ActiveMath can present it to a group of peers.

Interactive Concept Mapping

Supervisor: Martin Homik

Concept mapping is a well known cognitive technique for visualising and constructing relationships between different concepts. It bases on the assumption that the human brain consists of a (hierarchically) structured network of concepts and that concept maps are a visual externalisation of this structure. Hence, concept mapping is known for stimulating analytical and reflective thinking.

ActiveMath provides an interactive concept mapping tool called iCMap. In the past two years, we have evaluated it and acquired a lot of new insights which contributed to the decision to reimplement iCMap. Hence, we have some good starting point literature, however, this task is still ambitious.

A perfect candidate for this task is a computer scientist with either psychology or teaching education as minor subject. We plan to make evaluations towards the end of the thesis. There are a lot of questions to investigate, be it on usability or on learning effects.

The technical requierements are: we want a rich client Java application. Its architecture should be modular as much as possible to be able to extend it or to replace components by new ones. It should include a rule-based engine which simplifies authoring of inference rules for checking consistency, transitivity, etc. The rules should be stated in a declarative way. One possibility could be to use JBoss Rules, another to use some AI techniques. We'd like to adda more intuitive suggester mechanism and a more suitable layout algorithm (Piccolo framework could be nice). The tool should be sexy, so if you like to play with Java 2D, this is your playground.

Neue adaptive Szenarien für Lernumgebung ActiveMath

Supervisor: Erica Melis

Zu entwickeln ist ein Hausaufgabenszenario für das intelligente Lernsystem ActiveMath. Bei diesem werden nach mehreren Kriterien adaptiv Aufgaben für Hausaufgaben in Mathematik zusammengestellt. Dieses soll nach im SFB-Schwerpunkt 'Bildungsqualität Schule' erarbeiteten pädagogischen Vorstellungen modelliert werden.

Kontakte und Literatur sind gegeben: Mathematik Lehren 140, 2007, Thema Hausaufgaben. Zu erweitern ist das Lernszenario 'Prüfungsvorbereitung' so dass es nicht nur zu bestimmten Themen/Konzepten adaptiert wird, sondern auch bestimmte Typen von Aufgaben/Kompetenzen vorbereitet sowie auf eine Menge von potentiellen Prüfungsfragen/Aufgaben zugeschnitten werden kann. Dieses Szenario soll in einer Klasse getestet werden gegen das derzeitige nur auf Konzepte adaptierbare Prüfungsszenario.

Gender Related Adaptation

Supervisor: Erica Melis

AJAX-powered mathematical text-input

Supervisor: Paul Libbrecht

The objective is to realize a non-obstrusive text-input tool for semantic mathematical formulæ producing XML output in OpenMath or content-mathml with the following characteristics:

web-based, using a simple text-field sturdy to parse incompletely input formulae presenting asymchronously a 2d-rendering of the formula (using, e.g., ActiveMath's presentation from OpenMath) able to report errors and bring user to error location in input able to let user select a location or fragment in the rendering triggering a selection in the text-field, and vice-versa as much as browsers provid access to able to receive drop or paste of XML or URL-to-it and convert it to the input syntax extensible to new input patterns

This Bachelor work is stimulated by our experimental results which indicates that the easy access of plain-text-input goes beyond the confort of use of input-palettes.

We expect this work to yield an open-source product licensed with a non-viral license (for example under the Apache Public License) and expect demand at cooperating researchers.

Production and treatment of erroneous math-examples

Supervisor: Erica Melis

Exercise Generation with verbalization

Supervisor: Erica Melis

The generation of exercises in natural language will use templates for the mathematical content and enhance these by ontologies of several application domains (so that mathematically same exercises can b stated, e.g., for medical situations, technical situations, demographic situations etc). This is the basis for NL macro-planning and will be combined with a NL-micro planner that returns the actual natural language formulation of the exercise.

Added Value Mathematical Formulae Presentation

Supervisor: Paul Libbrecht

ActiveMath presents mathematical content to a browser from a semantically encoded content. Among others, this allows the mathematical formulae to be presented with tooltips, click-to-explain-the-symbol, or copy-and-paste, all from the simple fact of enconding semantically the formulae (using OQMath, OMDoc, and OpenMath). We wish to extend this "value-added presentation of formulae'' by considering supplementary input of the author. An example would be annotations about the variables in a well-defined context (e.g. "m is a the mass of the ball"), another avenue might be the usage of type-information about the symbols. The bachelor work should include the enrichment of a 10-15 pages "book" in order to convince mathematicians of the well-foundedness of the approach.

Semantic conversion of presentation mathematical documents

Supervisor: Paul Libbrecht

The goal is to use existing tools such as WebEQ, HELM's and manyothers to provide the automated conversion of presentation-encoded mathematical documents, typically easily convertible to MathML- presentation or TeX, to the semantic format of OMDoc, including OpenMath for mathematical expressions.

The breadth and generalizability of this conversion defines the grade:

In the case of a master's thesis, it would be expected that most of the wikipedia content is convertible. Alternatively, thoroughcoverage of thematic segments of the abstracts of Zentralblatt Mathshould be convertible (the latter being the biggest and oldest mathematical research litterature review service). n the case of a bachelor thesis, a subset of this is expected, for example a thematic part of wikipedia.