Use case & info

2010-03-06T05:17:00.000-08:00

Technologically and computationally the agents are based on Semantic Neural Networks. At the beginning, the end user gets his/her own agent, with which the user interacts. This personal media agent utilizes all the other agents available on the system.

There are two types of agents. First type of agents reads social media services and organizes the information into databases. The first types of agents are pre taught: they cannot learn more. Therefore end user interacts only with second type of agents.

Second type of agents interacts with end users and builds all personalized semantic networks. These personalized semantic networks consist of relations between concepts found from tags, titles and comments. The pieces of contents are connected into one or more concepts in this high level semantic network. These agents can learn by the feedback of evaluations made by the end user. In other words, second types of agents tries to match the high level conceptual structure and user taught conceptual structure. The idea behind the search is relatively close to image recognition with neural networks where images are replaced by concept maps.

The use case, in brief, is following: At first user describes the subject in focus by typing several key concepts (tags maybe) into UI’s definition field, for example ‘eyetracking and heatmaps’. Secondly, the first types of agents start to search all possible content related to keywords. Currently search is limited to YouTube, Flickr and Slideshare but new media services will be added on the list. The data received from these media services will be indexed and prepared for later use.

According to this raw data, the personal agent (second type) forms a high level semantics related to this task. Additionally to traditional tag cloud, this one is computed by applying cluster analysis in determining the places of the tags. This ensures, that neighbor tags are strongly related to each other. In traditional tag clouds, the tags are in random order. Furthermore, the tags that have strong explanative power are placed into red background.

The agent constructs a rank ordered list about content according to semantic relations (figure 4). The user can evaluate the search results by clicking + or – symbols. After evaluations, the semantics will be re-computed. In general, the computing related to semantics are relatively heavy processes. Therefore the semantics are not necessarily computed after every evaluation: the semantics are re-computed in every 3-5 minutes if possible. Evaluations are used in order to determine irrelevant tags and decrease their importance in semantic and in content rank.

The semantics learned by personal media agent evolves through all assigned tasks: The learned semantic context is always a background for new tasks. This feature can be used to make effective agents for a certain limited search domain. On the other hand, when we switch to completely different search domain, we can receive interesting results.

In figure 1 we have continued the previous tasks by assigning a new task related to ‘contemporary arts and Picasso’. From the visualization we can see that word contemporary is related to words entertainment and modern while arts -tag is related to context of martial arts. Picasso is related to tags like artist and painter. In the visualization, the causes of the previous tasks can be seen, but the red-background area remains the same in most cases.

Figure 1. Visualization about context formed by previous tasks and ‘contemporary arts and Picasso’ -task.

More: www.gameminer.fi

Interview at Checkpoint e-Learning, November 2009.

Semantics

2009-06-23T13:12:00.000-07:00

In our studies we have designed teachable agents that can learn conceptual structures in terms of conceptual learning. These agents are based on authors' previous work, AnimalClass, introduced at Online Educa Berlin 2006. In AnimalClass the learner can teach conceptual structures about mathematics, sciences, languages and arts to virtual pets (teachable agents). The main difference between Teachable Media Agents and AnimalClass is in philosophy of learning. When teachable agents in AnimalClass were taught in terms of inductive learning, the Teachable Media Agents are taught in terms of deductive learning.

Suffering from information overload?

2009-06-23T13:08:00.000-07:00

Social media services, such as YouTube, Flickr and Creative Commons, contain enormous number of content valuable for education. If the requested theme can be effectively searched or recognized, teacher can easily construct the course material from social media sources. However, the search engines are not optimal for educational purposes: Search engines can list numerous pieces of content that matches more or less perfectly to keywords. After search there are thousands of pieces of content to check manually if they really fit to the requested subject.