TaLK need to divide by the area of span? about talkconvolutions

[lioutasb]

We tried normalizing by the size of the summed area (effectively making it an adaptive average pooling) but this normalization made the training unstable. We believe this is because we optimize the parameters ᾶ^l and ᾶ^r for both the optimal window size as well as bringing the magnitude of the representation down.

[gaopengcuhk]

Thank you for your response.
However, without normalization, the output value will be very large as the window size increase.