Fig. 2.

ConvNet architecture. The input tweet is represented as a D × M matrix, where M is the length of the word vector representations. A collection of K kernels of size k_i × M (k_i equal to 2 and 3 in this illustration with multiple filters for each filter width) are convolved with the input to produce K outputs each of size D − k_i + 1. In the max pooling layer, the maximum of each convolution is collected into a single vector which is the input to a fully connected softmax layer used to compute the log-probability of the presence of an ADR in the input tweet