topic_shift
bool
2 classes
utterance
stringlengths
1
7.9k
session_id
stringlengths
7
14
false
Oh , you don't know. OK.
QMSum_120
false
Yeah.
QMSum_120
false
Alright.
QMSum_120
false
Um , yeah. So the points were the the weights how to weight the different error rates that are obtained from different language and and conditions. Um , it 's not clear that they will keep the same kind of weighting. Right now it 's a weighting on on improvement.
QMSum_120
false
Mm - hmm.
QMSum_120
false
Some people are arguing that it would be better to have weights on uh well , to to combine error rates before computing improvement. Uh , and the fact is that for right now for the English , they have weights they they combine error rates , but for the other languages they combine improvement. So it 's not very consistent. Um
QMSum_120
false
Mm - hmm.
QMSum_120
false
Yeah. The , um Yeah. And so Well , this is a point. And right now actually there is a thing also , uh , that happens with the current weight is that a very non - significant improvement on the well - matched case result in huge differences in in the final number.
QMSum_120
false
Mm - hmm.
QMSum_120
false
And so , perhaps they will change the weights to
QMSum_120
false
Hmm.
QMSum_120
false
Yeah.
QMSum_120
false
How should that be done ? I mean , it it seems like there 's a simple way
QMSum_120
false
Mm - hmm.
QMSum_120
false
Uh , this seems like an obvious mistake or something.
QMSum_120
false
Well , I mean , the fact that it 's inconsistent is an obvious mistake.
QMSum_120
false
Th - they 're
QMSum_120
false
But the but , um , the other thing
QMSum_120
false
In
QMSum_120
false
I don't know I haven't thought it through , but one one would think that each It it 's like if you say what 's the what 's the best way to do an average , an arithmetic average or a geometric average ?
QMSum_120
false
Mm - hmm.
QMSum_120
false
It depends what you wanna show.
QMSum_120
false
Mm - hmm.
QMSum_120
false
Each each one is gonna have a different characteristic.
QMSum_120
false
Yeah.
QMSum_120
false
So
QMSum_120
false
Well , it seems like they should do , like , the percentage improvement or something , rather than the absolute improvement.
QMSum_120
false
Tha - that 's what they do.
QMSum_120
false
Well , they are doing that.
QMSum_120
false
Yeah.
QMSum_120
false
No , that is relative. But the question is , do you average the relative improvements or do you average the error rates and take the relative improvement maybe of that ?
QMSum_120
false
Yeah. Yeah.
QMSum_120
false
And the thing is it 's not just a pure average because there are these weightings.
QMSum_120
false
Oh.
QMSum_120
false
It 's a weighted average. Um.
QMSum_120
false
Yeah. And so when you average the the relative improvement it tends to to give a lot of of , um , importance to the well - matched case because the baseline is already very good and , um , i it 's
QMSum_120
false
Why don't they not look at improvements but just look at your av your scores ? You know , figure out how to combine the scores
QMSum_120
false
Mm - hmm.
QMSum_120
false
with a weight or whatever , and then give you a score here 's your score. And then they can do the same thing for the baseline system and here 's its score. And then you can look at
QMSum_120
false
Mm - hmm.
QMSum_120
false
Well , that 's what he 's seeing as one of the things they could do.
QMSum_120
false
Yeah.
QMSum_120
false
It 's just when you when you get all done , I think that they pro I m I I wasn't there but I think they started off this process with the notion that you should be significantly better than the previous standard.
QMSum_120
false
Mm - hmm.
QMSum_120
false
And , um , so they said " how much is significantly better ? what do you ? " And and so they said " well , you know , you should have half the errors , " or something , " that you had before ".
QMSum_120
false
Mm - hmm. Hmm.
QMSum_120
false
Mm - hmm.
QMSum_120
false
Yeah.
QMSum_120
false
So it 's , uh , But it does seem like
QMSum_120
false
Hmm.
QMSum_120
false
i i it does seem like it 's more logical to combine them first and then do the
QMSum_120
false
Combine error rates and then
QMSum_120
false
Yeah.
