TRANSKRIBEERI 🔗
WEBVTT
00:00:08.300 --> 00:00:11.339
Opening AI for Language Learning explores artificial
00:00:11.339 --> 00:00:14.240
intelligence through informal conversations between
00:00:14.240 --> 00:00:16.500
two longtime colleagues in language education
00:00:16.500 --> 00:00:20.679
and digital technologies, Matt Schultze and Phil
00:00:20.679 --> 00:00:25.120
Hubbard. Language educators, learners, and enthusiasts
00:00:25.120 --> 00:00:27.920
are invited to an informed discussion where emerging
00:00:27.920 --> 00:00:32.560
technology meets human communication. Well, hello,
00:00:32.679 --> 00:00:35.399
Matt. Hello, listeners. So what are we doing
00:00:35.399 --> 00:00:43.630
today? Picking a topic that we've actually, that
00:00:43.630 --> 00:00:47.289
has come up before in somewhere or another, we're
00:00:47.289 --> 00:00:51.420
asking the questions, can chatbots tutor? Well,
00:00:51.460 --> 00:00:53.899
that's a simple question, but I'm guessing the
00:00:53.899 --> 00:01:00.439
answer is not so simple. Well, as we'd like to
00:01:00.439 --> 00:01:04.659
say, as academics like to say, it depends. Looking
00:01:04.659 --> 00:01:08.000
at the context of computer -assisted language
00:01:08.000 --> 00:01:11.019
learning, where this idea of tutor It's like
00:01:11.019 --> 00:01:14.500
Santa has come up at the very beginning, literally
00:01:14.500 --> 00:01:17.019
at the time of mainframe computers. There was
00:01:17.019 --> 00:01:20.319
a kind of big project, Plato, all the software
00:01:20.319 --> 00:01:22.280
they had also for language learning. They had
00:01:22.280 --> 00:01:26.319
many different subject areas in Plato, that old
00:01:26.319 --> 00:01:29.200
system. This was basically what we would call
00:01:29.200 --> 00:01:32.299
today kind of drill and kill exercises, very
00:01:32.299 --> 00:01:36.900
tutorial, very sequenced, very constrained, often.
00:01:37.719 --> 00:01:41.939
grammar exercises, vocabulary exercises. That
00:01:41.939 --> 00:01:44.599
was kind of the beginning of CALDE. I also gave
00:01:44.599 --> 00:01:46.840
computer -assisted language learning CALDE a
00:01:46.840 --> 00:01:50.859
name. We then have a fairly influential book
00:01:50.859 --> 00:01:54.799
by Mike Levy on computer -assisted language learning.
00:01:55.000 --> 00:01:57.719
To my knowledge, it was the first noticeable,
00:01:57.900 --> 00:02:01.140
at least, monograph on computer -assisted language
00:02:01.140 --> 00:02:06.480
learning in 97. So not that long ago for people
00:02:06.480 --> 00:02:10.400
like me. And he talked about the tool versus
00:02:10.400 --> 00:02:14.099
tutor distinction, looking at hardware and software,
00:02:14.280 --> 00:02:19.919
desktop computers at the time, right? Was the
00:02:19.919 --> 00:02:22.759
software acting as a tool? Was it acting as a
00:02:22.759 --> 00:02:29.180
tutor? And if we compare this to today's AI tools,
00:02:29.319 --> 00:02:32.400
that's the word I often use in order not to say.
00:02:32.810
Load full transcript
--> 00:02:36.090
Things like DAI, which I don't like to say, because
00:02:36.090 --> 00:02:38.849
I think it also is not the right metaphor very
00:02:38.849 --> 00:02:43.520
often. It's too broad. And he looked at the tutor
00:02:43.520 --> 00:02:46.120
and the main thing he said about the tutor was
00:02:46.120 --> 00:02:49.120
that evaluation plays a role there. So whenever
00:02:49.120 --> 00:02:53.039
the software also evaluates the learner response,
00:02:53.460 --> 00:02:55.840
then he called it a tutor. That was his criterion.
00:02:56.120 --> 00:02:58.879
And you, Phil, and you can talk better about
00:02:58.879 --> 00:03:01.520
it than I can, together with Claire Braden -Siskin
00:03:01.520 --> 00:03:05.800
wrote a paper in 2004 about... tutorial call
00:03:05.800 --> 00:03:11.280
at the time, 2004, when most of the research
00:03:11.280 --> 00:03:14.979
and development and use practice of computer
00:03:14.979 --> 00:03:17.439
assisted language learning was actually in computer
00:03:17.439 --> 00:03:19.919
mediated communication. I was kind of almost
00:03:19.919 --> 00:03:22.439
a paper that was saying, no, there's something
00:03:22.439 --> 00:03:28.240
else that's also going on. I wrote about the
00:03:28.240 --> 00:03:37.599
tutorial call in 2015 together with through the
00:03:37.599 --> 00:03:40.400
hive and then a little bit later on fairly recently
00:03:40.400 --> 00:03:44.960
actually came out in 25 in the context of AI
00:03:44.960 --> 00:03:47.800
basically looking at tutorial call and there
00:03:47.800 --> 00:03:50.740
was important to me to make the distinction of
00:03:50.740 --> 00:03:54.039
interaction with the computer which I see for
00:03:54.039 --> 00:04:01.099
tutorial call and via the computer in computer
00:04:01.099 --> 00:04:04.680
mediated communication. Now, the chatbots kind
00:04:04.680 --> 00:04:07.620
of make this a bit of a problem and we were revisiting
00:04:07.620 --> 00:04:09.979
the question also in some way because I think
00:04:09.979 --> 00:04:14.159
it's important also for teachers. Why? Because
00:04:14.159 --> 00:04:16.699
we're dealing with chatbots and that feels a
00:04:16.699 --> 00:04:19.139
little bit like computer mediated communication,
00:04:19.620 --> 00:04:23.600
but it surely is in a way interaction with a
00:04:23.600 --> 00:04:26.980
computer. There is no human. at the other end,
00:04:27.000 --> 00:04:29.019
basically, when you have a conversation with
00:04:29.019 --> 00:04:33.259
a chatbot. So let's see how we can disentangle
00:04:33.259 --> 00:04:36.680
this. And I think this will have a lot to do
00:04:36.680 --> 00:04:38.459
with teaching presence and things like that.
