609c5e5089a9aa967232aba2a4d03114-AuthorFeedback.pdf

Effect of UniLM: We observe obvious performance drop when using fine-tuned UniLM2 with fixed top-1 retrieved knowledge (-parameterized posterior in Table 3). (aligned with the columns of Table 1). Our model performs implicit knowledge selection on the input K14 knowledgesentences(concatenatedinasequence)inanend-to-endwaylikeDRD[52]. F1 on the validation set increases until the number of knowledge reaches 10, but22 stays stable when the number increases from 10 to 30. Intesttime,knowledge28 selection module is mainly to shorten the input sequence of knowledge candidates, so the performance drop is not29 significant.