Add time sensitive approach

[mcollardanuy]

No description provided.

[kasra-hosseini]

@mcollardanuy I reviewed the implementation of weighted_gauss function (as I think the other changes are related to using weighted_gauss in the pipeline)

My only comment is on these lines:

https://github.com/Living-with-machines/TargetedSenseDisambiguation/blob/try-gauss/tasks/wsd.py#L244-L245

I was wondering if it is easier to work with offset and scale directly without weighting them by (max_date-min_date). Here is one example if you want to try it yourself (well, it is based on your code):

import matplotlib.pyplot as plt
import numpy as np

origin = 1850
min_date = 1800
max_date = 1900
year_arr = np.linspace(min_date, max_date, 100)
year_arr = np.sort(year_arr)

offset = 10.0
scale = 10.0

#offset = offset * (max_date-min_date)
#scale = scale * (max_date-min_date)

d = (abs(year_arr - origin)).clip(offset, None)
expr = 2**(-((d-offset)/scale)**2)

plt.plot(year_arr, expr)
plt.axvline(origin - offset, 0, 1, color="r", ls="--")
plt.axvline(origin + offset, 0, 1, color="r", ls="--")
plt.axvline(origin - offset - scale, 0, 1, color="g", ls="--")
plt.axvline(origin + offset + scale, 0, 1, color="g", ls="--")
plt.show()

[kasra-hosseini]

The output of the above code is:

[kasra-hosseini]

which gives an easy way to control the flat top, scale and etc... Of course, the following also works:

offset = offset * (max_date-min_date)
scale = scale * (max_date-min_date)

but probably the first method is easier to control for this application?

[kasra-hosseini]

Otherwise, it is good from my side.