00:04:39.959 --> 00:04:44.000
Yeah, well, I'll try to first project myself
00:04:44.000 --> 00:04:47.720
back into 2003, 2004, when Claire and I were
00:04:47.720 --> 00:04:52.079
working on this. We were really reacting. Exactly
00:04:52.079 --> 00:04:55.120
as you said to this idea that CMC had come to
00:04:55.120 --> 00:04:57.660
dominate, especially in journal publications
00:04:57.660 --> 00:05:01.180
and books and conference presentations. A lot
00:05:01.180 --> 00:05:04.639
of that done at the university level, but software
00:05:04.639 --> 00:05:07.300
that was instructional software for language
00:05:07.300 --> 00:05:11.079
was still very widespread. We were in the multimedia
00:05:11.079 --> 00:05:15.500
era from, say, the mid -90s on, in particular,
00:05:15.500 --> 00:05:18.860
with CD -ROMs and school systems were buying
00:05:18.860 --> 00:05:21.910
copies of these, especially outside. the U .S.
00:05:21.930 --> 00:05:24.790
for English as a second language because it would
00:05:24.790 --> 00:05:28.810
be bringing in voice and graphics and so on,
00:05:28.870 --> 00:05:34.759
even video on occasion, bringing English to areas
00:05:34.759 --> 00:05:37.699
where maybe the teacher didn't have the level
00:05:37.699 --> 00:05:40.259
of proficiency that the students would like to
00:05:40.259 --> 00:05:44.860
see. So what we called tutorial call was still
00:05:44.860 --> 00:05:47.279
quite alive and well out in the world. It just
00:05:47.279 --> 00:05:56.740
wasn't as much in academia. So we took Mike Levy's
00:05:56.740 --> 00:06:00.079
tool tutor distinction and thought very highly
00:06:00.079 --> 00:06:02.959
of it, but we didn't feel that the tutor part
00:06:02.959 --> 00:06:05.879
went far enough because there's a lot more to
00:06:05.879 --> 00:06:08.540
instruction than evaluation. Oh, that's very
00:06:08.540 --> 00:06:13.019
true. And so that's why we came up with this
00:06:13.019 --> 00:06:19.480
concept of a teaching presence. I think I'll
00:06:19.480 --> 00:06:21.959
try and even read a little bit of a quote from
00:06:21.959 --> 00:06:26.519
our paper. So we said, tutorial call refers to
00:06:26.519 --> 00:06:29.779
the implementation of computer programs, disk,
00:06:29.959 --> 00:06:32.800
CD -ROM, web -based, et cetera, that include
00:06:32.800 --> 00:06:36.660
an identifiable teaching presence. And that means
00:06:36.660 --> 00:06:41.199
specifically for improving some aspect of the
00:06:41.199 --> 00:06:46.399
learner's proficiency. So it's aimed at, in a
00:06:46.399 --> 00:06:50.170
sense, at instruction and education. Not just
00:06:50.170 --> 00:06:52.389
evaluation, though evaluation is an important
00:06:52.389 --> 00:06:58.990
part. So to us, the point was there's a design
00:06:58.990 --> 00:07:02.910
decision made in this software. And that's back
00:07:02.910 --> 00:07:05.889
then that was always made by a human. So it's
00:07:05.889 --> 00:07:09.470
the designer who puts this stuff in and in a
00:07:09.470 --> 00:07:12.930
way projects their own ideas about teaching and
00:07:12.930 --> 00:07:17.350
learning into that software for the student to
00:07:17.350 --> 00:07:22.009
interact with then. across time and space, unlike
00:07:22.079 --> 00:07:28.870
the face -to -face classroom. So we also talked
00:07:28.870 --> 00:07:32.910
about this notion of teacher presence. And in
00:07:32.910 --> 00:07:36.189
that case, there had to be some identifiable,
00:07:36.350 --> 00:07:40.649
either through dialogue or often through some
00:07:40.649 --> 00:07:44.250
sort of avatar, someone representing a teacher
00:07:44.250 --> 00:07:47.610
and not just the instructional material itself.
00:07:47.670 --> 00:07:50.379
So think about... difference between a textbook
00:07:50.379 --> 00:07:54.660
and a teacher. So a teacher presence, an example
00:07:54.660 --> 00:08:00.500
of that, Van Stevens and I wrote a CD -ROM program.
