Algorithm 1.

DataSifterText - Masking and Prediction

1:	Input: W, L, blacklist, whitelist, pw, pn, and replacement method.
2:	Construct BERT on documents with downstream tasks as blank words prediction.
3:	for i = 1,…, n do
4:	Set nmask = 0, coef = 1.2.
5:	while nmask ≤ 0.5 * Ti do
6:	for t = 1,…, Ti do
7:	if wt ∈ whitelist then
8:	Set p = 1 − pw * coef.
9:	else if wt ∈ blacklist then
10:	Set p = 0.
11:	else
12:	Set p = 1 − pn * coef.
13:	end if
14:	Mask wt with probability p.
15:	if wt is masked then
16:	nmask = nmask+1
17:	coef = 1.2
18:	else if coef > 0.05 then
19:	coef = coef-0.05
20:	end if
21:	end for
22:	end while
23:	end for
24:	for i = 1,…, n do
25:	for t = 1,…, Ti do
26:	if wt = [MASK] then
27:	Use a trained BERT model to generate P(wt|Wi).
28:	Sample one token with the obtained distribution P(wt|Wi) and replace the [MASK] token at location t.
29:	end if
30:	end for
31:	end for
32:	if Replacement method ≠ (No obfuscation,0) then
33:	Construct D(W).
34:	Use Mini Batch K-means to classify documents into K clusters.
35:	Obtain RAKE(W) or TR(W) based on replacement method specification.
36:	for i = 1,...,n do
37:	Sample min(1,000,nWi) documents in the same cluster as Wi, where nWi is the number of documents in the same cluster as Wi, such that the document index j 6= i.
38:	for j = 1,...,1,000 do
39:	$dist(i,j)=\frac{D(\mathbf{W}{)}_{i}D(\mathbf{W}{)}_j}{\Vert D(\mathbf{W}{)}_i\Vert \Vert D(\mathbf{W}{)}_j\Vert }$
40:	end for
41:	Sample one document from {Wj : dist(i,j) within the smallest top 10% ∀j} as the replacement partner for document i and denote as Wj*.
42:	if Replacement method = (RAKE keyphrase,q) then
43:	Replace RAKE(Wi)q with RAKE(Wj*)q
44:	else if Replacement method = (RAKE index,1) then
45:	Replace all tokens from RAKE(Wi)1 to $W_{T_i}$ with RAKE(Wj*)1 to $W_{T_{j*}}$
46:	else
47:	Replace TR(Wi)q with TR(Wj*)q.
48:	end if
49:	end for
50:	end if
51:	Output: W* = W, and L* = L