[slig]

What about using the those three models and returning the best score between them?

[michalg82]

That's some really interesting idea. But what if it will make too many "false positives"? Maybe too many word pairs will be considered more related that one could expect.