00:08:00.720 --> 00:08:03.800
We didn't do the programming of it, but we wrote
00:08:03.800 --> 00:08:08.160
the script for it. And it was called Tracy Talk,
00:08:08.399 --> 00:08:11.879
T -R -A -C -I. And it was basically a mystery
00:08:11.879 --> 00:08:14.959
that students would solve through voice interaction.
00:08:15.139 --> 00:08:19.279
But Tracy stood for teacher ranging across the
00:08:19.279 --> 00:08:24.959
computer interface. She had her own little character.
00:08:25.160 --> 00:08:29.800
So whenever you called up help or wanted a hint
00:08:29.800 --> 00:08:32.539
or something, it was her voice that came through
00:08:32.539 --> 00:08:34.519
and she had her own little avatar that showed
00:08:34.519 --> 00:08:37.240
up sometimes too. So that's a teacher presence.
00:08:37.399 --> 00:08:40.399
And we see that a lot. We used to see it even
00:08:40.399 --> 00:08:43.460
in the old days in kids programs, there would
00:08:43.460 --> 00:08:47.240
almost always be some sort of Whether it was
00:08:47.240 --> 00:08:50.360
Kermit the Frog or Mickey Mouse or some character
00:08:50.360 --> 00:08:53.279
designed for it, there would be something playing
00:08:53.279 --> 00:08:57.659
that teacher role. So because tutorial programs
00:08:57.659 --> 00:09:01.980
required programming skills, some of it was done
00:09:01.980 --> 00:09:04.720
by teachers with that or by teachers who knew
00:09:04.720 --> 00:09:07.340
someone who could program and help them. But
00:09:07.340 --> 00:09:10.399
an awful lot was done by corporations. And I
00:09:10.399 --> 00:09:13.139
think that's why it fell out of favor to a large
00:09:13.139 --> 00:09:18.750
extent with... with academia. Carol Chappell
00:09:18.750 --> 00:09:23.769
and Joan Jameson did some work with an English
00:09:23.769 --> 00:09:27.389
multimedia program that was very widespread at
00:09:27.389 --> 00:09:31.590
the time. And so you can see how they interacted
00:09:31.820 --> 00:09:35.799
with this idea that tutorial was still something
00:09:35.799 --> 00:09:40.279
viable. But as you say, a lot of others just
00:09:40.279 --> 00:09:43.940
did computer -mediated communication. And over
00:09:43.940 --> 00:09:51.250
time, And as Rosetta Stone became less popular,
00:09:51.509 --> 00:09:56.649
I would say, and people started noticing and
00:09:56.649 --> 00:10:00.129
being quite
aware of what the limitations of
00:10:00.129 --> 00:10:03.509
this type of material was and the expense involved,
00:10:03.889 --> 00:10:08.549
it began to, I think, to fall from favor even
00:10:08.549 --> 00:10:11.049
more. And then suddenly we had mobile learning
00:10:11.049 --> 00:10:16.320
and apps. And now... Even people with somewhat
00:10:16.320 --> 00:10:18.720
limited programming skills could often put together
00:10:18.720 --> 00:10:21.539
an app that would do something like particularly
00:10:21.539 --> 00:10:25.759
pronunciation and vocabulary work. And now apps
00:10:25.759 --> 00:10:28.360
are everywhere and they have a teaching presence.
00:10:30.820 --> 00:10:37.899
Right. It's interesting. The intersections of
00:10:37.899 --> 00:10:43.059
tutorial call and AI, that was what drew me to
00:10:43.059 --> 00:10:48.220
the area. uh like talking about projects of of
00:10:48.220 --> 00:10:52.340
yesteryear there also is the asina project which
00:10:52.340 --> 00:10:54.980
involved natural language processing was done
00:10:54.980 --> 00:10:58.639
at mit but there were also applied linguists
00:10:58.639 --> 00:11:00.980
on board at least for a short while like claire
00:11:00.980 --> 00:11:03.980
crunch who was involved in that project and i
00:11:03.980 --> 00:11:07.940
can only imagine that Claire Gramsch would have
00:11:07.940 --> 00:11:12.480
advocated for sound applied linguistics approaches
00:11:12.480 --> 00:11:16.659
in there. She's not exactly a computer person,
00:11:16.820 --> 00:11:24.460
basically, in that way. Because at the same time,
00:11:24.480 --> 00:11:28.500
this was also the time of the community of language
00:11:28.500 --> 00:11:32.279
teaching approach, basically, and that also kind
00:11:32.279 --> 00:11:37.289
of... put a damper on this tutorial call. Going
00:11:37.289 --> 00:11:41.789
back to AI, we now have GenAI, and it brought
00:11:41.789 --> 00:11:46.110
to the fore this communicative interaction. I
00:11:46.110 --> 00:11:47.750
mean, we're dealing with chatbots on the one
00:11:47.750 --> 00:11:50.610
hand, and on the other hand, and this is where
00:11:50.610 --> 00:11:53.870
the contradictions seem to start, a clear interaction
00:11:53.870 --> 00:11:58.850
with the computer. What I find interesting, and
00:11:58.850 --> 00:12:01.429
that seems to be something that a lot of people
00:12:01.429 --> 00:12:05.429
are not aware of. Within AI in general, there
00:12:05.429 --> 00:12:10.009
is one branch of artificial intelligence research
00:12:10.009 --> 00:12:13.070
and development that is literally called intelligent
00:12:13.070 --> 00:12:16.950
tutoring systems. And there are intelligent language
00:12:16.950 --> 00:12:20.429
tutoring systems that were written over the years
00:12:20.429 --> 00:12:23.809
to very success. None of them ever entered the
00:12:23.809 --> 00:12:27.860
real mainstream. Very few were used practically
00:12:27.860 --> 00:12:30.940
as in with students, usually at a university
00:12:30.940 --> 00:12:33.740
and not at a high school. There were much fewer.
