I am studying machine learning as part of my degree in information science, and I’m taking this course in order to learn practical approach to deep learning.
I am currently working on trying to implement a model presented this paper.
In short, they are using CNNs on a dataset of statements taking from politifact.com (dataset combines textual statements with various metadata features - representing speaker identity and affiliation). The suggested model is a hybrid CNN, combining on one hand the text (using word embeddings based on pretrained vectors from word2vec) and the metadata features, which are transformed through embedding and a convolutional layer and concatenated to create the complete model.
I am trying to reconstruct the model, based on the material from Lesson 4-5, and trying to fully understand their description of the model in order to replicate their results.
The information given in the paper regarding the model is as follows:
The best filter sizes for the text-only CNN model was (2,3,4). In all cases, each size has 128 filters. The dropout keep probabilities was optimized to 0.8. For the hybrid model, we use 3 and 8 as filter sizes, and the number of filters was set to 10. We considered 0.5 and 0.8 as dropout probabilities.
How would you repm lesson 4-5 for IMDB dataset for this purpose, but so far resent this model? I’m trying to modify the model fromy results fail to achieve any result significantly better than random (although the paper also has fairly low accuracy in their results).