QMSum_120
false
Yeah. Well
QMSum_120
false
Yeah.
QMSum_120
false
But there is this this is this still this problem of weights. When when you combine error rate it tends to give more importance to the difficult cases , and some people think that
QMSum_120
false
Oh , yeah ?
QMSum_120
false
well , they have different , um , opinions about this. Some people think that it 's more important to look at to have ten percent imp relative improvement on well - matched case than to have fifty percent on the m mismatched , and other people think that it 's more important to improve a lot on the mismatch and So , bu
QMSum_120
false
It sounds like they don't really have a good idea about what the final application is gonna be.
QMSum_120
false
l de fff ! Mmm.
QMSum_120
false
Well , you know , the the thing is that if you look at the numbers on the on the more difficult cases , um , if you really believe that was gonna be the predominant use , none of this would be good enough.
QMSum_120
false
Yeah. Mmm. Yeah.
QMSum_120
false
Nothing anybody 's
QMSum_120
false
Mm - hmm.
QMSum_120
false
whereas you sort of with some reasonable error recovery could imagine in the better cases that these these systems working. So , um , I think the hope would be that it would uh , it would work well for the good cases and , uh , it would have reasonable reas soft degradation as you got to worse and worse conditions. Um.
QMSum_120
false
Yeah. I I guess what I 'm I mean , I I was thinking about it in terms of , if I were building the final product and I was gonna test to see which front - end I 'd I wanted to use , I would try to weight things depending on the exact environment that I was gonna be using the system in.
QMSum_120
false
But but No.
QMSum_120
false
If I
QMSum_120
false
Well , no well , no. I mean , it isn't the operating theater. I mean , they don they they don't they don't really know , I think.
QMSum_120
false
Yeah.
QMSum_120
false
I mean , I th
QMSum_120
false
So if if they don't know , doesn't that suggest the way for them to go ? Uh , you assume everything 's equal. I mean , y y I mean , you
QMSum_120
false
Well , I mean , I I think one thing to do is to just not rely on a single number to maybe have two or three numbers ,
QMSum_120
false
Yeah.
QMSum_120
false
you know ,
QMSum_120
false
Right.
QMSum_120
false
and and and say here 's how much you , uh you improve the , uh the the relatively clean case and here 's or or well - matched case , and here 's how here 's how much you ,
QMSum_120
false
Mm - hmm.
QMSum_120
false
uh
QMSum_120
false
So not
QMSum_120
false
So.
QMSum_120
false
So not try to combine them.
QMSum_120
false
Yeah. Uh , actually it 's true.
QMSum_120
false
Yeah.
QMSum_120
false
Uh , I had forgotten this , uh , but , uh , well - matched is not actually clean. What it is is just that , u uh , the training and testing are similar.
QMSum_120
false
The training and testing.
QMSum_120
false
Mmm.
QMSum_120
false
So , I guess what you would do in practice is you 'd try to get as many , uh , examples of similar sort of stuff as you could , and then ,
QMSum_120
false
Yeah.
QMSum_120
false
uh So the argument for that being the the the more important thing , is that you 're gonna try and do that , but you wanna see how badly it deviates from that when when when the , uh it 's a little different.
QMSum_120
false
So
QMSum_120
false
Um ,
QMSum_120
false
so you should weight those other conditions v very you know , really small.
QMSum_120
false
But No. That 's a that 's a that 's an arg
QMSum_120
false
I mean , that 's more of an information kind of thing.
QMSum_120
false
that 's an ar Well , that 's an argument for it , but let me give you the opposite argument. The opposite argument is you 're never really gonna have a good sample of all these different things.
QMSum_120
false
Uh - huh.
QMSum_120
false
I mean , are you gonna have w uh , uh , examples with the windows open , half open , full open ? Going seventy , sixty , fifty , forty miles an hour ? On what kind of roads ?
QMSum_120
false
Mm - hmm.
QMSum_120
false
With what passing you ? With uh , I mean ,
QMSum_120