00:12:33.840 --> 00:12:37.299
So this was more an area of research where this
00:12:37.299 --> 00:12:41.639
was alive in the small fringe group of computer
00:12:41.639 --> 00:12:44.139
assistant language learning that I belonged to
00:12:44.139 --> 00:12:49.700
for quite a while. Well, I think maybe we'll...
00:12:50.259 --> 00:12:52.960
Get back to intelligent tutoring systems in a
00:12:52.960 --> 00:12:55.200
moment, but I'd like to return to the question
00:12:55.200 --> 00:12:59.200
that we started this particular podcast with,
00:12:59.279 --> 00:13:04.759
and that is, can chatbots tutor? And if you look
00:13:04.759 --> 00:13:07.080
at it, there seems to be kind of a continuum
00:13:07.080 --> 00:13:14.019
from zero to a lot if you take this notion of
00:13:14.019 --> 00:13:16.899
teaching presence, which I think is still a valid
00:13:16.899 --> 00:13:20.539
one to apply here. So there are conversation
00:13:20.539 --> 00:13:25.940
partner tools out there. And of course, you can
00:13:25.940 --> 00:13:29.039
prompt yourself to have a conversation partner
00:13:29.039 --> 00:13:35.679
chat GPT or any of the others. And if it's just
00:13:35.679 --> 00:13:38.419
a conversation partner, that's kind of zero teaching
00:13:38.419 --> 00:13:42.460
presence. But you or someone else can also prompt
00:13:42.460 --> 00:13:45.720
a conversational interaction that could have,
00:13:45.799 --> 00:13:49.879
for example, recasting. built into it so that
00:13:49.879 --> 00:13:54.860
if you say something that's not whatever the
00:13:54.860 --> 00:13:59.960
AI believes is, there I go with AI believe, but
00:13:59.960 --> 00:14:06.919
the AI determines is not what the target language
00:14:06.919 --> 00:14:10.799
form should be, then it can, in its next round
00:14:10.799 --> 00:14:13.559
of that conversation, put in the correct form
00:14:13.559 --> 00:14:19.590
without explicit correction. You can have a conversation
00:14:19.590 --> 00:14:22.509
and then get some feedback at the end. You can
00:14:22.509 --> 00:14:26.190
prompt for that. I've done it with asking, for
00:14:26.190 --> 00:14:28.950
example, for a word list at the end of what was
00:14:28.950 --> 00:14:31.889
in the conversation. You can have a conversation
00:14:31.889 --> 00:14:35.870
partner that has ongoing explicit corrections
00:14:35.870 --> 00:14:39.980
and feedback. And you can have an explicit tutor
00:14:39.980 --> 00:14:45.600
role. We talked with the CEO of Speakology when
00:14:45.600 --> 00:14:50.690
we were at Calico. And Speakology has a tutor
00:14:50.690 --> 00:14:54.629
avatar. In fact, you can choose among a collection
00:14:54.629 --> 00:14:57.929
of tutor avatars that there is a visible, animated,
00:14:58.190 --> 00:15:03.289
human -like presence taking that teacher presence
00:15:03.289 --> 00:15:08.169
or a tutor presence. A quick example of this
00:15:08.169 -->
00:15:10.470
Google Translate practice mode that I've talked
00:15:10.470 --> 00:15:18.980
about before, it has a sequence. where either
00:15:18.980 --> 00:15:23.320
you can pick one of the topics it wants to talk
00:15:23.320 --> 00:15:25.679
about or you can present your own, but then it
00:15:25.679 --> 00:15:30.360
generates a conversation that you first, well,
00:15:30.440 --> 00:15:32.679
first you look and see that there's a word list.
00:15:33.200 --> 00:15:36.980
So that's some pre -learning. Then you listen
00:15:36.980 --> 00:15:40.179
to a conversation and you are quizzed on the
00:15:40.179 --> 00:15:47.399
content of that. Then you do a simulated conversation
00:15:47.399 --> 00:15:52.850
yourself with the chat bot and then you get feedback
00:15:52.850 --> 00:15:56.840
at the end on how your chat went. So I did one
00:15:56.840 --> 00:16:00.940
in Spanish recently and the feedback I got at
00:16:00.940 --> 00:16:04.980
the end, it was fairly limited because I did
00:16:04.980 --> 00:16:08.899
okay. But it did say I used a particular phrase
00:16:08.899 --> 00:16:14.090
and more commonly I should use this other phrase.
