Get coherence score for PLSA

[lyriccoder]

PLSA and other methods gives strange coherence score:

PLSA(n_components=3).fit(data_vec).coherence()
PLSA(n_components=4).fit(data_vec).coherence()

n=5, -894.0931521853117
n=4, -846.5056881515624
n=1000, -548.1772075123278

When I use gensim, I get quite a good score:

2 & 0.4492 
3 & 0.4257 
4 & 0.4308 
5 & 0.4443 
6 & 0.4625 
7 & 0.455 
8 & 0.4791 
9 & 0.4897 
10 & 0.5354 
11 & 0.5165 
12 & 0.5149 
13 & 0.5382 
14 & 0.5546 
15 & 0.5669 
16 & 0.5633 
17 & 0.5323 

Could you please tell whether there is a bug?