by Achilles Dougalis and Dimitris Plexousakis (ICS-FORTH)
In science fiction, artificial agents are portrayed as being capable of interacting with and helping humans. This aid could take the form of holding intelligent conversations and even acting as teachers and coaches. Some progress has been made in this direction in real life. Indeed, systems utilising intelligent agents, such as duolingo™, have proven capable of acting as personal tutors. These “intelligent tutoring systems” (ITS) emulate a human tutor by using AI techniques to adapt instructions and teaching according to each individual learner’s background and progress but also guide the learner through an exercise by providing hints and feedback.
These systems are results of a combination of multiple disciplines, such as computer science, cognitive psychology, human-robot interaction and educational research. Some advantages of ITS are that they are location-independent, easily accessible, and offer great flexibility, allowing students to learn at their own pace and not to have to rely on rigid classroom schedules.
Researchers have found that a learning session can be improved if the teacher is empathetic to the emotions of the learner. Such improvements could take the form of an ITS offering help when it detects that the learner is confused, or by offering the learner motivating remarks when it detects boredom. ITS systems that make use of emotions are known as affective tutoring systems (ATS). ATS combine tutoring strategies and emotion sensing techniques into a single system. Also, evaluations have shown that they can contribute positively to the user’s learning experience [1].
A disadvantage of these systems is that they are designed for teaching a specific subject to specific users, making reusability difficult or even impossible. Moreover, the more diverse and complicated a course is, the more difficult it is to manage it. In other words, there is a need for a universal, well-defined structure in order to facilitate the design of ITS. Fortunately, similar problems in different domains have been dealt with successfully using a knowledge representation and reasoning approach. Such approaches use declarative logic and logical formalisms known as action languages in order to formalise the problem and use AI techniques, such as projection and planning, in order to solve it. ITS that make use of this approach already exist and are known as cognitive tutoring systems. However, these systems are always modelled around a specific course or problem, they don’t take any preferences of the user into account and none of them have been used for affective input. In order to tackle the user’s preference problem, researchers have created the field of “adaptive learning systems”. These systems use machine learning techniques in order to adapt the content of a given course according to the preferences of their current user. However, these systems do not offer feedback to the user (emotional or cognitive) as they are mainly concerned about the course’s presentation.
Figure 1: A course as a directed graph. Each subchapter is depicted as a node, with one or more tutorials represented as vectors linking it with other nodes. A chapter can be traversed either by one tutorial (e.g. tutorial8 for chapter1) or by a selection of smaller tutorials.
In order to address these limitations of both fields, we have built AFFLOG, an adaptive cognitive affective tutoring system that uses answer set programming and the event calculus action language [2] in order to represent the various components and actions of an ATS, and perform reasoning tasks such as planning for creating a course suitable to the current user, as well as offering the tools to create a new course from scratch. The main contributions of this work are:
- Course representation: We represent a course and its chapters in such a way that it can be effectively used by our system while also providing the author of the course with flexibility in terms of which modalities to use and also the design, structure and size of the course.
- User representation: Which parts of a specific course the user has learned, his/her preferred learning styles, as well as his/her emotional state regarding a course or a part of it.
- The integration of emotions and the user’s learning style in the tutor’s decision making process (Figure 2).
Figure 2: Emotional Strategies according to the user’s current emotion.
We use the term “tutorial” to describe the main building block for a course. A tutorial can be plain text, a picture, an audio file or a video file. Important semantic properties of a tutorial written in ASP include the subchapters that denote its beginning and end, as well as its modality, relative difficulty, its duration, and how suitable it is for the current user according to his/her learning style. A “test” is the method the system uses in order to understand whether the user understood a part of the course.
A “course” is the material that the tutor uses to teach in one or more “sessions”. It is essentially a collection of tutorials and tests that are presented to the user. Each course consists of a number of “chapters” and each chapter consists of a number of “subchapters”. Subchapters are traversed using tutorials and can be described as nodes in a graph where the tutorials act as the vectors (Figure 2). Once the tutorials that comprise a chapter have been presented to the user, the tutor presents him/her with a test to determine whether the user understood the chapter. If the user passes the test, the tutor assumes that the user understood the chapter and proceeds to the next chapter. If not, the tutor assumes that the user has not understood one or more of the tutorials of the particular chapter, but since it cannot differentiate between which tutorial was misunderstood and which not, all the tutorials of that chapter are labelled as wrong. Consequently, the course must be modified using “planning” by replacing the wrong tutorials with unused tutorials for that chapter. When all the chapters are completed and all the tutorials are correct, the course has been completed successfully.
The work here represents a part of a PhD in emotional adaptive tutoring systems. Future work includes system evaluations, linking the system with the semantic web via the RDF language, and using machine learning methods to offer more personalised tutoring.
References:
[1] B. Kort, R. Reilly, R. W. Picard: “An affective model of interplay between emotions and learning: Reengineering educational pedagogy-building a learning companion, 2001.”
[2] M. Sergot, R. Kowalski: “A logic-based calculus of events”, New Generation Computing 4.1 (1985): 67-95.
Please contact:
Achilles Dougalis
FORTH-ICS, Greece
