matrix_positions in manual_bigram_penalty
Closed this issue · 5 comments
when evaluating the "manual_bigram_penalty" score, it seems that the key positions are interpreted in the inverse way : [to_position, from_position] and not in intended and mentioned in the comments [from_position, to_position]
I think the culprit is this line in manual_bigram_penality.rs:
.map(|(((x1, y1), (x2, y2)), w)| (((*x2, *y2), (*x1, *y1)), *w)),
It should be ? :
.map(|(((x1, y1), (x2, y2)), w)| (((*x1, *y1), (*x2, *y2)), *w)),
Please let me know if I'm wrong,
Thanks and have a nice day
In case of the manual bigram penalties, the "mirrored" bigrams to those configured ones are added automatically.
This is because usually both directions of a bigram are bad and this way, you only configure one direction. The other ist included automatically with the same weight.
We could consider making this automatic addition optional, though.
ok,
For my case, I was using it for easy to type bigrams, with a negative score. and I made sure to configure manually both directions (usually with different values).
When artificially elevating a single bigram score, it was its inverse that showed on "worst", and the evolution algorithm favored its inverse bigram. Thats why I thought it was reading in inverse.
I will make the auto-inclusion of the mirrored versions configurable, when I'll find the time.
One thing to note (as it may not be obvious): The "worst" elements are in fact those with the highest absolute value of the cost. So in the case of negative costs, it would actually be "best". I decided to do it that way because otherwise, you would only see elements with zero score in that case, which would not give much insight.
It is better with absolute value, as it can work on both ways.
For my use case, I just changed the global weight to a negative value, while keeping each matrix_position's weight positive.
I was aware about the part where "worst" being the "best" :D, but for some reasons when I was testing, it was the mirrored version that kept being shown (probably some values were overwritten by the automatic mirroring of manually mirrored positions)
anyway, I commented that part of the code for now
an option would be more than helpful.
and thanks again for the great work
There is now a configuration parameter add_mirrored in the manual_bigram_penalty metric.