00:16:14.250 --> 00:16:20.059
So this comes down to... Is that tutoring? It
00:16:20.059 --> 00:16:23.769
felt like tutoring. So this is one of those eye
00:16:23.769 --> 00:16:27.789
of the beholder things. Right. Yeah. It's often
00:16:27.789 --> 00:16:31.850
what it is that you want as the learner. What
00:16:31.850 --> 00:16:37.289
is it you need as a learner? Talking about these
00:16:37.289 --> 00:16:40.649
recalls, this was one of the things that worked
00:16:40.649 --> 00:16:44.629
quite early on. kind of the early models that
00:16:44.629 --> 00:16:51.330
were available in 23. I remember trying out question
00:16:51.330 --> 00:16:55.269
prompts, basically, that had deliberate spelling
00:16:55.269 --> 00:16:57.730
errors and other grammatical errors in there,
00:16:57.850 --> 00:17:00.929
like asking what's the capital of Germany, really
00:17:00.929 --> 00:17:03.149
kind of beginner stuff in a way, also contents
00:17:03.149 --> 00:17:07.970
-wise. But jumbling that sentence, misspelling
00:17:07.970 --> 00:17:11.359
something or whatever, not only would you get
00:17:11.359 --> 00:17:14.309
the answer, In some way in this conversation,
00:17:14.609 --> 00:17:18.670
but because this chatbot is such a sycophant,
00:17:18.750 --> 00:17:20.950
of course, it also tries to get it right. So
00:17:20.950 --> 00:17:24.910
in the answer, just looking at spelling errors,
00:17:25.069 --> 00:17:27.109
it would have been spelled correctly. So if I
00:17:27.109 --> 00:17:30.549
spell capital with an E at the end, basically,
00:17:30.549 --> 00:17:34.660
or something like that, it would not do that
00:17:34.660 --> 00:17:37.500
when answering. So if I notice those things as
00:17:37.500 --> 00:17:39.829
a learner. then I can learn something and then
00:17:39.829 --> 00:17:46.450
it becomes a tutoring, a teaching presence, basically.
00:17:47.809 --> 00:17:50.789
We're always talking about Gen AI. When we talked
00:17:50.789 --> 00:17:53.690
about SAPD, we talked about Gen AI. I gave a
00:17:53.690 --> 00:18:00.549
presentation at the Eurocall conference in September
00:18:00.549 --> 00:18:03.890
and also kind of made a stance and saying it's
00:18:03.890 --> 00:18:07.390
problematic. to call these chatbots tutors, basically.
00:18:07.569 --> 00:18:10.150
I was asked a question from the floor, which
00:18:10.150 --> 00:18:12.250
basically suggested, well, but they sometimes
00:18:12.250 --> 00:18:14.769
can be tutors, a little bit like you laid out
00:18:14.769 --> 00:18:19.259
that spectrum. My main point of departure has
00:18:19.259 --> 00:18:23.559
always been the kind of wider concept of AI with
00:18:23.559 --> 00:18:26.440
the intelligent tutoring systems that I already
00:18:26.440 --> 00:18:31.559
mentioned. So basically, of course, AI can be
00:18:31.559 --> 00:18:34.839
a tutor. And really shorthand metaphors there.
00:18:35.319 --> 00:18:38.079
It's a little more challenging, as we are trying
00:18:38.079 --> 00:18:41.900
to tease out in our conversation today, in Gen
00:18:41.900 --> 00:18:44.599
AI. It's a kind of what role does the chatbot
00:18:44.599 --> 00:18:47.599
actually play? How much does this kind of go
00:18:47.599 --> 00:18:51.960
into tutoring, basically? Yeah. And as you point
00:18:51.960 --> 00:18:55.619
out, AI has been around a lot longer in the form
00:18:55.619 --> 00:18:59.910
of intelligent tutoring systems. Maybe you could
00:18:59.910 --> 00:19:02.349
tell us a little bit more about that. What are
00:19:02.349 --> 00:19:06.869
the parts of an ITS? That's an interesting thing.
00:19:06.930 --> 00:19:10.410
And again, those things become quite complicated
00:19:10.410 --> 00:19:16.650
now with the advent of Gen AI. Because you need
00:19:16.650 --> 00:19:19.130
an interface for an intelligent tutoring system,
00:19:19.329 --> 00:19:22.170
right? The student, the learner actually interacts
00:19:22.170 --> 00:19:23.950
with an interface. That kind of goes without
00:19:23.950 --> 00:19:28.440
saying. But what's in the box is free. models
00:19:28.440 --> 00:19:31.819
is the expert model is the student model and
00:19:31.819 --> 00:19:34.720
the tutor model and they basically all three
00:19:34.720 --> 00:19:36.799
of them do what it says on the tin basically
00:19:36.799 --> 00:19:40.279
so the expert model is all the expert knowledge
00:19:40.279 --> 00:19:43.819
about the subject area that can be the grammar
00:19:43.819 --> 00:19:45.839
of a language if you're teaching grammar for
00:19:45.839 --> 00:19:48.900
example so that knowledge has to be encoded in
00:19:48.900 --> 00:19:53.180
an intelligent tutoring system and there were
00:19:53.180 --> 00:19:55.420
other expert systems around that were not for
00:19:55.420 --> 00:19:58.500
tutoring so people are aware within the artificial
00:19:58.500 --> 00:20:03.019
intelligence world about creating these complicated
00:20:03.109 --> 00:20:06.589
databases if you want about expert knowledge
00:20:06.589 --> 00:20:09.470
in some way and then drawing inference from those
