e4dd5528f7596dcdf871aa55cfccc53c-AuthorFeedback.pdf

We thank all reviewers for their detailed and constructive comments. "problem [...] is relevant and important," "dataset is original," Apologies for the confusion; we will clarify. We will include results for the upper bound in Table 2 as requested by R2 . R1: Contribution of stage 2: If we remove stage 2 and zero out weights for text embedding, acc. is only 0.677. R1: "Sweet spot" for text data: We will include an experiment that trains with the first k sentences (varying k).