In part seven of this series I talked about how empathy for the students plays a key role in the philosophy behind this application. By that, I primarily meant that the user must continually monitoring the students’ actions and behaviors in order to try and estimate why, or why not, they might be paying attention at any given time. In this post I will describe exactly how the the virtual students’ attention system functions. I will also describe how the user of this simulation can, in real time, both monitor and effect those levels.
In the previous updates I mentioned the behaviors students might display when their attention levels are low. I also mentioned the pencil meter at the top right of the screen that would be a graphic display of attention. What I have not talked about yet are the actual numbers and logic I am using to keep track of, and change, attention over time. For an attention system to work we first needed to make some reasonable assumptions about what might effect a student’s attention in real life. On top of those assumptions, we also have to consider what kind of interventions methods, or actions, the user might have at their disposal to effect or change a student’s attention. These actions are dependent upon what kind of input data we can capture using both the Kinect and Speech Recognition software. With all that in mind, we initially came up with three reasonable assumptions. Each assumption is paired with a primary method of interaction to control the attention levels of the students.
The first reasonable assumption we made was that students who a have an unobstructed view of the teacher are likely to have a higher attention level than those who do not. The interaction method, the user is provided with, is the ability to move front to back and side to side in front of the Kinect. As they do this, they will need to take notice of which students they can see and which students can see them.
The second reasonable assumption we made was that students who are in close proximity to the teacher are likely to have a higher attention level than those who are far away. The interaction method provide for this is the same as mentioned before. As the user moves around in front of the Kinect, changing their position in the virtual space, they will need to monitor the behaviors of not only the students they are near but also of those who are further away.
The third reasonable assumption we made was that when a teacher talks directly to an individual student, that student’s attention level will be higher than if the teacher was addressing the entire class at once. The interaction method here gives the user the ability to, at anytime, use their voice to call on a student by name. When they do so, they will see an immediate rise in attention and change in behavior.
Now let’s talk a little about the numbers and functions that are used in response to the user’s interactions. First, I want to mention, that I was careful to build these functions with flexibility. At some point in our application all three interactions will be active but in other stages maybe only one will be active. This will depend on what concept we might want to drive home for the user. Also, the effect of one interaction, on a student’s attention level, might vary depending on the particular situation and on the particular student. What that means is there are no hard numbers as to how much any one user interaction effects attention levels. However, a reasonable example of how these numbers might break down during a regular teaching situation,when all three interactions are in play, is as follows.
When a user and student have an established a clear line of sight with each other, the student will reach a minimum attention level of 25%. As the physical distance between the user and a student decreases, the student will receive some relative portion of an additional 50% of attention. This means, if the user is very near a student, and they can see each other, the student’s attention level could be as high as 75%. Should a user call on an individual student, and talk to them directly, that student’s attention will be set to 90% regardless of the user’s physical location in the room.
Using that logic, the user now has the ability to interact with the application and alter student attention levels. However, you might notice, that in this example, the user has no way of raising the attention levels above 90%. This is because I am reserving that last 10% of attention for an additional type of interaction, which is word choice.
Now, to follow the process of what we have done so far, before we can justify another type of interaction we should first make sure it fits our two established criteria. One, it can be done using the Kinect and Speech Recognition and two, it has some reasonable assumption tied to it as to being able to effect student’s attention in real life. Obviously, saying words fits in with Speech Recognition, but how does the choice of words effect attention?
A fourth reasonable assumption we made is that by choosing to use kind and respectful words while talking to students, attention levels would be higher than if those words were left out. Conversely, one could make the argument that by choosing to use not so kind and harsh words, student attention levels might also be high. However, this would probably be the wrong message to pass onto future teachers. So, in this case, we are choosing to recognize and reward preferable behavior. Let’s see how this new interaction works in our example.
When a user calls on a particular student, e.g. “Bridget,” Bridget’s attention level is set to 90%. However, with our new interaction in place, that allows the user to get an additional 10% of attention just by choosing to use kind and or respectful words, the user could say something like “Good Morning Bridget,” to set Bridget’s attention level at 100%. The first 90% comes from saying “Bridget” and the second 10% comes from saying “Good Morning” while addressing her. There are several other contextual examples of this interaction included in the program. Words like “please” and “thank you” can always be used. When the user chooses to use these words, in addition to the attention bonus, they will often see a physical reaction from the students in the form of a smile.
