首先分组和聚合必要的字段：

res = df.groupby(["reference", "subject"]).agg({"page": min, "ids": list, "word": lambda l: [[ll] for ll in l]}).reset_index

  reference subject  page                                         ids              word
0     apple   fruit     1  [[bndv, asasa, swdsd], [bsnm, dfsd, dgdf]]    [[is], [text]]
1     apple    name     1                       [[aaaa, bbbbb, cccc]]           [[app]]
2       bat   place     1                           [[fsf, sff, fss]]          [[that]]
3       bat   thing     1      [[asas, ddfgd, ff], [sds, fsss, ssfd]]  [[sport], [was]]
4       cat   fruit     1                          [[sds, dffd, gdg]]         [[color]]
5       dog    name     2                          [[fffds, gd, sdg]]           [[mud]]
6       dog   thing     2                         [[dsfd, fds, gfdg]]           [[kit]]
7       egg   place     2          [[dfff, sdf, vcv], [ddd, fg, dfg]]    [[gun], [hut]]

请注意，这也将每个word值 Package 在一个列表中，就像您希望在所需的输出中所做的那样。我也只是假设在每个组中取最小的page值，因为您没有提到该变量的规则。您可以将agg函数中的min值更新为您认为合适的任何值。
然后，如果length为1，则可以清除列表：

res["word"] = res["word"].apply(lambda l: l[0] if len(l) == 1 else l)
res["ids"] = res["ids"].apply(lambda l: l[0] if len(l) == 1 else l)

  reference subject  page                                         ids              word
0     apple   fruit     1  [[bndv, asasa, swdsd], [bsnm, dfsd, dgdf]]    [[is], [text]]
1     apple    name     1                         [aaaa, bbbbb, cccc]             [app]
2       bat   place     1                             [fsf, sff, fss]            [that]
3       bat   thing     1      [[asas, ddfgd, ff], [sds, fsss, ssfd]]  [[sport], [was]]
4       cat   fruit     1                            [sds, dffd, gdg]           [color]
5       dog    name     2                            [fffds, gd, sdg]             [mud]
6       dog   thing     2                           [dsfd, fds, gfdg]             [kit]
7       egg   place     2          [[dfff, sdf, vcv], [ddd, fg, dfg]]    [[gun], [hut]]