00:20:09.470 --> 00:20:12.670
pieces of information student model the same
00:20:12.670 --> 00:20:15.589
thing you're gathering data about individual
00:20:15.589 --> 00:20:18.230
students their learning behavior their learning
00:20:18.230 --> 00:20:22.140
outcomes and things that And you put this in
00:20:22.140 --> 00:20:24.339
a student profile, you make this into a data
00:20:24.339 --> 00:20:27.319
structure, and then it becomes a model because
00:20:27.319 --> 00:20:30.619
you're drawing inferences from it. So you take
00:20:30.619 --> 00:20:33.460
two, to simplify, to take two bases of information
00:20:33.460 --> 00:20:37.700
and you can infer a third that is not in the
00:20:37.700 --> 00:20:41.440
profile. Tutor model, the easiest way to describe
00:20:41.440 --> 00:20:46.259
it is it's the instructional sequences. It's
00:20:46.259 --> 00:20:49.410
when to give feedback and to kind of. do the
00:20:49.410 --> 00:20:52.069
feedback pipeline in some way. That's what the
00:20:52.069 --> 00:20:55.150
tutor model looks after. And an intelligent tutoring
00:20:55.150 --> 00:20:59.430
system and ITS has all of those. If you compare
00:20:59.430 --> 00:21:03.869
this with LLM -based, large language model -based
00:21:03.869 --> 00:21:08.029
chatbots, basically, they can produce fluent
00:21:08.029 --> 00:21:11.990
and plausible feedback. But it might not be necessarily
00:21:11.990 --> 00:21:14.950
that they identified the mistake correctly or
00:21:14.950 --> 00:21:18.920
even the feedback might not be as correct as
00:21:18.920 --> 00:21:21.759
you would want it to be. That's the advantage
00:21:21.759 --> 00:21:26.579
of an ITS. That would be programmed in a hard
00:21:26.579 --> 00:21:31.819
way, basically. Teachers need to look at those
00:21:31.819 --> 00:21:34.299
claims carefully. I think that's the pragmatic
00:21:34.299 --> 00:21:39.410
way of dealing with this. Of course, there are
00:21:39.410 --> 00:21:42.210
huge improvements in the feedback these chatbots
00:21:42.210 --> 00:21:45.390
can give. If I remember the times of ChatGPT
00:21:45.390 --> 00:21:48.769
3 .5, it was very different to what I can do
00:21:48.769 --> 00:21:52.769
now with 5 .6 or 6 .0, basically. The one thing
00:21:52.769 --> 00:21:54.869
that remained the same is the overconfidence
00:21:54.869 --> 00:21:57.470
of the machine. It always has the same level.
00:21:57.549 --> 00:21:59.849
No, I'm right and you're wrong. And then it starts
00:21:59.849 --> 00:22:04.079
apologizing, basically. What has improved with
00:22:04.079 --> 00:22:08.099
those chatbots basically is what they call memory.
00:22:08.259 --> 00:22:11.279
So what it remembers from the current chat session,
00:22:11.440 --> 00:22:14.440
but also sometimes what it remembers across chat
00:22:14.440 --> 00:22:17.039
sessions, what you put in the overall configuration
00:22:17.039 --> 00:22:22.150
of your tool. Caution there, that's not a student
00:22:22.150 --> 00:22:26.150
model because it does not really yet, that might
00:22:26.150 --> 00:22:28.829
change very soon, draw inferences basically.
00:22:31.150 --> 00:22:34.990
Of course, you can prompt a chatbot such that
00:22:34.990 --> 00:22:37.690
you get Socratic dialogues. And of course, you
00:22:37.690 --> 00:22:41.369
learn there is a teaching presence in a Socratic
00:22:41.369 --> 00:22:42.970
dialogue. I think that's kind of the name says
00:22:42.970 --> 00:22:45.549
it already. That's why Socrates did those things.
00:22:46.730 --> 00:22:51.230
But does this necessarily create the best possible
00:22:51.230 --> 00:22:55.970
or even a sound instructional sequence is doubtful.
00:22:56.619 --> 00:22:58.640
Often the case, but doesn't have to be the case.
00:22:58.660 --> 00:23:01.480
So it's check -in and double check -in. So you
00:23:01.480 --> 00:23:03.619
can't leave the learner alone, basically, in
00:23:03.619 --> 00:23:09.000
some way. Yeah, and if we go back to a recent
00:23:09.000 --> 00:23:14.019
one of our podcasts about Sam Illingworth's slow
00:23:14.019 --> 00:23:17.720
AI, he talks about honest comparison classes.
00:23:17.819 --> 00:23:23.480
So is in the absence of a true intelligent tutoring
00:23:23.480 --> 00:23:29.359
system that... essentially does potentially replace
00:23:29.359 --> 00:23:33.410
a human teacher because it does everything the
00:23:33.410 --> 00:23:36.309
human teacher can do. Given that we're not there
00:23:36.309 --> 00:23:42.569
yet, can we find something out there that is
00:23:42.569 --> 00:23:47.710
good enough to be useful again? And I gave the
00:23:47.710 --> 00:23:54.579
example of having a chat with a... ChatGPT playing
00:23:54.579 --> 00:23:58.880
the role in live mode of a German friend who
00:23:58.880 --> 00:24:04.420
was in Bremen with me and was explaining the
00:24:04.420 --> 00:24:09.539
Bremen soccer team and the players and so on
00:24:09.539 --> 00:24:12.819
there as we went to a sports bar to have this
00:24:12.819 --> 00:24:19.509
chat. And in that interaction, I learned some...