Everything we have talked about so far refers to the abilities the user has to raise attention levels. What we also need in order for the attention system to be complete, and function more like real life, are reasons and logic for, decline in attention levels.
Obviously if the user neglects to use any of the four interactions I have described so far, attention levels will never rise as high as they could be. Even if the user does use the interactions it would not be realistic for attention to maintain an achieved level indefinitely. Here is where we must introduce a time factor into our attention system and not just a simple decline of attention over time. Instead, each student is equipped with several attributes that dictate how their attention level fluctuates. These attributes provide the needed variety and the necessary realism for how a unique individual might start to become inattentive.
The first attribute each student has is an attention span. This is how long a student will maintain attention before starting to decline. The attention span can be set to something different for each individual however, currently the students in the back row are set to have a 30 second attention span while the students in the front row have a 60 second attention span. Again, we are making some reasonable assumptions here. Assumptions that are based on interactions the program is providing to the user. The user can clearly see there are two rows of students, and it is reasonable to assume that those in the back will be quicker to fall out of attention than those in the front. At this time, the program does not provide any deeply personal information or phycological profiles that a might indicate a reason that a student has a short attention span. We must keep our assumptions and interactions closely alined to keep our users engaged in the simulation.
Each student also has a true/false attribute that indicates if the student is currently paying attention. This attribute is set to true whenever a student rises above 50% attention level and false when they are below 50%. The user is provided a physical indicator of this when a student stops randomly looking around the room and begins to look at them. The user will probably quickly realize that just reaching 50% attention is not good enough. In fact, just because a student reaches this attention 50% tipping point, does not necessarily mean their attention span kicks in. The assumption being, the interaction the user did to raise the student’s attention may not have been significant enough to hold their attention.
To deal with this, each interaction the user makes passes a “significance” parameter to the student involved. If the interaction is defined as significant it will trigger the student’s attention span, and whatever attention level was reached as a result of the interaction will hold for the duration of the span before any decline. This is done to add extra variety and realism to the simulation. It is also done to place emphasis on certain interactions. The assumption here is while just walking by a student, being in close proximity for a moment (not significant), might spike their attention level, as soon as I walk away that attention will likely be lost. As opposed to, if I call directly on a student (significant), their attention level will most likely remain high for a period of time.
When an attention span expires, or when a student is defined as not paying attention, their attention must fall into a state of decline. To do this, each student has a unique rate of decline attribute. Sticking with the same assumptions we made before, the back row students have a generally faster rate of decline than front row students. For additional bit of assumed realism I have made the rate of decline happen on an ease out curve instead of a linear decline. This means when attention first starts to decline it declines slower. As decline continues uninterrupted over time it declines faster and faster until eventually reaching zero. The curve of decline is also adjustable on a per student basis.
I think that about covers the thought process and logic behind the student attention system. It really is the backbone of the whole simulation because it gives the user methods of interaction to deal with realistic, but simulated, attention issues. While it certainly doesn’t cover every issue that might cause attention fluctuation, hopefully it covers enough that the user gets the point.
One last thing I would like to mention, and I will talk a bit more about this in the next post, is how the attention levels of the students can be used in assessment of the user’s success with the simulation. During a pure teaching stage in the simulation, a module where the user must get through some sort of lesson content, each student will be keeping an average attention level. When the user finishes delivering the content, hopefully while using the interactions mentioned in this article, they will be able to administer a quiz with questions pertaining to the lesson. Each student’s score on the quiz will be based on what their average attention level was during the lesson.
This direct relationship between attention level and quiz score is based on another reasonable assumption. It is also based on one of the technical requirements of the Speech Recognition software. One could argue that it might be more accurate to measure the attention level of a student at the moment during the lesson when the user speaks an answer to an eventual quiz question. If the attention was high at that moment, we could then reasonably assume the student heard, and now knows, that information and therefore could answer the quiz question correctly. The difficulty with this method comes with how Speech Recognition works. In order for Speech Recognition to work it must identify a recognized phrase. While you can build in some flexibility to the exact wording of the phrase you can not separate out a phrase from a larger body of spoken words. This means recognized phrases need to be spoken in the clear, with a short pause before and after. While this works well for short, command like, phrases it does not work well during a lengthy speech like delivering the content of a lesson. For these reasons, the most reasonable way for us to judge if a student has learned the content of a lesson is to keep track of their average attention during the lecture and relate that directly back to how well they score on a quiz. The reasonable assumption being, higher attention levels lead to higher quiz scores.
Leave a Reply
You must be logged in to post a comment.