Beam Search Error Analysis

During the 3rd week of the Sequence Models, Andrew Ng talked on Error Analysis of Beam search. He talks of y_asterisk as the probability of the human translation given the x. However, he doesn’t comment (or I didn’t get it :slight_smile: ) how this probability is actually calculated.

Thanks a lot for this exquisite series of courses on DL. It has helped me a lot across my research.

Hi naposto,

Based on Andrew Ng’s disscussion of human level performance in this video, I would say y_asterisk can be determined by asking humans with particular language skills to translate the sentence. The different translations that are produced can be kept track of and the most likely translation (by humans) can be determined.

Thanks @reinoudbosch for your answer. I thought that y_asterisk was computed through the model but I couldn’t understand how. Now, it makes much more sense.