How can I calculate "probability" for a result segment?

[toanhn]

I see in result segments, each word has the value "probability", but in each segment, I do not see the value "probability".
How can I calculate "probability" for a result segment?
Thanks

[phineas-pta]

it's complicated

[MahmoudAshraf97]

you need to define what does the probability for each segment means or represents before you ask how to calculate it

[dgoryeo]

I guess there is no one-size-fits-all solution and each language or context may have its own characteristics. I found the question useful/relevant, however I have no subject matter experise so I asked GPT-o1-preview for possible approaches. FWIW, here is a summary of the output:

`
Here are some scientifically sound methods to compute the segment probability:

1. Product of Probabilities (Joint Probability)
   Method: Calculate the joint probability of the segment by multiplying the probabilities of all words in the segment.

2. Sum of Log Probabilities
   Method: Convert word probabilities to log probabilities, sum them up, and use the total log probability as the segment's confidence score.

3. Average Log Probability
   Method: Calculate the average of the log probabilities of the words in the segment.

4. Geometric Mean of Probabilities
   Method: Calculate the geometric mean of the word probabilities.

`

[gongouveia]

@toanhn @phineas-pta It is not complicated, in the output of the inference, you have access to the logits of each word, with this logit you can convert it to a probability of each word (type of confidence). Mind that the logit of the probability is the logarithm of the odds.
Now you can do the average of the probabilities :)

Let me know if this message was helpful.

[toanhuynhnguyen]

Thanks bro