Evaluation of Algorithm. Figure 5.3 and figure 5.4 are examples generated by visualizing the dimension weights for sense “obtain.v.01”, which has a definition of “come into possession of”, and sense “have.v.01”, which has a meaning of “have or possess, whether in concrete or abstract way”, after applying our algorithm on the ELMo model with a dimension of 512, using the second layer as the output layer. As shown in the graph, the value of dimensions varies with some dropped to 0. The values of dimension weights can be interpreted as the value of that certain dimension in representing the current sense “obtain.v.01”. The red line in the graph is the threshold where every dimension with weights below the threshold will be masked out, which is set to 0.6 here. Figure 5.3 and figure 5.6 show the mean word embedding of sense group “obtain.v.01” and “have.v.01” with orange dots representing the dimensions that have been masked out to 0. Figure 5.7 and figure 5.7 are graphs based on the ELMo embedding model with a dimension size of 512 and the third network layer as the output layer. Figure 5.7 is a graph of 20 selected sense groups with 100 embeddings each. Figure 5.8 is a graph of 20 selected sense groups with 100 embeddings each and the dimensions of the embedddings with a weight value smaller than
