A language model is tasked with finding an answer span within a text. After processing the text, it produces the following log-probabilities for each token being the start or the end of the answer. To find the best span, one must find the start index ($j_1$) and end index ($j_2$) that maximize the sum of their log-probabilities, with the constraint that the start index cannot be after the end index ($j_1 \le j_2$).

| Index | Token | Start Log-Prob | End Log-Prob | |-------|-----------|-----

Analyzing the Span Prediction Constraint

Diagnosing a Span Prediction Error