00:24:19.930 --> 00:24:23.210
New words. And I had words in German that I thought
00:24:23.210 --> 00:24:27.430
were probably correct. In one case, correct that
00:24:27.430 --> 00:24:30.349
it was incorrect. What I thought, I thought the
00:24:30.349 --> 00:24:35.029
word for goal was goal instead of tour. But I
00:24:35.029 --> 00:24:38.670
also had things confirmed. So the word for sports
00:24:38.670 --> 00:24:45.329
bar is sports bar. And that my ability to interact
00:24:45.329 --> 00:24:50.539
with that material and that character. that I
00:24:50.539 --> 00:24:53.799
had essentially brought into being through my
00:24:53.799 --> 00:24:59.710
prompt, it gave me lightweight tutorial information.
00:24:59.910 -->
00:25:03.609
I had actually asked it not to correct me during
00:25:03.609 --> 00:25:07.069
the talk, and it did once anyway, because that's
00:25:07.069 --> 00:25:14.809
what AIs do. So, yeah, I remember that, our conversation
00:25:14.809 --> 00:25:21.079
in a previous episode about soccer. So it's a
00:25:21.079 --> 00:25:23.720
good example. It basically shows, again, what
00:25:23.720 --> 00:25:25.819
we've already mentioned, just to kind of sum
00:25:25.819 --> 00:25:30.049
up a little bit, the teacher, whenever you have,
00:25:30.089 --> 00:25:33.710
even if you have a chatbot for an AI task, what
00:25:33.710 --> 00:25:37.869
Sam Ellingworth calls an AI task, there has to
00:25:37.869 --> 00:25:43.750
be a teacher presence before, during, and or
00:25:43.750 --> 00:25:47.369
after, basically, the actual task. There is an
00:25:47.369 --> 00:25:52.170
involvement there in many ways. Yeah, so looking
00:25:52.170 --> 00:25:56.109
at what teachers could do with all this, well,
00:25:56.210 --> 00:26:00.349
first, we've said it before, experience this
00:26:00.349 --> 00:26:04.049
chatbot in some sort of tutorial mode yourself.
00:26:04.849 --> 00:26:07.829
And you can do it in the language that you're
00:26:07.829 --> 00:26:10.450
teaching, where you play the role of a lower
00:26:10.450 --> 00:26:12.990
level student than you actually are. I've done
00:26:12.990 --> 00:26:15.450
that for English. And then, of course, do it
00:26:15.450 --> 00:26:19.769
in a language you're trying. to learn or a language
00:26:19.769 --> 00:26:22.910
that you don't know well yet so that you can
00:26:22.910 --> 00:26:26.210
experience it directly from the learner's perspective
00:26:26.210 --> 00:26:31.990
there are all sorts of commercial products out
00:26:31.990 --> 00:26:34.150
there that are trying to do this some of them
00:26:34.150 --> 00:26:37.650
working on very specific areas like just spoken
00:26:37.650 --> 00:26:41.430
communication others more more general kind of
00:26:41.430 --> 00:26:44.549
like duolingo claiming to you know to teach language
00:26:45.220 --> 00:26:48.200
pretty much as a whole. These are fine -tuned
00:26:48.200 --> 00:26:50.799
for language learning. They often involve multiple
00:26:50.799 --> 00:26:56.200
gen AIs and other forms of AI brought together.
00:26:57.299 --> 00:27:00.420
So compare those versus the ones you or your
00:27:00.420 --> 00:27:03.680
students can prompt just using chat GPT or Gemini
00:27:03.680 --> 00:27:09.079
or Claude or something. In both cases, you ask
00:27:09.079 --> 00:27:12.099
yourself, is the teaching presence what you want?
00:27:12.759 --> 00:27:17.170
Another thing, don't expect perfection. Actually,
00:27:17.210 --> 00:27:20.509
we don't get it from humans, so it's not surprising
00:27:20.509 --> 00:27:25.220
that we don't yet get it from the AIs. And again,
00:27:25.319 --> 00:27:28.599
the ultimate question is, is a synthetic tutor
00:27:28.599 --> 00:27:33.000
useful enough to warrant sending your students
00:27:33.000 --> 00:27:38.039
out to work with it? And of course, it's up to
00:27:38.039 --> 00:27:41.000
you, as Matt says. You should be in the loop
00:27:41.000 --> 00:27:45.740
before, during, if possible, but certainly before
00:27:45.740 --> 00:27:52.130
and after. the students' interactions. One way
00:27:52.130 --> 00:27:55.069
for the during, of course, is to train students
00:27:55.069 --> 00:27:57.930
in how to get the most from a chatbot, especially
00:27:57.930 --> 00:28:01.849
as a conversation partner. I continue to be a
00:28:01.849 --> 00:28:05.089
little disappointed that I don't see much in
00:28:05.089 --> 00:28:07.869
the way of research studies on that. It's always
00:28:07.869 --> 00:28:10.849
turn the students loose with this, maybe show
00:28:10.849 --> 00:28:14.309
them how technically to use it, but not spend
00:28:14.309 --> 00:28:18.630
time helping them. you know, be their own, in
00:28:18.630 --> 00:28:21.529
a sense, be their own teacher presence themselves
00:28:21.529 --> 00:28:25.210
while they're engaging with this. And finally,
00:28:25.250 --> 00:28:28.049
remember that the chatbots aren't just for second
00:28:28.049 --> 00:28:31.529
language conversation. You can use a prompt to
00:28:31.529 --> 00:28:34.690
create a tutor persona for giving writing feedback,
00:28:34.849 --> 00:28:38.460
for example, by saying something like, you know,
00:28:38.539 --> 00:28:43.920
you are an expert language teacher who is teaching
00:28:43.920 --> 00:28:46.289
a writing class, and then you get the feedback
00:28:46.289 --> 00:28:49.630
more in that sort of form. So those are just
00:28:49.630 --> 00:28:53.069
a few thoughts here. Those are very good goals,
00:28:53.170 --> 00:28:58.380
I think, that I could set myself as a teacher,
00:28:58.619 --> 00:29:01.299
as a language teacher, and language teachers
00:29:01.299 --> 00:29:05.200
can set for themselves. I would also like to
00:29:05.200 --> 00:29:10.730
draw the attention to what we always talk about,
00:29:10.849 --> 00:29:14.609
SIPD. Sustained Integrated Professional Development.
