lzr | Laser show projection software , and ILDA | Game Engine library

The solution comprises two steps:

1. Group words to closest keyword (I wouldn't call it clustering as the centers of the groups are already given here, as opposed to clustering where you try to find clusters with no a priori known locations)
2. Remove outliers that don't seem to really belong to this keyword, although this keywords is the closest by distance.

Grouping is straightforward by assigning keyword numbers by distance using vector quantization. The only thing we have to bear in mind here is that the keyword numbers in the original dataframe don't appear in ordered sequence, but in vq groups are numbered in sequence starting with 0. That's why we have to map the new keyword group numbers to the given keyword numbers.

Removing outliers can be done in different ways, and there were no strict requirements in the question on how the keyword groups should be formed. I chose a very simple approach: take the mean and standard deviation of the distances from the keyword to all keyword group members and consider words with distances greater than Mean + x * StdDev as outliers. A choice of x = 1.5 gives good results.