TER scores are not scaled

[DhruvAwasthi]

I am trying to calculate the TER score that should be in the range of 0-100.
According to the new update, scores are scaled automatically.
The code that I am using to calculate scaled TER scores is:

metric = sacrebleu.metrics.TER()  
cal_score = metric.corpus_score(hypotheses, [references])
score = cal_score.score

But sometimes the score comes out to be greater than 100.
Can you please tell me is it something that I am doing wrong or is it something else?

Thank you!

[ozancaglayan]

Hi,

As far as I know TER scores can be larger than 100 if the hypotheses are way too far from the references. This is also the case with the reference TERCOm library.

[DhruvAwasthi]

Ah okay!

Thank you for the prompt response, appreciate it!

[DhruvAwasthi]

@ozancaglayan @mjpost Even though TER scores can be larger than 100 then what is the meaning of "scaled TER scores"?
Ideally, when we say scaled it should be within the scale.

[ozancaglayan]

Before, they were between 0-1, now all metrics are multiplied by 100, scaling was simply this :)

[DhruvAwasthi]

Oh okay, thank you for the clarificaiton!