00:29:15.349 --> 00:29:18.470
There are two principles that are particularly
00:29:18.470 --> 00:29:20.750
relevant here. They're always all kind of in
00:29:20.750 --> 00:29:24.170
some way relevant or in another way. This SIPD
00:29:24.170 --> 00:29:28.130
can also be pronounced as SIPT. Do this as microlearning.
00:29:28.250 --> 00:29:32.869
Do this in small steps. Don't try to kind of
00:29:32.869 --> 00:29:35.250
turn the chatbot into a tutor. Kind of immediately
00:29:35.250 --> 00:29:39.119
learn. whenever necessary, whenever possible,
00:29:39.220 --> 00:29:44.039
in small spurts, basically. And the same on what
00:29:44.039 --> 00:29:46.440
I would call the implementation level when you
00:29:46.440 --> 00:29:49.660
try it out in class. Yes, not that much has been
00:29:49.660 --> 00:29:52.759
done yet in research. I agree. I haven't seen
00:29:52.759 --> 00:29:56.880
it either. But talking from my own experience
00:29:56.880 --> 00:30:00.779
as a classroom teacher, I'm also careful as to
00:30:00.779 --> 00:30:04.180
how much I put in, how much I change. So keep
00:30:04.180 --> 00:30:07.960
it simple. Keep it in the control. Keep in charge.
00:30:08.140 --> 00:30:10.440
The teaching and learning is always much more
00:30:10.440 --> 00:30:14.500
important than the introduction of a piece of
00:30:14.500 --> 00:30:18.160
new technology, basically. Okay. Well, I guess
00:30:18.160 --> 00:30:23.259
that's all for this podcast. Goodbye. And one
00:30:23.259 --> 00:30:28.519
recommendation, read Matt's Panta Rai blog post
00:30:28.519 --> 00:30:31.940
on this topic from last December. We will put
00:30:31.940 --> 00:30:35.220
the link into the show notes. You're so kind.
00:30:35.299 --> 00:30:38.579
I did not see that coming. This is nice. Yes.
00:30:38.859 --> 00:30:42.619
Have a look at the blog. This concludes our episode
00:30:42.619 --> 00:30:46.700
18. We're kind of very happy about it. In some
00:30:46.700 --> 00:30:49.700
countries, by law, we've reached maturity with
00:30:49.700 --> 00:30:53.880
18. We're very happy about the reception of this
00:30:53.880 --> 00:30:56.859
podcast. So far, I checked yesterday, actually,
00:30:56.940 --> 00:31:01.849
we had... 1800 downloads in the previous seven
00:31:01.849 --> 00:31:04.769
months the podcast has been around now for about
00:31:04.769 --> 00:31:07.829
seven months we hope you find our conversations
00:31:07.829 --> 00:31:12.839
useful and learn as much as we do when we're
00:31:12.839 --> 00:31:15.920
talking basically when you're listening basically
00:31:15.920 --> 00:31:21.980
that you learn as much as we do and why don't
00:31:21.980 --> 00:31:25.880
you share the link to this podcast with One colleague
00:31:25.880 --> 00:31:28.700
or more, if you like, to give you the link again,
00:31:28.759 --> 00:31:35.019
because it's very, very simple. RSS .com, RSS
00:31:35.019 --> 00:31:40.079
.com slash podcasts, obviously one word, slash
00:31:40.079 --> 00:31:47.900
again, while O -A -I -L -L for opening AI for
00:31:47.900 --> 00:31:50.140
language learning. Send this to a colleague,
00:31:50.319 --> 00:31:52.839
fellow grad student, somebody you think might
00:31:52.839 --> 00:31:56.579
be interested. in what we are talking about,
00:31:56.660 --> 00:32:00.079
basically, and interested in learning more about
00:32:00.079 --> 00:32:03.400
artificial intelligence, because remember, artificial
00:32:03.400 --> 00:32:06.579
intelligence is no substitute for natural ignorance.
00:32:07.599 --> 00:32:17.970
Take care. We are grateful for the support for
00:32:17.970 --> 00:32:20.849
opening AI for language learning by the Language
00:32:20.849 --> 00:32:23.109
and Applied Research Center at San Diego State
00:32:23.109 --> 00:32:25.650
University and the Southern Area International
00:32:25.650 --> 00:32:28.930
Languages Network, SALEN, which is part of the
00:32:28.930 --> 00:32:32.319
California World Languages Project. Mario Condefinol
00:32:32.319 --> 00:32:35.279
is the production coordinator of Wild. Our theme
00:32:35.279 --> 00:32:38.779
music is by Tillman Spiegel. Our editor is Chris
00:32:38.779 --> 00:32:42.000
Brown. Live conversations are moderated and promoted
00:32:42.000 --> 00:32:44.880
by me, Shanaza Madian. Until next time.