Randomly extract non-overlapping sets

Generate content for a word game. Input is data about relations (A-to-B strength). Output is distant relationships. 

For example: 

GOOD:  apple, airplane, dog, house

BAD: banana, cherry, peach, strawberry. 

You do NOT need to speak very much English. This is purely data. 

I have two source files. A list of ranked relationships between words, and a separate list of words which may be in the first file. This is real English word data, similar to a thesaurus. https://www.powerthesaurus.org/

Task is to randomly output sets of lines from the 2nd file, which are NON-overlapping concepts. 

An ideal algorithm would create a multidimensional mesh, and then randomly extract distant nodes. I don't know how to do that. i.e. output sets of words which are all distant in vector space. See: https://dzone.com/articles/introduction-to-word-vectors

A non-ideal algorithm would randomly pull lines from file 2, measure similarity to other output lines, if dissimilar, keep and remove from file 2. If a line in file 2 is similar to too many test lines, remove it and return to file 2. i.e. a "bag of coins", and you keep randomly testing/replacing coins so they are all different. 

No word pair should be more than 0.3 similar; and the total similarity of all words between sets should be <0.5. 

Preferred programming language is: ruby, perl, python. 

Two input files: 

A) relations.txt

    #aaa [syn]: aab | aac; [syn-score]: 100.0 | 8.0;

    #aab [syn]: aaa | aac; [syn-score]: 75.0 | 5.0;

    #bbb [syn]: bba | bbc; [syn-score]: 50.0 | 4.3;

    #bba [syn]: bbb | bbc; [syn-score]: 150.0 | 1.2;

    #ccc [syn]: ccd | ccz; [syn-score]: 150.0 | 0.4;

    ... etc.

B) lists.txt

    #aaa = aab | aac     

    #bbb = bbd | bba

    #bba = bbd | bbx

    #ccc = cca | ccz

    #cca = ccd | cce

    #ddd = dda | ddb 

    ... etc.    

The real file A is 300+ MB, with 855k lines. 

The real file B is ~15k lines.     

I will want to be able to set N, number of sets; and Y number of packs. N will typically be around 25 sets; and Y will likely be 1000 packs.     

Output, with N=2:

    #aaa = aab | aac     

    #cca = ccd | cce

    #bbb = bbd | bba

    #ddd = dda | ddb 

    #bba = bbd | bbx

    #ccc = cca | ccz

Output, with N=3:

    #aaa = aab | aac     

    #ccc = cca | ccz

    #bba = bbd | bbx

    #bbb = bbd | bba

    #ddd = dda | ddb 

    #cca = ccd | cce